I Mapping the Field

Legal procedure determines how legal problems are processed. Many areas of procedure also raise issues of rights, which are established by substantive law and overarching principles, and allocated in the process of dispute resolution. More broadly, legal procedure reflects how authorities can impose a conflict settlement when the individuals involved are unable to do so.

Criminal procedure is an example of legal processing that has evolved over time and developed special characteristics. The state asks the alleged victim to stand back and allow the people to prosecute an individual’s wrongdoing. The state also grants the defendant rights when accused by the people. However, new developments are demanding that criminal procedure adapt in order to maintain its unique characteristics. Adjustments may have to be made as artificial intelligence (AI)Footnote ¹ robots enter criminal investigations and courtrooms.

In Chapter 6, Sara Sun Beale and Hayley Lawrence describe these developments, and using previous research into human–robot interaction,Footnote ² they explain how the manner in which these developments are framed is crucial. For example, human responses to AI-generated evidence will present unique challenges to the accuracy of litigation. The authors argue that traditional trial techniques must be adapted and new approaches developed, such as new testimonial safeguards, a finding that also appears in other chapters (see Section II.B). Beale and Lawrence suggest that forums beyond criminal courts could be designed as sandboxes to learn more about the basics of AI-enhanced fact-finding.

If we define criminal procedural law broadly to include all rules that regulate an inquiry into whether a violation of criminal law has occurred, then the relevance of new developments such as a “Robo-Judge” become even clearer (see Section II.D). Our broad definition of criminal procedure includes, e.g., surveillance techniques enabled by human–robot interaction, as well as the use of data generated by AI systems for criminal investigation and prosecution or fact-finding in court. This Introduction to Part II of the volume will not address the details of other areas such as sentencing, risk assessment, or punishment, which form part of the sanction regime after a verdict is rendered, but relevant discussions will be referred to briefly (Section II.D).

II The Spectrum of Procedural Issues

AI systems play a role in several areas of criminal procedure. The use of AI tools in forensics or predictive analysis reflects a policy decision to utilize new technology. Other areas are affected simply by human–robot cooperation in everyday life, because law enforcement or criminal investigations today make use of data recorded in everyday activities. This accessible data pool is growing quickly as more robots constantly monitor humans. For example, a modern car records manifold data on its user, including infotainment and braking characteristics.Footnote ³ During automated driving, driving assistants such as lane-keeping assistants or drowsiness-detection systems monitor drivers to ensure they are ready to respond to a take-over request if required.Footnote ⁴ If an accident occurs, this kind of alert could be used in legal proceedings in various ways.

III Privacy and Fairness Concerns

The use of human–robot interaction in criminal matters raises manifold privacy and fairness concerns, only some of which can be highlighted here.

IV The Larger Perspective

It seems inevitable that human–robot interaction will impact criminal proceedings, just as it has other areas of the law. However, the exact nature of this impact is unclear. It may help to prevent crime before it happens or it might lead to a merciless application of the law.

Legal scholars primarily point to the risks of AI systems in criminal justice and the need to have adequate safeguards in place. However, many agree that certain robots have the potential to make criminal proceedings faster, and possibly even fairer. One big, not yet fully scrutinized issue will be whether we can and will trust systems that generate information where the decision-making process is opaque to humans, even when it comes to criminal verdicts.Footnote ⁸⁹

Future lawmakers drafting criminal procedure must keep in mind what Tatjana Hörnle pointed out in her introduction to Part I of the volume, that humans tend to blame other humans rather than machines.Footnote ⁹⁰ The same is true for bringing charges against humans as opposed to machines, as explained by Jeanne Gaakeer.Footnote ⁹¹ Part of the explanation for this view lies in the inherent perspectives of substantive and procedural law.Footnote ⁹² Criminal justice is tailored to humans, and it is much easier, for reasons rooted in human understanding and ingrained in the legal framework, to prosecute a human.Footnote ⁹³ This appears to be the case when a prosecution can be directed against either a human or a human–robot cooperation,Footnote ⁹⁴ and it would most probably also be the case if one had to choose between prosecuting a visible human driver or a robot that guided automated driving.

With human–robot interaction now becoming a reality of daily life and criminal justice, it is time for the legal community to reconcile themselves to these challenges, and engage in a new conversation with the computer scientists, behavioral scholars, forensic experts, and other disciplines that can provide relevant knowledge. The digital shift in criminal justice will be manifold and less than predictable. Human–robot interaction might direct more blame in the direction of humans, but it might also open up various new ways to reconstruct the past and possibly assist in exonerating falsely accused humans. A basic condition for benefiting from these developments is to understand the different aspects of human–robot interaction and their ramifications for legal proceedings.

I Introduction

In the courtroom, the phrases artificial intelligence (AI) and robot witnesses (“robo-witnesses”) conjure up images of a Star Wars-like, futuristic world with autonomous robots like C3PO taking the witness stand. Although testimony from a robo-witness may be possible in the distant future, many other kinds of evidence produced by AI are already becoming more common.

Given the wide and rapidly expanding range of activities being undertaken by robots, it is inevitable that robot-generated evidence and evidence from human witnesses who interacted with or observed robots will be presented in legal forums. This chapter explores the effects of human psychology on human–robot interactions (HRIs) in legal proceedings. In Section II, we review the research on HRI in other contexts, such as market research and consumer interactions. In Section III, we consider the effect the psychological responses detailed in Section II may have in litigation.

We argue that human responses to robot-generated evidence will present unique challenges to the accuracy of litigation, as well as ancillary goals such as fairness and transparency, but HRI may also enhance accuracy in other respects. For our purposes, the most important feature of HRI is the human tendency to anthropomorphize robots. Anthropomorphization can generate misleading impressions, e.g., that robots have human-like emotions and motives, and this tendency toward anthropomorphization can be manipulated by designing robots to make them appear more trustworthy and believable. The degree of distortion caused by anthropomorphization will vary, depending on the design of the robot and other situational factors, like how the interaction is framed. The effects of anthropomorphization may be amplified by the simulation heuristic, i.e., how people estimate the likelihood that something happened based on how easy it is for them to imagine it happening, and the psychological preference for direct evidence over circumstantial evidence.Footnote ¹ Moreover, additional cognitive biases may distort fact-finding or attributions of liability when humans interact with or observe robots.

On the other hand, robot-generated evidence may offer unique advantages if it can be presented as direct evidence via a robo-witness, because of the nature of a robo-witness’s memory compared to that of a human eyewitness. We have concerns, however, about the degree to which the traditional methods of testing the accuracy of evidence, particularly cross-examination, will be effective for robot-generated evidence. It is unclear whether lay fact-finders, who are prone to anthropomorphize robots, will be able to understand and evaluate the information generated by complex algorithms, particularly those using unsupervised learning models.

Although it has played a limited role in litigation, AI evidence has been used in other legal forums. Section IV compares the use of testimony from autonomous vehicles (AVs) in litigation with the use of similar evidence in alternative dispute resolution (ADR) and the National Transportation Safety Board (NTSB). These contrasting legal infrastructures present an opportunity to examine AI evidence through a different lens. After comparing and contrasting AI testimony in ADR and NTSB proceedings with traditional litigation, the chapter suggests that the presence of expert decision-makers might help mitigate some of the problems with HRI, although other aspects of the procedures in each forum still raise concerns.

II The Psychology of HRI in Litigation

Although there is no universally agreed-upon definition of “robot,” for our purposes, a robot is “an engineered machine that senses, thinks, and acts.”Footnote ² Practically speaking, that means the robot must “have sensors, processing ability that emulates some aspect of cognition,” and the capacity to act on its decision-making.Footnote ³ A robot must be equipped with programming that allows it to independently make intelligent choices or perform tasks based on environmental stimuli, rather than merely following the directions of a human operator, like a remote controlled car.Footnote ⁴ Under our definition, robots need not be embodied, i.e., they need not occupy physical space or have a physical presence. Of course, the fictitious examples of R2D2 and C3P0 fit our definition, but so too do the self-driving, guided steering, or automatic braking features in modern cars.

III The Impact of the Psychology of HRI in Litigation

In this section, we assume that the psychological phenomena described above will occur outside the laboratory setting and, more specifically, in the courtroom. This is a significant assumption because it is difficult to perfectly extrapolate real-world behavior from experimental studies.Footnote ⁵¹

The cognitive errors associated with people’s tendency to anthropomorphize robots could distort the accuracy and fairness of the litigation process in multiple ways. The current prevalence of these errors may lead to the conclusion that the distortions arising from robot-generated evidence are no greater than those arising from other forms of evidence. Indeed, in some respects, robot-generated evidence might contribute to accuracy because it would be less subject to certain cognitive errors. There remain, however, difficult questions about how well the tools traditionally used to test accuracy in litigation can be adapted to robot-generated evidence, as well as questions about the distributional consequences of developing more persuasive robots.

IV AI Testimony in Other Legal Proceedings

In this section, we consider the impact of HRI in legal proceedings other than litigation, specifically on ADR, with a focus on arbitration, and the specialized procedures of the NTSB. We do so for two reasons. First, in the United States, litigation is relatively rare, and most cases are now resolved by some form of ADR. That is likely to be true of disputes involving robo-witnesses and evidence about the actions of robots as well. Second, these alternatives address what Sections II and III identify as the critical problem in using robot-generated evidence in litigation: the tendency of humans, especially laypersons, to anthropomorphize robots and to misunderstand how human memory functions. In contrast, the arbitration process and the NTSB’s procedures assign fact-finding either to subject matter experts or to decision-makers chosen for their sophistication and their ability to understand the complex technology at issue. In this section, we describe the procedures employed by the NTSB and in arbitration and consider how these forums might address the potential distortions discussed in Sections II and III.

V Conclusion

Human responses to robot-generated evidence will present unique challenges to the accuracy of litigation, as well as the transparency of the legal system and the perceptions of its fairness.

Robot design and framing have the potential to distort fact-finding both intentionally and unintentionally. Robot-generated evidence may be undervalued, e.g., because it is not direct evidence. But such evidence may also be overvalued because of design choices intended to thwart or minimize a robot’s liability or perceived responsibility, and thus the liability of its designers, manufacturers, and owners. Although there are human analogs involving witness selection and coaching, they are subject to natural limits, limits which largely do not apply to the ex ante design-a-witness problem we may see with robots. Additionally, cognitive biases may distort assessments of blame and liability when human and robot actors are both at fault, leading to the failure to impose liability on the designers and producers of robots.

Testing the accuracy of robot-generated evidence will also create new challenges. Traditional cross-examination is ill-suited to this evidence, which may lead to both inaccurate fact-finding and a lack of transparency in the process that could undermine public trust. Cognitive biases can also distort the evaluation of evidence concerning algorithms. The high cost of accessing the most sophisticated robots and mounting the means to challenge them can exacerbate concerns about the fairness and accuracy of the legal system, as well as accessibility to justice. Accordingly, traditional trial techniques need to be adapted and new approaches developed, such as new testimonial safeguards.Footnote ¹¹²

But the news concerning litigation is not all bad. If it is possible to reduce the distorting effects arising from cognitive errors, robot-generated evidence could improve the accuracy of litigation, capturing more data initially and preserving it without the many problems that distort and degrade human memory.Footnote ¹¹³

Finally, alternative forums, such as arbitration and agency proceedings, can be designed to minimize the evaluation of evidence and the imposition of liability on the basis of fact-finding by individuals who lack familiarity with the technology in question.

I Introduction

Criminal prosecutions now routinely involve technologically sophisticated tools for both investigation and proof of guilt, from complex software used to interpret DNA mixtures, to digital forensics, to algorithmic risk assessment tools used in pre-trial detention, sentencing, and parole determinations. As Emily Silverman, Jörg Arnold, and Sabine Gless’s Chapter 8 explains, these tools offer not merely routine measurements, but also “evaluative data” akin to expert opinions.Footnote ¹ These new tools, in critical respects, are a welcome addition to less sophisticated or more openly subjective forms of evidence that have led to wrongful convictions in the past, most notably eyewitness identifications, confessions, and statements of source attribution using “first generation”Footnote ² forensic disciplines of dubious reliability, such as bite marks.Footnote ³

Nonetheless, this new generation of evidence brings new costs and challenges. Algorithmic tools offer uniformity and consistency, but potentially at the expense of equitable safety valves to correct the unjust results that would otherwise flow from mechanistic application of rules. Such tools also may appear more reliable or equitable than they are, as fact-finders fail to identify sources of error or bias because the tools appear objective and are shrouded in black box secrecy. Even with greater transparency, some results, such as the decisions of deep neural networks engaged in deep learning, will not be fully explainable without sacrificing the very complexity that is the ostensible comparative advantage of artificial intelligence (AI). The lack of explainability as to the method and results of sophisticated algorithmic tools has implications for accuracy, but also for public trust in legal proceedings and participants’ sense of being treated with dignity. As Sara Sun Beale and Haley Lawrence note in their Chapter 6 of this volume, humans have strong reactions to certain uses of robot “testimony” in legal proceedings.Footnote ⁴ Absent proper regulation, such tools may jeopardize key systemic criminal justice values, including the accuracy expressed by convicting the guilty and exonerating the innocent, fairness, public legitimacy, and softer values such as mercy and dignity.

In furtherance of these systemic goals, this chapter argues for four overarching principles to guide the use of digital and machine evidence in criminal justice systems: a right to front-end safeguards to minimize error and bias; a right of access both to government evidence and to exculpatory technologies; a right of contestation; and a right to an epistemically competent fact-finding process that keeps a human in the loop. The chapter offers legal and policy proposals to operationalize each principle.

Three caveats are in order. First, this chapter draws heavily on examples from the United States, a decentralized and adversarial system in which the parties themselves investigate the case, find witnesses, choose which evidence to introduce, and root out truth through contestation. Sabine Gless has described the many differences between the US and German approaches to machine evidence, distinguishing their adversarial and inquisitorial approaches, respectively.Footnote ⁵ Nonetheless, the principles discussed here are relevant to any system valuing accuracy, fairness, and public legitimacy. For example, although many European nations have a centralized, inquisitorial system, proposed EU legislation evinces concern over the rights of criminal defendants vis-à-vis AI systems, specifically the potential threat AI poses to a “fair” trial, the “rights of the defense,” and the right to be “presumed innocent,” as guaranteed by the EU Charter of Fundamental Rights.Footnote ⁶ As noted in Chapter 10 of this volume by Bart Custers and Lenneke Stevens, European nations are facing similar dilemmas when it comes to the regulation of digital evidence in criminal cases.Footnote ⁷

The second caveat is that digital and machine evidence is a wide-ranging and definitionally vague concept. Erin Murphy’s Chapter 9 in this volume offers a helpful taxonomy of such evidence that explains its various uses and characteristics, which in turn determine how such evidence implicates the principles in Section II.Footnote ⁸ Electronic communications and social media, e.g., implicate authentication and access concerns, but not so much the need for equitable safety valves in automated decision-making. Likewise, biometric identifiers may raise more privacy concerns than use of social media posts as evidence. The key characteristics of digital evidence as cataloged by Murphy also affect which principles are implicated. For example, data created by a private person, and possessed by Facebook, might implicate the right to exculpatory information and the Stored Communications Act,Footnote ⁹ while resiliency or lack of data such as body-worn camera footage might require the state to adopt more stringent preservation and storage measures, and to allow defendants access to e-discovery tools. So long as the principles are followed when they apply, the delivery of justice can be enhanced rather than jeopardized by digital and machine proof.

The third caveat is that this chapter does not write on a blank slate in setting forth principles to govern the use of technology in rendering justice. A host of disciplines and governing bodies have adopted principles for “ethical or responsible” use of AI, from the US Department of Defense to the Alan Turing Institute to the Council of Europe. Recent meta-studies of these various sets of principles have identified recurring values, such as beneficence, autonomy, justice, explainability, transparency, fairness, responsibility, privacy, expert oversight, stakeholder-driven legitimacy, and “values-driven determinism.”Footnote ¹⁰ More specifically, many countries already have a detailed legal framework to govern criminal procedure. In the United States, e.g., criminal defendants already have a constitutional right to compulsory process, to present a defense, to be confronted with the witnesses against them, to a verdict by a human jury, and to access to experts where necessary to a defense. But these rights were established at a time when cases largely depended on human witnesses rather than machines. The challenge here is not so much to convince nations in the abstract to allow a right to contest automated decision-making, but to explain how existing rights, such as the right of confrontation or right to pre-trial disclosure of the bases of expert testimony, might apply to this new type of evidence.

II The Principles

• Principle I: The digital and machine evidence used as proof in criminal proceedings should be subject to front-end development and testing safeguards designed to minimize error and bias.

• Principle I(a): Jurisdictions should acknowledge the heightened need for front-end safeguards with respect to digital and machine evidence, which cannot easily be scrutinized through case-specific, in-trial procedures.

To understand why the use of digital and machine evidence merits special front-end development and testing safeguards that do not apply to all types of evidence, jurisdictions should acknowledge that the current real-time trial safeguards built for human witnesses, such as cross-examination, are not as helpful for machine-generated proof.

A critical goal of any criminal trial is to ensure verdict accuracy by minimizing the chance of the fact-finder drawing the wrong inferences from the evidence presented. There are several different levers a system could use to combat inferential error by a jury. First, the system could exclude unreliable evidence so that the jury never hears it. Second, the system could implement front-end design and production safeguards to ensure that evidence is as reliable as it can be when admitted, or that critical contextual information about its probative value is developed and disclosed when the fact-finder hears it. Third, the system could allow parties themselves to explore and impeach, or attack the credibility/reliability of the evidence. Fourth, the system could adopt proof standards that limit the fact-finder’s ability to render a verdict absent a proof threshold such as beyond a reasonable doubt, or type or quantum of evidence.

For better or worse, the American system of evidence pursues accuracy almost entirely through trial and back-end safeguards, the third and fourth levers described above. Although the United States still clings to the rule excluding hearsay, understood as out-of-court statements offered for their truth, that rule has numerous exceptions. And while US jurisdictions used to have stringent competence requirements for witnesses, these have given way to the ability to impeach witnesses once they testify or once their hearsay is admitted.Footnote ¹¹ The parties conduct such impeachment through cross-examination, physical confrontation, and admission of extrinsic evidence such as a witness’s prior convictions or inconsistent statements. In addition, the United States has back-end proof standards to correct for unreliable testimony, such as corroboration requirements for accomplice testimony and confessions. The US system has a similarly lenient admission standard with regard to physical evidence, requiring only minimal proof that an item such as a document or object is what the proponent says it is.Footnote ¹²

Nonetheless, there are particular types of witness testimony that do require more front-end safeguards, ones that could work well for digital and machine evidence too. One example is eyewitness identifications. If an identification is conducted under unnecessarily suggestive circumstances, a US trial court, as a matter of constitutional due process, must conduct a hearing to determine whether the identification is sufficiently reliable to be admitted against the defendant at trial.Footnote ¹³ Moreover, some lower US courts subject identification testimony to limits or cautionary instructions at trial, unless certain procedures were used during the identification, to minimize the risk of suggestivity.Footnote ¹⁴ Likewise, expert testimony is subjected to enhanced reliability requirements that question whether the method has been tested, has a known error rate, has governing protocols, and has been subject to peer review.Footnote ¹⁵ To a lesser extent, confession evidence is also subject to more stringent front-end safeguards, such as the requirement in some jurisdictions that stationhouse confessions be videotaped.Footnote ¹⁶

The focus on front-end safeguards in these specific realms is not a coincidence. Rather, it stems from the fact that the problems with such testimony are largely cognitive, subconscious, or recurring, rather than a matter of one-off insincerity, and therefore not meaningfully scrutinized solely through cross-examination and other real-time impeachment methods.Footnote ¹⁷ These categories of testimony bear some of the same process-like characteristics that make digital and machine evidence difficult to scrutinize through cross-examination alone.

Even more so than these particular types of human testimony, digital and machine evidence bear characteristics that call for robust front-end development and testing safeguards before it gets to the courtroom. First, the programming of the algorithms that drive the outputs of many of the categories of proof discussed by Erin Murphy, including location trackers, smart tools, and analytical software tools, does not necessarily change from case to case.Footnote ¹⁸ Repeatedly-used software can be subject to testing to determine its accuracy under various conditions. Second, unlike eyewitnesses and confessions, where the declarant in some cases might offer significant further context through testimony, little further context can be gleaned from in-court scrutiny of any of the categories of proof Murphy describes.Footnote ¹⁹ To be sure, a programmer or inputter could take the stand and explain some aspects of a machine’s output in broad strokes. But the case-specific “raw data,” “measurement data,” or “evaluative data”Footnote ²⁰ of the machine is ultimately the product of the operation of the machine and its algorithms, not the programmer’s own mental processes, and it is the machine’s and algorithm’s operation that must also be scrutinized. In short, the accoutrements of courtroom adversarialism, such as live cross-examination, are hardly the “greatest legal engine ever invented for the discovery of truth”Footnote ²¹ of the conveyances of machines.

• Principle I(b): Jurisdictions should implement and enforce, through admissibility requirements, certain minimal development and testing procedures for digital and machine evidence.

Several development and testing safeguards should be implemented for any software-driven system whose results are introduced in criminal proceedings. The first is robust, independent stress testing of the software. Such standards are available,Footnote ²² but are typically not applied, at least in the United States, to software created for litigation. For example, a software expert reviewing the code of the Alcotest 7110, a breath-alcohol machine used in several US states, found that it would not pass industry standards. He documented 19,500 errors, nine of which he believed “could ultimately [a]ffect the breath alcohol reading.”Footnote ²³ A reviewing court held that such errors did not merit excluding the reading, in part because the expert could not say with “reasonable certainty” that the errors caused a false reading in the case at hand,Footnote ²⁴ but the court did require modifications of the program for future use.Footnote ²⁵ In addition, Nathaniel Adams, a computer scientist and expert in numerous criminal cases in the United States, has advocated for forensic algorithms to be subject to the industry-standard testing standards of the Institute of Electrical and Electronic Engineers (IEEE).Footnote ²⁶ Adams notes that STRMix, one of the two primary probabilistic genotyping programs used in the United States, had not been tested by a financially independent entity,Footnote ²⁷ and the program’s creators have disclosed more than one episode of miscodes potentially affecting match statistics, thus far, in ways that would underestimate but not overestimate a match probability.Footnote ²⁸ Professor Adams’ work helped to inspire a recent bill in the US Congress, the Justice in Forensic Algorithms Act of 2021, which would subject machine-generated proof in criminal cases to more rigorous testing, along with pre-trial disclosure requirements, defense access, and the removal of trade secret privilege from proprietary code.Footnote ²⁹ And exclusion aside, a rigorous software testing requirement reduces the chance of misleading or false machine conveyances presented at trial.

Jurisdictions should also enact mandatory testing and operation protocols for machine tools used to generate evidence of guilt or innocence, along the lines currently used for blood-alcohol breath-testing equipment.Footnote ³⁰ Such requirements need not be a condition of admission; in the breath-alcohol context, the failure to adhere to protocols goes to weight, not admissibility.Footnote ³¹ Even so, the lack of validation studies showing an algorithm’s accuracy under circumstances relevant to the case at hand should, in some cases, be a barrier to admissibility. Jurisdictions should subject the conclusions of machine experts to validity requirements at the admissibility stage, similar to those imposed on experts at trial. Currently, the Daubert and Frye reliability/general acceptance requirements apply only to human experts; if the prosecution introduces machine-generated proof without a human interlocutor, the proof is subject only to general authentication and relevance requirements.Footnote ³²

Requiring the proponent to show that the algorithm is fit for purpose through developmental and internal validation before offering its results is key not merely for algorithms created for law enforcement but for algorithms created for commercial purposes as well. For example, while Google Earth results have been admitted as evidence of guilt with no legal scrutiny of their reliability,Footnote ³³ scientists have conducted studies to determine its error rate with regard to various uses.Footnote ³⁴ While error is inevitable in any human or machine method, this type of study should be a condition of admitting algorithmic proof.Footnote ³⁵

Such testing need not necessarily require public disclosure of source code or other levels of transparency that could jeopardize intellectual property interests. Instead, testing algorithms for forensic use could be done in a manner similar to testing of potentially patentable pharmaceuticals by the US Food and Drug Administration.Footnote ³⁶ Others have made the point that scrutiny by “entrusted intermediate parties,” behind closed doors, would avoid any financial harm to developers.Footnote ³⁷ Of course, for algorithms that are open source, such concerns would be lessened.

One limit on validation studies as a guarantor of algorithmic accuracy is that most studies do not speak to whether an algorithm’s reported score or statistic, along a range, is accurate. Studies might show that a software program boasts a low false positive rate in terms of falsely labeling a non-contributor as a contributor to a DNA mixture, but not whether its reported likelihood ratio might be off by a factor of ten. As two DNA statistics experts explain, there is no “ground truth” against which to measure such statistics:

But systems are not helpless in testing the accuracy of algorithm-generated credit scores or match statistics. Rather, such results must be scrutinized using other methodologies, such as more complex studies that go beyond simply determining false positive rates, stress testing of software, examination of source code by independent experts, and assessment of whether various inputs, such as assumptions about the values of key variables, are appropriate.

• Principle I(c): Jurisdictions should explicitly define what is meant by algorithmic fairness for purposes of testing for, and guarding against, bias.

Algorithms should also be tested for bias. The importance of avoiding racial and other bias in algorithmic decision-making is perhaps obvious, given that fairness is an explicitly stated value in nearly all promulgated AI standards in the meta-studies referenced in the introduction to this chapter. In addition, racial, gender, and other kinds of bias might trigger legal violations as well as ethical or policy concerns. To be sure, the Equal Protection Clause of the Fourteenth Amendment to the US Constitution guards only against state action that intentionally treats people differently because of a protected status, but if an algorithm simply has a disparate impact on a group, it will likely not be viewed as an equal protection violation. However, biased algorithms used in jury selection could violate the requirement that petit juries be drawn from a fair cross section of the population, and biased algorithms used to prove dangerousness or guilt at trial could violate statutory anti-discrimination laws or reliability-based admissibility standards.

In one highly publicized example of algorithmic bias from the United States, Pro Publica studied Northpointe’s post-trial risk assessment tool Correctional Offender Management Profiling for Alternative Sanctions (COMPAS) and determined that the false positive rates, i.e., rates of those labeled “dangerous,” but who did not reoffend, for Black subjects was much higher than for White subjects.Footnote ³⁹ At the same time, however, other studies, including by Northpointe itself, noted that the algorithm is, in fact, racially non-biased if the metric is whether race has any predictive value in the model in determining dangerousness.Footnote ⁴⁰ As Northpointe notes, Black and White subjects with the same risk score present the same risk of reoffending under the model.Footnote ⁴¹ The upshot was not that Pro Publica was wrong in noting the differences in false positive rates; it was that Pro Publica judged the algorithm’s racial bias by only one particular measure.

The COMPAS example highlights the problems of testing algorithms for fairness without defining terms. As others have explained, it is impossible to have both equal false positive rates and predictive parity where two groups have different base rates.Footnote ⁴² So, in determining whether the algorithm is biased, one needs to decide which measure is the more salient indicator of the type of bias the system should care about. Several commentators have noted possible differences in definitions of algorithmic fairness as well.Footnote ⁴³ Deborah Hellman argues that predictive parity alone is an ill-suited measure of algorithmic fairness because it relates only to beliefs, not outcomes.Footnote ⁴⁴ In Hellman’s view, a disparate false positive rate between groups is highly relevant to proving, though not dispositive of, normatively troubling unfairness.Footnote ⁴⁵ While not all jurisdictions will agree with Hellman’s take, the point is that algorithm designers should be aware of different conceptions of fairness, be deliberate in choosing a metric, and ensure that algorithms in criminal proceedings are fair under that metric. Jurisdictions could require what Osagie Obasogie has termed “racial impact statements” in the administrative law context,Footnote ⁴⁶ to determine the effect of a shift in decision-making on racial groups. The Council of Europe has made a similar recommendation, calling on states to conduct “human rights impact assessments of AI applications” to assess “risks of bias/discrimination … with particular attention to the situation of minorities and vulnerable and disadvantaged groups.”Footnote ⁴⁷

Finally, in determining algorithmic fairness, decision-makers should judge algorithms not in a vacuum, but against existing human-driven decision-making processes. For example, court reporters have been known to mistakenly transcribe certain dialects, such as African American Vernacular English (AAVE), in ways that matter to fact-finding in criminal proceedings.Footnote ⁴⁸ If an AI system were to offer a lower, even if non-zero, error rate with regard to mistranscriptions of AAVE, the shift toward such systems, at least a temporary one subject to continued testing and oversight, might reduce, rather than exacerbate, bias.Footnote ⁴⁹

• Principle II: Before trial or other relevant proceeding, the parties should have meaningful and equitable access to digital and machine evidence material to the proceeding, including exculpatory technologies and data.

• Principle II(a): Pretrial disclosure requirements related to expert testimony should apply to digital and machine conveyances that, if asserted by a human, would be subject to such requirements.

Because digital and machine evidence cannot be cross-examined, parties cannot use the in-court trial process itself to discover the infirmities of algorithms or possible flaws in their results or opinions. As Edward Cheng and Alex Nunn have noted, enhanced pre-trial discovery must in part take the place of in-court discovery with regard to process-based evidence like machine conveyances.Footnote ⁵⁰ Such enhanced discovery already exists in the United States for human experts, precisely because in-court examination alone is not a meaningful way for parties to understand and prepare to rebut expert testimony. Specifically, parties in criminal cases are entitled by statute to certain information with regard to expert witnesses, including notice of the basis and content of the expert’s testimony and the expert’s qualifications.Footnote ⁵¹ Disclosure requirements in civil trials are even more onerous, requiring experts to prepare written reports that include the facts or data relied on.Footnote ⁵² Moreover, proponents of expert testimony must not discourage experts from speaking with the opposing party,Footnote ⁵³ and in criminal trials, proponents must also disclose certain prior statements, or Jencks material, of witnesses after they testify.Footnote ⁵⁴ These requirements also facilitate parties’ ability to consult with their own experts to review the opposing party’s evidence or proffered expert testimony.

Using existing rules for human experts as a guide, jurisdictions should require that parties be given access to the following:

1. (1) The evidence and algorithms themselves, sufficient to allow meaningful testing of their assumptions and running the program with different inputs. One probabilistic genotyping software company, TrueAllele, offers defendants access to its program, with certain restrictions, albeit only for a limited time and without the source code.Footnote ⁵⁵ This sort of “black box tinkering” not only allows users to “confront” the code “with different scenarios,” thus “reveal[ing] the blueprints of its decision-making process,”Footnote ⁵⁶ but also approximates the posing of a hypothetical to a human expert. Indeed, the ability to tinker might be just as important as access to source code; data science scholars have written about the limits of transparency and the superior promise of reverse engineering in understanding how inputs relate to outputs.Footnote ⁵⁷ Along these lines, Jennifer Mnookin has argued that a condition for admissibility of computer simulations should be that “their key evidence-based inputs are modifiable,” allowing the opposing party to “test the robustness of the simulation by altering the factual assumptions on which it was built and seeing how changing these inputs affects the outputs.”Footnote ⁵⁸

2. (2) The training or software necessary to use or test the program. In the United States, criminal defendants have reported that certain trainings are off limits to non-law-enforcement; e.g., for using the Cellebrite program to extract digital evidence from a cell phone, or for using DNA genotyping software. Moreover, only certain defense experts are able to buy the software for their own use, and some academic researchers have been effectively denied research licenses to study proprietary forensic software. Instead, the defense and academic communities should presumptively be given a license to access to all software used by the government in generating evidence of guilt, to facilitate independent validity testing.

3. (3) A meaningful account of the assumptions underlying the machine’s results or opinion, as well as the source code and prior output of software, where necessary to a meaningful understanding of those assumptions. Human experts can be extensively questioned both before and during trial, offering a way for parties to understand and refute their methods and conclusions. Digital and machine evidence cannot be questioned in the same way, but proponents should be required to disclose the same type of information about methods and conclusions that a machine expert witness would offer, if it could talk. Likewise, Article 15(1)(h) of General Data Protection Regulation (GDPR)Footnote ⁵⁹ gives a data subject the right to know of any automated decision-making to which he is subject, and if so, the right to “meaningful information about the logic involved.” While the GDPR may apply only to private parties rather than criminal prosecutions, the subject’s dignitary interest in understanding the machine’s logic would presumably be even greater in the criminal realm.

   In particular, where disclosure of source code is necessary to meaningful scrutiny of the accuracy of machine results,Footnote ⁶⁰ the proponent must allow access. As discussed in Principle I, source code might be important in particular to scrutinize scores or match statistics, where existing studies reveal only false positive rates. A jurisdiction should also require disclosure of prior output of the machine, covering the same subject matter as the machine results being admitted.Footnote ⁶¹ For human witnesses, such prior statements must be disclosed in many US jurisdictions to facilitate scrutiny of witness claims and impeachment by inconsistency. For machines, parties should have to disclose, e.g., the results of all prior runs of DNA software on a sample, all potentially matching reference fingerprints reported by a database using a latent print from a crime scene,Footnote ⁶² or calibration data from breath-alcohol machines.Footnote ⁶³

4. (4) Access to training data. Defendants and their experts should have access to underlying data used by the machine or algorithm in producing its results. In countries with inquisitorial as compared to adversarial systems, defendants should have access to “any data that is at the disposal of the court-appointed expert.”Footnote ⁶⁴ For example, for a machine-learning model labeling a defendant a “sexual psychopath” for purposes of a civil detention statute, the defendant should have access to the training dataset. Issues of privacy, i.e., the privacy of those in the dataset, have arisen, but are not insurmountable.Footnote ⁶⁵

To be sure, access alone does not guarantee that defendants will understand what they are given. But access is a necessary condition to allowing defendants to consult with experts who can meaningfully study the algorithms’ performance and limits.

• Principle II(b): Jurisdictions should not allow claims of trade secret privilege or statutory privacy interests to interfere with a criminal defendant’s meaningful access to digital and machine evidence, including exculpatory technologies and data.

While creators of proprietary algorithms routinely argue that source code is a trade secret,Footnote ⁶⁶ this argument should not shield code from discovery in a criminal case, where the code is material to the proceedings.Footnote ⁶⁷ Of course, if proprietors can claim substantive intellectual property rights in their algorithms, those rights are still enforceable through licensing fees and civil lawsuits.

Likewise, criminal defendants should have meaningful access to exculpatory digital and machine evidence, including the ability to subpoena witnesses who can produce such evidence in criminal proceedings where such evidence is material. Rebecca Wexler has explored the asymmetries inherent in US statutes such as the Stored Communications Act, which shields electronically stored communications from disclosure and has an exception for “law enforcement,” but not for criminal defendants, however material the communications might be to establishing innocence. Such asymmetries are inconsistent not only with basic adversarial fairness, but arguably also with the Sixth Amendment compulsory process.Footnote ⁶⁸

• Principle II(c): Jurisdictions should apply a presumption in favor of open-source technologies in criminal justice.

In the United States, the public has a constitutional right of access to criminal proceedings.Footnote ⁶⁹ With regard to human witnesses, the public can hear the witnesses testify and determine the strength and legitimacy of the state’s case. The public should likewise be recognized as a stakeholder in the development and use of digital and machine evidence in criminal proceedings. The Council of Europe’s guidelines for use of AI in criminal justice embrace this concept, requiring Member States to “meaningfully consult the public, including civil society organizations and community representatives, before introducing AI applications.”Footnote ⁷⁰

The most direct way to ensure public scrutiny of such evidence would be through open-source software. Scholars have discussed the benefits of open-source software in terms of facilitating “crowdsourcing”Footnote ⁷¹ and “ruthless public scrutiny”Footnote ⁷² as means of testing models and algorithms for hidden biases and errors. Others have gone further, arguing that software should be open source whenever used in public law.Footnote ⁷³ Public models would have the benefit of being “transparent” and “continuously updated, with both the assumptions and the conclusions clear for all to see.”Footnote ⁷⁴ States could encourage adoption of open-source software through drastic means, excluding output from adjudication, or more modest means, such as offering monetary incentives or prizes for development of open source replacements.

• Principle II(d): Jurisdictions should make investigative technologies equally available to criminal defendants for potential exculpatory purposes, regardless of whether the state used the technology in a given case.

As Erin Murphy notes in Chapter 9 of this volume, defendants have two compelling needs with regard to digital and machine evidence: a meaningful chance to attack the government’s proof, and a meaningful chance to discover and present “supportive defense evidence.”Footnote ⁷⁵ Just as both defendants and prosecutors have the ability to interview and subpoena witnesses, defendants should have an equal ability to wield new technologies that are paid for by the state when prosecutors seek to use them. If a defendant is accused of a crime based on what he believes to be a human analyst’s erroneous interpretation of a complex DNA mixture, the defendant should be given the ability to use a probabilistic genotyping program, like TrueAllele, to attack these results. Of course, this access would be costly, and might reasonably be denied in cases where it bears no relevance to the defense, as determined ex parte by a judge. But if defendants have a due process right to access to defense experts where critical to their defense,Footnote ⁷⁶ they should have such a right of access to exculpatory algorithms as well.

• Principle III: Criminal defendants should have a meaningful right of contestation with respect to digital and machine evidence including, at a minimum, a right to be heard on development and testing procedures and meaningful access to experts.

Much has been written about a right of contestation by data subjects with regard to results of automated decision-making processes.Footnote ⁷⁷ In the US criminal context, defendants already enjoy, at least in theory, a right to present a defense, encompassing a cluster of rights, including the right to be confronted by the witnesses against them, to testify in their own defense, and to subpoena and present witnesses in their favor. In the United States, a criminal defendant’s right of contestation essentially encompasses everything already discussed with regard to access to the state’s evidence, as well as to some exculpatory electronic communications. In addition, the US Supreme Court has held that the right to present a defense exists even where the government presents scientific evidence of guilt that a trial judge might deem definitive. The fact that an algorithm offers compelling evidence of guilt cannot preclude a defendant from offering a defense case.Footnote ⁷⁸

In addition to pre-trial access to the evidence itself, and information about its assumptions and processes, other rights that are key to a meaningful ability to contest the results of digital and machine evidence include the ability to consult experts where necessary. David Sklansky has argued that a right to such expert testimony, and not merely in-court cross-examination, should be deemed a central part of the Sixth Amendment right of confrontation.Footnote ⁷⁹

The importance of a right of contestation in the algorithmic design process might be less obvious. But in a changing world in which machine evidence is not easily scrutinized at the trial itself, the adversarialism upon which common law systems are built might need to partially shift from the trial stage to the design and development stage. Carl DiSalvo has coined the term “adversarial design”Footnote ⁸⁰ to refer to design processes that incorporate political contestation among different stakeholders. While adversarial design would not be a case-specific process, it could still involve representatives from the defense community. Others have suggested appointing a “defender general” in each jurisdictionFootnote ⁸¹ who could inject adversarial scrutiny into various recurring criminal justice issues at the front end. Perhaps such a representative could oversee defense involvement in the design, testing, and validation of algorithms. This process would supplement, not supplant, case-specific machine access and discovery.

The right of contestation with regard to sophisticated AI systems, the methods of which may well never be meaningfully understood by the parties, might also need to incorporate a right to delegated contestation, in the form of the right to another machine’s scrutiny of the results. Other scholars have noted the possibility of “reversible” algorithms that would audit themselves or each other,Footnote ⁸² or have suggested that one machine opinion alone should be deemed legally insufficient for a conviction, in the absence of corroboration from a second expert system.Footnote ⁸³

At the trial itself, the right of contestation should first include the right to argue for exclusion of the evidence on reliability (Frye/Daubert) and/or authenticity grounds. In the US federal system, proponents of digital and machine evidence must present sufficient evidence to persuade the fact-finder that the evidence is what the proponent says it is, e.g., that an email is from a particular sender.Footnote ⁸⁴ In China, courts have used blockchain technology to facilitate authentication of electronically stored information.Footnote ⁸⁵ Jurisdictions’ authenticity method might reasonably change as the ability for malfeasors to falsify evidence changes in the future. Likewise, litigants should have the right to insist on exclusion of machine evidence if inputs are not proven accurate. For example, in the United Kingdom, a “representation” that is made “other than by a person” but that “depends for its accuracy on information supplied (directly or indirectly) by a person” is not admissible in criminal cases without proof that the “information was accurate.”Footnote ⁸⁶ In some cases, this showing will require testimony from the inputter.Footnote ⁸⁷

• Principle IV: Criminal defendants should have a right to a factfinding process that is epistemically competent but that retains a human in the loop, so that significant decisions affecting their liberty are not entirely automated.

• Principle IV(a): While parts of the criminal process can be automated, human safety valves must be incorporated into the process to ensure a role for equity, mercy, and human moral judgment.

Both substantive criminal law and criminal procedure in the United States have become more mechanical over the past few decades, from mandatory arrest laws, to sentencing guidelines, to laws criminalizing certain quantities of alcohol in drivers’ blood.Footnote ⁸⁸ The more mechanical that the system becomes on the front end via, e.g., mandatory arrest, prosecution, liability rules, and sentencing, the more that safety valves such as prosecutorial, fact-finder, and sentencing discretion become critical to avoid inequities, i.e., results that are legal but unjust.Footnote ⁸⁹ Moreover, mechanical regimes reduce the possibility of mercy, understood to mean leniency or grace, beyond what a defendant justly deserves. While mercy may be irrational, it is a pedigreed and “important moral virtue” that shows compassion and a shared humanity.Footnote ⁹⁰

As digital and machine evidence accelerate the mechanization of justice, jurisdictions should ensure that human actors are still able to exercise equity and mercy at the charging, guilt, and/or punishment stages of criminal justice. Not only are humans needed to ensure that laws are not applied mechanically. They are needed because they are literally human – they bring a human component to moral judgment that is necessary, if not for dignity, then at least for public legitimacyFootnote ⁹¹ and, in turn, for enforcement of criminal law.Footnote ⁹² In the United States, scholars have written since the 1970s of the illegitimacy of verdicts based solely on “naked statistical evidence,” based on personhood concerns.Footnote ⁹³ Moreover, humans add to the fact-finding process as well, rendering AI systems fairer without having to make such systems less accurate through simplification.Footnote ⁹⁴ Corroborating these observations, recent AI guidelines and data privacy laws reflect the public’s desire to keep humans in the loop with regard to automated decision-making, from the Council of Europe’s call to “ensure that the introduction, operation and use of AI applications can be subject to effective judicial review,”Footnote ⁹⁵ to the EU Directive prohibiting processes that produce an “adverse legal effect” on a subject “based solely on automated processing,” without appropriate “safeguards for the rights and freedoms of the data subject, at least the right to obtain human intervention on the part of the controller.”Footnote ⁹⁶

More concretely, criminal liability should not be based solely on an automated decision. Red light cameras are the closest the United States has come to fully automated liability, but thus far, such violations end only in a mailed traffic ticket rather than a criminal record. Moreover, in jurisdictions with juries, the power of jury nullification should continue undisturbed. It may well be that jurors’ ability to decide historical fact, e.g., “was the light red?”, could be curtailed, so long as their ability to decide evaluative data, e.g., “did the defendant drive ‘recklessly’?”, is preserved.Footnote ⁹⁷ Indeed, some historical fact-finding might be removed from lay jurors, if they lack the “epistemic competence” to assess the evidence’s probative value.Footnote ⁹⁸ Jurisdictions could still ensure that humans remain in the loop by disallowing machine experts from giving dispositive testimony on ultimate questions of fact,Footnote ⁹⁹ prohibiting detention decisions based solely on a risk assessment tool’s score, and requiring a human expert potentially liable for injustices caused by inaccuracies to vouch for the results of any machine expert, before introducing results in a criminal proceeding.

• Principle IV(b): Jurisdictions should ensure against automation complacency by developing effective human–machine interaction tools.

Keeping a human in the loop would be useless if that human deferred blindly to a machine. For example, if sentencing judges merely rubber-stamped scores of risk assessment tools, there would be little reason to ensure that judges remain in the loop.Footnote ¹⁰⁰ Likewise, if left to their own devices, juries might irrationally defer to the apparent objectivity of machines.Footnote ¹⁰¹ A human in the loop requirement should entail the development of tools to guard against automation complacency. One underused tool in this regard is jury instructions. For example, where photographs are admitted as silent witnesses, the jury hears little about lens, angle, speed, placement, cameraperson bias, or other variables that might lead it to draw a false inference from the evidence. The jury should be educated about the effect of these variables on the image they are assessing.Footnote ¹⁰² Ultimately, jurisdictions should draw from the fields of human factors engineering, and human–computer interaction and collaboration, in designing ways to ensure a systems approach that keeps humans in the loop while leveraging the advantages of AI.

• Principle IV(c): Jurisdictions should establish a formal means for stakeholders to challenge uses of digital and machine evidence that are fundamentally inconsistent with principles of human-delivered justice.

Keeping a human in the loop also necessarily means taking steps to ensure against inappropriate uses of AI that threaten softer systemic values like dignity. For example, certain machines might be condemned as inherently dehumanizing, such as the penile plethysmographFootnote ¹⁰³ or deception detection.Footnote ¹⁰⁴ Just as some modes of obtaining evidence are rejected as violating substantive due process, such as forcibly pumping a suspect’s stomach to find evidence of drug use,Footnote ¹⁰⁵ modes of determining guilt should be rejected if the public views them as inhumane. Other jurisdictions might decide that the “right to explanation” is so critical to public legitimacy that overly complex AI systems must be abandoned in criminal adjudication, even if such systems promise more accuracy.Footnote ¹⁰⁶ Whatever approach jurisdictions adopt regarding these issues, they should resolve such issues first, and only then look for available technological enhancements of proof, rather than vice versa. Numerous scholars have written about the seduction of quantification and measurement,Footnote ¹⁰⁷ and the Council of Europe expressly included in its guidelines for the use of AI in criminal justice that Member States should “ensure that AI serves overall policy goals, and that policy goals are not limited to areas where AI can be applied.”Footnote ¹⁰⁸

III Conclusion

The principles for governing digital and machine evidence articulated in this chapter attempt to move beyond the adversarial/inquisitorial divide, and incorporate the thoughtful recent work of so many scholars, policymakers, and stakeholders worldwide in promulgating guidelines for the ethical and benevolent use of AI in decision-making affecting peoples’ lives. Applied to a common law adversarial criminal system such as that in the United States, these principles may manifest in existing statutory and constitutional rights, albeit in new ways. Applied to other nations’ systems, these principles will manifest differently, perhaps because such systems already recognize the need for “out of court evidence gathering”Footnote ¹⁰⁹ to ensure meaningful evaluation of complex evidence. On the other hand, as Sabine Gless has suggested, continental systems might find that party-driven examinations have an underappreciated role to play in ensuring reliability of machine evidence.Footnote ¹¹⁰

As AI becomes more sophisticated, one key goal for all justice systems will be to ensure that AI is not merely given an objective to accomplish, such as “determine whether this witness is lying” or “determine if this person contributed to this DNA mixture,” but is programmed to continually look to humans to express and update their preferences. If the former occurs, AI will preserve itself at all costs, and may engage in behavior antithetical to human values, to get there.Footnote ¹¹¹ Only if machines are taught to continually seek feedback can AI remain benevolent. We cannot simply program machines to achieve the goals of criminal justice – public safety, social cohesion, equity, the punishment of the morally deserving, and the vindication of victims. We will have to ensure that humans have the last word on what justice means and how to achieve it.

I Drowsy at the Wheel?

In 2016, the Swiss media reported a collision involving a sports car and a motor scooter that resulted in serious injuries to the rider of the scooter.Footnote ¹ Charges were brought against the car driver on the grounds that he was unfit to operate his vehicle. Driving a motor vehicle while unfit to do so is a crime pursuant to the Swiss Traffic CodeFootnote ² and one for which negligence suffices to establish culpability.Footnote ³ Although the accused denied consciously noticing that he was too tired to drive, prosecuting authorities claimed that he should have been aware of his unfitness, as the car’s driving assistants had activated alerts several times during the journey.Footnote ⁴ Media coverage of the event did not report whether or how the accused defended himself against these alerts.

We refer to these alerts as “evaluative data” because they combine data with some form of robot evaluation. We argue that acknowledging this novel category of evidence is necessary because driving assistants and other complex information technology (IT) systems outfitted with artificial intelligence (AI) do more than simply employ sensors that engage in relatively straightforward tasks such as measuring the distance between the vehicle and lane markings. Driving assistants also evaluate data associated with indicators that they deem potential signs of fatigue, such as erratic steering movements or a human driver’s drooping eyelids. They interpret this data and decide autonomously whether to alert the driver to drowsiness. When introduced into a criminal proceeding, this evaluative data can be referred to as a kind of robot testimony because it conveys an assessment made by a robot based on its autonomous observation.

This chapter aims to alleviate deficits in current understandings of the contributions such testimony can make to truth-finding in criminal proceedings. It explains the need to vet robot testimony and offers a taxonomy to assist in this process. In addition to a taxonomy of robot testimony, the chapter proposes a standardized approach to the presentation and evaluation of robot testimony in the fact-finding portion of criminal trials. Analysis focuses on a currently hypothetical criminal case, in which a drowsiness alert is proffered as evidence in a civil law jurisdiction such as Switzerland or Germany.

The chapter first introduces robot testimony and outlines the difficulties it poses when offered as evidence in criminal proceedings (Section II). Second, we propose a taxonomy for and a methodical way of using the results of a robot’s assessment of human conduct (Section III). Based on traditional forensic science, robot testimony must first be grounded in the analog world, using a standardized approach to accessibility, traceability, and reproducibility. Then, with the help of forensic experts and the established concepts of source level and activity level, the evidence can be assessed on the offense level by courtroom actors, who are often digital laypersons (Section IV). As robot witnesses cannot be called to the stand and have their assessments subjected to cross-examination, the vetting of robot testimony in the courtroom poses a number of significant challenges. We suggest some ways to meet these challenges in Section V. In our conclusion, we call for legislatures to address the lacunae regarding the use of robot testimony in criminal proceedings, and we consider how criminal forensics might catch up with the overall debate on the trustworthiness of robots, an issue at the core of the current European debate regarding AI systems in general (Section VI). An outline of questions that stakeholders might want to ask when vetting robot testimony via an expert is presented in the Appendix.

II Introducing Robot Testimony

A core problem raised when defending oneself against a robot’s evaluation of one’s conduct, not to mention one’s condition, is the overwhelming complexity of such an assessment. A car driver, as a rule, does not have the tools necessary to challenge the mosaic of components upon which the robot’s evaluation is based, including the requisite knowledge of raw data, insights into source code, or the capacity for reverse engineering; this is certainly the case in a driving assistant’s assessment that the human driver is drowsy.Footnote ⁵

III A Taxonomy for the Use of Robot Testimony

Robot testimony is a potentially useful addition to the evidentiary process, but only if its meaning for a case can be communicated to the fact-finder in a comprehensible way. In order to facilitate this communication, we propose a taxonomy of robot testimony. The taxonomy distinguishes between three types of machine-readable data, beginning with the least complex form and ending with the most complex form. We also suggest how the taxonomy can be used in practice, by differentiating circumstantial information, which refers to the context in which the data is found (Section III.B), from information content, the forensically relevant information that the expert can deduce from the properly identified data (Section III.C).

IV A Standardized Approach to Interpreting Robot Testimony

Having established a tri-part taxonomy for the use of robot testimony, we now suggest a standardized approach regarding its interpretation in a court of law. Legal actors can draw on existing conceptsFootnote ³⁸ in concert with the new taxonomy proposed here, but the traditional approach will have to be modified so as to accommodate the special needs of assessing evaluative data for fact-finding in a criminal case. A sort of tool kit is needed to test whether a robot generates trustworthy evidence. In our hypothetical, the question can be framed as whether a drowsiness detection system reliably detects reasonable parameters related to a human driver’s fitness to operate a vehicle.

In principle, the general rules for obtaining and presenting evidence in a criminal case apply to robot testimony. In our hypothetical, the vehicle involved in the accident will be seized. Subsequently, the search for analog evidence will follow existing provisions of the applicable code of criminal procedure regarding the admissibility and reliability of potential evidence. As far as digital evidence is concerned, various modifications stemming from the particularities of using bits and bytes for fact-finding will apply,Footnote ³⁹ and specific risks of error will have to be addressed. For known problems, such as the loss of information during the transmission of data, solutions may already be at hand.Footnote ⁴⁰ But new problems arise, including, e.g., the sheer volume of data that may be relevant if it becomes necessary to validate a specific alert issued by a vehicle’s drowsiness detection system. In such cases, it is essential for stakeholders to understand what is meant by accessibility (Section IV.A) and traceability (Section IV.B) of relevant data, as well as the reproducibility (Section IV.C) and interpretation (Section IV.D) of results provided by the expert.

V Unique Challenges Associated with Vetting Robot Testimony

The proposed standardized approach to proffering evaluative data as evidence in criminal proceedings illustrates the need for a sound methodology. It also simultaneously highlights the limits of the traditional approach with robot testimony. One of the parties may want to use an alert issued by a drowsiness detection system as evidence of a human driver’s unfitness to operate a vehicle, but forensic experts may not be able to offer sufficient insights to verify or refute the system’s evaluation. Crucial questions of admissibility or weight of the evidence are left unanswered when experts can attest only that the drowsiness detection system issued an alert before the accident occurred. If experts cannot retrieve sufficient data or sufficient circumstantial information, they may not be able to provide the fact-finder with the information necessary to assess the evidentiary value of the alert. The fact-finder cannot simply adopt the driving assistant’s evaluation, as doing so would fail to satisfy the judicial task of conclusively assessing evidence. The question as to the grounds upon which judges can disregard such evidence remains an open one.Footnote ⁵⁷

The problems raised in vetting robot testimony become even clearer when the defense’s ability to challenge the trustworthiness of observations and evaluations generated by a robot are compared to the alternatives available to check and question measurement data generated by traditional forensic tools. If, e.g., the defense wants to question the results of a radar gun in a speeding case, the relevant measurement data, i.e., the whole dataset of frequency values, calculated speed values, and the additional measurements performed by the radar gun, can be accessed. This information can reveal whether or not a series of measurements appears to be robust.Footnote ⁵⁸ Furthermore, if the defense wishes to cast doubt on an expert’s findings and develop another proposition to explain the results of the radar gun, the court could require law enforcement authorities to offer a second dataset based on an independent measurement method, e.g., a videotaping of the radar gun’s measurement and its environment. This would allow for independent verification and would make it possible to check for factors that may have distorted the measurements, such as truck trailers parked on the street or the surface reflections of buildings.Footnote ⁵⁹

In contrast, if the defense wishes to challenge robot testimony such as a drowsiness detection system’s alert, new and unresolved issues with regard to both facts and law may arise.Footnote ⁶⁰ As mentioned above, driving assistants are consumer gadgets designed to enhance road safety. They are neither approved nor certified forensic tools designed to generate evidence for criminal proceedings. It is currently left to the programmer of the driving assistance system or the manufacturer of the car to develop a robust machine learning process for the system that leads to the establishment of a threshold for issuing an alert and to determine what information to store for potential evaluation, ex ante, of the robot’s assessment. The decision-making power of the programmer or producer regarding the shaping of a smart product’s capacity to observe and record is limited only if there are regulations that require the storage of particular data in a particular form.

Parties challenging drowsiness alerts can try their luck by challenging different kinds of data. Measurement data, which generally describes physical facts in a transparent way, appears to be the most objective information, and the corresponding information content seems relatively safe from legal attack. In contrast, evaluative data, including records of decisions taken or interventions launched by a robot, appears to be much closer to the contested legal questions and thus a more appropriate target for legal challenge. Counsel could argue that the dataset containing information about the incident does not allow for robust testing of alternative scenarios, or that no validation exists for the thresholds for issuing an alert set by machine learning, thereby rendering an expert’s probability ratios worthless, or that someone might have tampered with the data. These arguments show that in order to do their jobs properly, lawyers must be capable of understanding not only how data is generated, retrieved, and accessed, but also how evidence can be evaluated, interpreted, verified, and vetted with regard to its information content and to the integrity of the data.

VI A Look to the Future

I Introduction

In courtrooms across the world, criminal cases are no longer proved only through traditional means such as eyewitnesses, confessions, or rudimentary physical evidence like the proverbial smoking gun. Instead, prosecutors increasingly harness technologies, including those developed and used for purposes other than law enforcement, to generate criminal evidence.Footnote ¹

This kind of digital data may take different forms, including raw data, data that is produced by a machine without any processing; measurement data, data that is produced by a machine after rudimentary calculations; and evaluative data, data that is produced by a machine according to sophisticated algorithmic methods that cannot be reproduced manually.Footnote ² These distinctions are likewise evident in the array of consumer products that can now be tapped to produce evidence in a criminal case. A mobile phone can be used to track the user’s location via raw data in the form of a readout of which tower the cell phone “pinged,” via measurement data reflecting the triangulation of data towers accessed along a person’s route, or with evaluative data generated by a machine-learning algorithm to predict the precise location of a person, such as a specific shop in a shopping mall.Footnote ³ In all three forms, the use of such data presents new evidentiary challenges, although it is the evaluative data that raises the most issues as a result of both its precision and impenetrability.

As scholars begin to tackle the list of questions raised by these new forms of evidence, one critical perspective is often omitted: the view of the criminal defendant. Yet, just as digital evidence serves to prove the guilt of an accused, so too can it serve the equally important role of exculpating the innocent. As it stands now, law fails to adequately safeguard the rights of a criminal defendant to conduct digital investigations and present digital evidence. In a world increasingly reliant on technological forms of proof, the failure to afford full pre-trial rights of discovery and investigation to the defense fatally undermines the presumption of innocence and the basic precepts of due process.

Persons accused of crimes have two compelling needs with regard to digital evidence. First, criminal defendants must be granted the power to meaningfully attack the government’s digital proof, whether offered by the government to prove its affirmative case or to counter evidence tendered by the defense.Footnote ⁴ For example, the defense might challenge cell site location records that purport to show the defendant’s location at the scene of the crime. Or they might contest cell site records offered by the government to undermine a defense witness’s claim to have witnessed the incident. The defendant is attacking the government’s proffered digital evidence in both cases, but in the first variation, the attack responds to the government’s evidence in its case-in-chief, whereas the second variation responds to evidence proffered by the government to counter a defense claim or witness.

The defense’s use of digital evidence in this way differs from the second category, which might be called supportive defense evidence. A defendant must be able to access and introduce the defendant’s own digital proof, in order to support a defense theory or to attack the government’s non-digital evidence. For example, the defendant might use digital data to show that the defendant is innocent, to reinforce testimony offered by a defense witness, or to support a claim that another person in fact committed the offense. Classic examples of such use would be DNA evidence that proves there was another perpetrator, or surveillance footage that reveals the perpetrator had a distinguishing mark not shared by the defendant. A defendant might also use such evidence to bolster a legal claim. In the United States, e.g., the defendant might use digital proof to argue that evidence must be suppressed because it was obtained in violation of the Constitution.Footnote ⁵ Or a defendant might use digital evidence to attack the non-digital evidence in the government’s case, like a defendant who introduces the cell-site records that show that the government’s witness was not at the scene, or offers the victim’s social media posts to prove that the victim still possessed the property the defendant allegedly stole. What links these examples of supportive defense evidence is that the defense introduces digital proof of its own; it does not just attack the digital proof offered by the government.

In both cases – when the defense aims to attack government digital proof, or when it aims to introduce its own digital proof – the defendant cannot effectively mount a defense without access to and the ability to challenge complex forms of digital proof. Yet, in all too many jurisdictions, the legal system has embraced the government’s use of technological tools to inculpate a defendantFootnote ⁶ without reckoning with the equivalent needs of the accused.Footnote ⁷ Baseline principles such as those enshrined in the Fifth and Sixth Amendments to the US Constitution and Article 6 of the European Convention on Human Rights sketch broad rights, but how those rights are actually implemented, and the governing rules and statutes that embody those values, may vary dramatically.Footnote ⁸ In the United States, criminal defendants have few positive investigatory powers,Footnote ⁹ and are largely dependent on rules that mandate government disclosure of limited forms of evidence or the backstop of the constitutional rights of due process, confrontation, and compulsory process.Footnote ¹⁰

Even when the defense is entitled to certain information, existing legal tools may be inadequate to effectively obtain and utilize it. Criminal defendants must typically rely on either a court order or subpoena to obtain information from third parties, but both of those mechanisms are typically understood as intended for the purpose of presenting evidence at trial, not conducting pre-trial investigation.Footnote ¹¹ And even sympathetic courts struggle to determine whether and how much to grant requests. As one high court observed when addressing a defendant’s request for access to a Facebook post, “there is surprisingly little guidance in the case law and secondary literature with regard to the appropriate inquiry.”Footnote ¹²

Finally, generally applicable substantive laws may also thwart defense efforts to use technological evidence. For example, privacy statutes in the United States typically include law enforcement exceptions,Footnote ¹³ but as Rebecca Wexler has observed, those same statutes effectively “bar defense counsel from subpoenaing private entities for entire categories of sensitive information,” and in fact “[c]ourts have repeatedly interpreted [statutory] silence to categorically prohibit defense subpoenas.”Footnote ¹⁴

Without robust reconsideration of the rights necessary to empower defendants in each of these endeavors, the digitalization of evidence threatens to bring with it the demise of due process and accurate fact-finding. The first step in articulating these critical defensive rights, however, is to identify and classify the scope of such evidence and its pertinent features. Such analysis serves two purposes. First, it crystallizes the need for robust defense pre-trial rights, including rights to discovery, compelled process, and expert assistance, as well as substantive and procedural entitlements to confront such evidence and mount an effective defense at trial. Second, cataloging these technologies helps point the way toward a comprehensive framework for defense access and disclosure, one that can account for the many subtle variations and features involved in each technology – one that is wholly lacking now.

To facilitate deeper inquiry into the proper scope and extent of the criminal defendant’s interest in digital proof, this chapter presents a taxonomy of defensive use of technological evidence. Section II identifies and provides examples for seven categories of such data: location trackers, electronic communications and social media, historical search or cloud or vendor records, the “Internet of Things” and smart tools, surveillance cameras, biometric identifiers, and analytical software tools. Although the examples in this chapter are drawn primarily from legal cases in the United States, these technologies are currently in broad use around the world. Section III then considers ten separate characteristics that attend these technologies, and how each may affect the analysis of the proper scope of defense access. Section IV concludes.

II A Taxonomy of Digital Proof

The first step in articulating the issues that confound defense access to digital proof is to outline the general categories into which such technologies fall. Of course, digital information is used throughout the criminal justice process, e.g., in pre-trial bail and detention risk assessments and post-conviction at the time of sentencing. This chapter, however, focuses only on the use of such digital evidence to investigate and prove or disprove a defendant’s guilt.

In addition, although it might at first glimpse be appealing to attempt to draw sharp distinctions between consumer products and forensic law enforcement technologies, those categories prove illusory in this context.Footnote ¹⁵ The line between consumer and law enforcement either collapses, or is simply arbitrarily drawn, when it comes to defense investigation. For example, what difference does it make if law enforcement uses surveillance video from a police camera versus security footage from a Ring doorbell-camera or private bank? What does it matter if the facial recognition software is used on a repository of high school yearbooks versus police mugshots? Are questions of access so different when DNA testing was done via a public lab versus by a private lab, or whether the search was in a commercial versus law enforcement database?

Even if such a line were drawn, it may be difficult to defend in principle. Suppose law enforcement obtains data from an X (formerly Twitter) account, and then uses a proprietary law-enforcement software to do language analysis of the account. Is that a consumer product or law enforcement tool? Or if law enforcement secretly signs an agreement with a consumer DNA database to enable testing and searches for police purposes, is that a consumer tool or law enforcement tool? All too often, the lines between the two will break down as increasingly public–private cooperation generates evidence pertinent for a criminal case.

Of course, concerns about the reliability of evidence may differ when the evidence derives from a consumer product used by the general public as opposed to a forensic tool used only by law enforcement. Regulatory regimes and market incentives exercise an oversight function for commercial applications, and the financial incentives that ensure reliability for commercial products may be lacking in the law enforcement context. But those safeguards are not a substitute for a defendant’s opportunity to access and challenge technological evidence, because reliability of the government’s proof is not the only value at stake. The defense must have a meaningful right to access or challenge technological evidence, as a means of testing the non-digital aspects of the government’s proof as well as bolster its own case. Thus, in taxonomizing digital evidence, this chapter acknowledges but does not differentiate between technology created and used by general consumers versus those created primarily or exclusively by police.

III Characteristics of Digital Proof

The preceding section provides a general overview of the digital shift in evidence in criminal prosecutions, and identifies the ways in which such evidence might likewise be critical to the defense. And as the preceding section demonstrates, it is not the case that the data’s reliability is a defendant’s only concern. The defense also requires access to digital proof for the same reasons that the government does – because digital evidence can help find or bolster witnesses, establish a critical fact, or impeach a claim. But the preceding part also reveals just how little guidance exists, either as a matter of rules-based guidance or judicial opinion, for those attempting to craft a meaningful defense right to access and use this material. This part identifies the critical questions that must be answered, and values that must be weighed, in devising a regime of comprehensive access to such information for defensive or exculpatory purposes.

1. Who “owns” and who “possesses” the data. Perhaps the most important question with regard to technological forms of evidence is who owns and who possesses the data. Ownership is critical because, as with physical items, the right to share or disclose data often rests in the hands of the owner. Possession is also critical because, as with physical items, a possessor may disclose information surreptitiously without permission or knowledge of the owner.

The most straightforward cases involve physical items owned and possessed by the accused or a person sympathetic to the accused’s interest, in which the data is stored locally in the instrument. Such might be the case for a security camera owned by the defendant, or an electronic device with stored files. In these cases, the owner can make the evidence available to the defense.

But many technological forms of evidence will not be accessed so simply. As a general matter, ownership might be in public, governmental hands, such as police or public housing surveillance footage, or private hands, such as a private security camera. Even the category of private ownership is complex – ownership may be as simple as belonging to a private individual, or as complex as ownership held by a publicly traded large corporate entity. In some cases, ownership may even cross categories. Digital information in particular may have multiple “owners,” e.g., a user who uploads a picture to a social media site may technically own the intellectual property, but the terms of service for the site may grant the site-owner a broad license to use or publicize the material.Footnote ⁶⁹

When possession is divorced from ownership, a new suite of problems arises. Even when an owner is sympathetic or willing to share information with the defense, access may nonetheless be thwarted by an entity or person in possession of the data. Possession may also make access questions difficult because the reach of legal process may not extend to a physical site where information is kept. Possessors may also undermine the right of owners to exclude, e.g., if a security company grants visual access to the interior of a home against the wishes of the owner.

In both cases, obtaining data from third-party owners or possessors runs the further risk of disclosing defense strategies or theories. The government may have the capacity to hide its use of technology behind contracts with non-disclosure clauses or vague references to “confidential informants.” Human Rights Watch has labeled this practice “parallel construction” and documented its use.Footnote ⁷⁰ But a criminal defendant may not be able to operate stealthily, relying instead on the goodwill of a third party not to disclose the effort or a court’s willingness to issue an ex parte order.

2. Who created the data. In many cases, answering questions of ownership and possession will in turn answer the question of creation. But not always. An email may be drafted by one person, sent to a recipient who becomes its legal “owner,” and then possessed or stored by a third party. Or an entity may “create” information or data by collecting or analyzing material owned or possessed by another, e.g., a DNA sample sent for processing through analytical software; the data is created by the processing company, and the physical sample tested may be “owned” or possessed by the government.

Data created by the defendant perhaps poses only ancillary obstacles when it comes to a defendant’s access to information. If anything, a defendant’s claim to having created data may bolster their claim to access, even if they neither own nor possess it.Footnote ⁷¹ Some jurisdictions even specifically bestow upon an individual the right to access, correct, and delete data.Footnote ⁷² But it may also be the case that the diffusion of claims in data may complicate rules for access and use by the defense. Imagine a piece of technology or evidence possessed by one party, owned by another person, and created by still another person – with disputes between the parties about whether or not to release the information.

3. For what purpose was the data created. Another factor that must be considered in contemplating defense access is the source of the data and the purpose for which it was created. Much of the information that the defense may seek to access for exculpatory purposes will have likely been created for reasons unrelated to the criminal matter. For example, an automated vacuum may record the placement of the furniture in the room so that it can efficiently clean, but such data might be useful to show the layout at the time of the robbery. Or a search engine may store entered searches to optimize results and targeted advertising, but the record may suggest that a third party was the true killer. Such purpose need not be singular, either. The person searching the internet has one goal, but the internet search engine company has a different objective. The critical point is that the reason the information is there may help shed light on the propriety of defense access.

By way of example, the most compelling case for unconstrained access to the defense might be for data that was created specifically for a law enforcement purpose. In the national security context in the United States, a statutory frame exists to resolve some of these claims.Footnote ⁷³ Conversely, the most difficult case for access might be for information privately created for personal purposes unrelated to the criminal case. Although no single factor should determine the capacity and scope of access, the extent to which data or information is created expressly with a criminal justice purpose in mind may shed light on the extent to which such information should also be made accessible to an accused.

4. With what permissions was the data created. A related point arises with regard to how much of the general public is swept into a data disclosure, and the extent to which participants implicated by its disclosure are aware of the risks posed by broader dissemination. Open access to surveillance footage is troubling because it has the potential to implicate the privacy rights of persons other than the accused, who have no relation to the crime and who may not even know that they appear in the footage. Although we might tolerate those rights being compromised when it is only law enforcement who will access the information, or when used by a private operator with little incentive to exploit it, giving access to defense attorneys may create cause for concern. Persons in heavily policed neighborhoods may fear that, after viewing an image in surveillance, attorneys will be incentivized to accuse a third party of the crime simply out of expedience rather than in good faith. But such concerns may be minimized when creators or owners voluntarily provide data to law enforcement for law enforcement purposes.

5. How enduring or resilient is the data and who has the authority to destroy it. A central concern about defense access to data is that, without prompt and thorough access, such data will be destroyed before the defendant has a chance to request its preservation, if not disclosure. Some forms of evidence may be incredibly resilient. For example, cloud computing services or biometric identifiers of known persons may be highly resilient to destruction or elimination.

But other forms of data may be transient in nature, or subject to deliberate interference by an unwilling owner or holder of the data. Surveillance cameras notoriously run on short time loops, automatically erasing and retaping data in limited increments. Social media or other sites may promise total erasure of deleted material, not just superficial elimination from a single device.

Meaningful defense access to technological tools for exculpatory purposes requires attentiveness to timing, such that a defendant is able to access the material before its destruction. Even if the entitlement extends no farther than preservation, with actual access and use to be decided later, that would significantly impact a defendant’s capacity to make use of this information.

6. The form, expertise, and instrumentation required to understand or present the data. Generally speaking, evidentiary form is likely to be a less pertinent consideration in any framework for defense access than are questions related to ownership or possession. What import is it if the data is on a hard drive or flash drive? What matters is who owns it and who has it.

Nevertheless, any comprehensive scheme for meaningful defense access must consider form inasmuch as certain forms at the extreme may entail greater or lesser burdens on the party disclosing the information. Data diffused over a large and unsearchable system may provide important information to a defense team, but even turning over that data may present a significant challenge to its holder. In the Catalan case above, the surveillance video that exonerated him was physically held by the stadium officials and the production company. Fortunately, it was rather confined, as it covered one day and one game. It was the defense attorney who pored through the footage, isolating the exculpatory images.

But what if the records go beyond a single episode, or require special instrumentation to interpret. Information that requires that a holder devote significant time or resources to make the data available, or that is not readily shareable or accessible without expertise or instrumentation, may pose much more significant hurdles to open defense access. Some defense claims may actually be requests for access to services, rather than disclosure of information. For example, a defense request to run a DNA profile in the national database or to query a probabilistic genotyping system with a different set of parameters is less about traditional disclosure than about commandeering the government’s resources to investigate a defense theory. The same could be true for location data from a witness’s phone or search records from a particular IP address. The sought information is less an item than a process, a process to be conducted by a third party, not the defense.

7. What are the associated costs or expenses, and are there even available experts for the defense. A critical logistical, if not legal, hurdle to defense access to digital and technological evidence is the cost associated with seeking, interpreting, and introducing such evidence. Most of the forms of evidence described require some degree of expertise to extract, interpret, and understand, much less to explain to a judge, attorney, or juror. To the extent that the information also seeks an operational process or other search measure, the owner or possessor of the information may justly charge a fee for such services. Even more troubling, some vendors may restrict access to the government, or there may not be an available defense expert to hire given the lack of a robust market.

In this way, cost alone can preclude equitable access. For example, even assuming the defendant could get access to the probabilistic software used to interpret a complicated DNA crime sample, and assuming the vendor who contracted with the government would agree to run a defense query, the vendor may nonetheless charge for the service. Routine costs like copying fees or hourly rates can quickly put even routine investigative efforts beyond the capacity of a criminal defense lawyer, as the vast majority of defendants are indigent and there may be insufficient public funds available or such funds may be jealously guarded by judicial officials.Footnote ⁷⁴

The introduction of this evidence also may require payment of expert fees so that the attorney is able to understand and clearly present the findings. Such costs can make defense lawyers reluctant even to pursue exculpatory evidence, because actually obtaining and using it appears insurmountable.

One still more troubling possibility is that some subpoenaed parties will actively choose to defy orders rather than comply. As discussed above, social media companies Facebook and Twitter (now X) both refused to turn over posts requested by the defense in a criminal case, leading the judge to hold them in contempt and fine each $1,000 – the maximum allowed under the law.Footnote ⁷⁵ With fines capped statutorily, a company wealthy enough or unlikely to be a repeat player might simply choose non-compliance.

8. What are the privacy implications of divulging the data. Perhaps the most apparent and central concern raised by defense access to digital data relates to privacy. The nature of the material sought and the scope of what it reveals, along with the number of persons implicated by defense disclosure, is perhaps equal only to concerns about unnecessary “fishing expeditions” or wasted resources as a basis for the reluctance to provide generous access to the defense. Whereas the government is bound to act in the interest of the public, and thus in theory should minimize harm to innocent third parties in the course of its investigations, the defense is entitled to act only in furtherance of the interest of the accused.

At one end of the spectrum, some technological evidence will reveal deeply private or personal data belonging to a person wholly unrelated to the criminal offense. The DNA sequence or entire email history of a witness, or extensive public surveillance of a small community, obviously implicate profound interests. But at the other end of the spectrum, discrete bits of information created by the defendant him- or herself pose little concern when sought by the same defendant.

And of course, non-digital forms of evidence can raise the same concerns. As such, there are already mechanisms available to limit the privacy impact of revealing information to the defense. In prosecution investigations, it is not unusual to have a “taint team” that reviews sensitive information and passes along only the incriminating material to the prosecutor in the case. Or a judge can take on the responsibility to review material in camera, i.e., outside of the view of the parties and their attorneys, and disclose only evidence that is relevant to the defense.

In short, privacy is understandably a central and driving concern, but it should not be a definitive reason to close the door on broad defense access to exculpatory or defensive material.

9. What legal restrictions, whether substantive, procedural, or jurisdictional, limit access or use. The final critical inquiry incorporates some aspects of the privacy concerns just discussed, but goes beyond them. Namely, any comprehensive effort to provide defense access to digital and technological evidence must square with existing legal regimes surrounding disclosure and use of such evidence, whether as a matter of comprehensive or targeted privacy laws, intellectual or physical property, trade secret, or evidence. At a basic level, the data may straddle jurisdictions – created in one place, processed in another, and then used somewhere else. Legal restrictions may also be loosely lumped into substantive and procedural limitations, differentiating between substantive constraints such as privacy laws and procedural impediments such as jurisdictional rules.

Background statutory regimes that may conflict with defense access are imperative to consider, because jurisdictions increasingly have adopted such restrictions in response to complaints about privacy.Footnote ⁷⁶ Although law enforcement is routinely afforded exceptions to privacy statutes,Footnote ⁷⁷ there is rarely any mention of any equivalent route of access for a criminal defendant.Footnote ⁷⁸ Moreover, even outside the realm of privacy law, there may be other statutory limitations on disclosure or access, such as legal non-disclosure agreements. Or jurisdictions may point to regulatory regimes aimed at reliability as sufficient to safeguard all of the defendant’s interests.Footnote ⁷⁹

IV Conclusion

Digital proof is here, and it is here to stay. Such proof has already assumed a prominent place in the prosecution of criminal suspects. But all too often, the ability of the defense to access and utilize such evidence depends on happenstance rather than formal right. By cataloging and characterizing this critical form of proof, the chapter hopes to support efforts to formalize and standardize a defendant’s ability to marshal defense evidence for exculpatory and adversarial purposes as readily as the government does to inculpate.

I. IntroductionFootnote ^*

Technology has rapidly changed our society over the past decades. As a result of the ubiquitous digitalization of our society, people continuously leave digital traces behind. Some have already referred to this as “digital exhaust.”Footnote ¹ People are often monitored without being aware of it, not only by camera surveillance systems, but also by their own smartphones and by other devices they use to access the internet.

Information about the whereabouts, behavior, networks, intentions, and interests of people can be very useful in a criminal law context. It is used mainly for guiding criminal investigations, as it may provide clues on potential suspects, witnesses, etc., but it can also constitute evidence in courts, as the data may confirm specific actions and behavior of actors. In other words, digital data can be used to find out exactly what happened, understood in the legal context as finding the truth, and try to prove what happened, understood in the legal context as providing evidence. This chapter focuses on the use of digital data as evidence in criminal courts. The large amounts of potentially useful data now available may cause a shift in the types of evidence presented in courts, in that there may be more digital data as evidence, in addition to or at the cost of other types of evidence, such as statements from suspects, victims, and witnesses.Footnote ²

However, in many jurisdictions, the legal provisions setting the rules for the use of evidence in criminal courts were formulated long before these digital technologies existed. As a result of ongoing technological developments, there seems to be an increasing discrepancy between legal frameworks and actual practices. The chapter investigates this disconnect by analyzing the relevant legal frameworks in the European Union for processing data in criminal courts and then comparing and contrasting these with actual court practices.

The relevant legal frameworks are criminal law and data protection law. Data protection law is mostly harmonized throughout the European Union, via the General Data Protection Regulation (GDPR)Footnote ³ and by regulation more specifically tailored to the criminal law context, via Directive 2016/680, also known as the Law Enforcement Directive (LED).Footnote ⁴ Criminal law, however, is mostly national law, with limited harmonization throughout the European Union. For this reason, criminal law is considered from a national perspective in this chapter. Criminal law in the Netherlands is taken as an example to illustrate the issues that may arise from using data as evidence in criminal courts.

Although Dutch criminal law may not be representative for all EU Member States, the discrepancies between EU data protection law and Dutch criminal law may be similar to other EU Member States. As such, the Netherlands may serve as a helpful example of how legal provisions dealing with the use of evidence in criminal courts is not aligned with developments in data as evidence.

We also think that reviewing the use of data as evidence in courts in the Netherlands may be interesting for other jurisdictions, because it can provide some best practices as well as identify caveats and pitfalls that can perhaps be avoided in other countries. We see two major arguments supporting such a claim. First, the issues of using data as evidence in courts are likely to be the same across Europe, as the technologies available are not confined to one or particular jurisdictions. This point also applies to the forensic standards that are applied, as these also have an international scope and nature, either because they are established by international standardization organizations such as ISO,Footnote ⁵ CEN-CENELEC,Footnote ⁶ and ETSIFootnote ⁷, or, if created on a national level, are at least aligned among forensics experts from different countries. Second, the legal frameworks for using data as evidence in courts are highly comparable. This is particularly the case for data protection law, which is highly harmonized across the European Union. Criminal law may not be harmonized that much across the European Union, but the norms and standards for evidence and fair trial are fleshed out in large part by the European Convention on Human Rights (ECHR) and Court of Justice of the European Union (CJEU) case law. All this means that the basic situation regarding technology and forensic practices and the relevant legal boundaries are more or less the same across the European Union, although national interpretations and practices within these confines may vary.

There are two other reasons to use the Netherlands as an example in this chapter, both related to the fact that the Netherlands is in the forefront of relevant regulation. First, international legal comparisons show that the Netherlands is a front runner in privacy and data protection law in several aspects.Footnote ⁸ The Netherlands implemented national legislation with higher levels of data protection than strictly necessary for compliance with EU data protection laws. Typical examples are data breach notification laws and mandatory privacy impact assessments that already existed in the Netherlands before the GDPR came into force in 2018.Footnote ⁹ Also, when looking at the criminal law context, the Netherlands was among the first countries to have specific acts for the police and the judiciary dealing with the processing of personal data in criminal law, long before EU Directive 2016/680 (the LED, see section III.C) came into force.Footnote ¹⁰ If there exists a disconnect between legal frameworks and actual practices with regard to data as evidence in criminal courts in a country that seems to be a regulatory front runner, in this case the Netherlands, similar problems may also exist in other EU Member States.

Second, the Netherlands is among the front runners in digital forensics and cybercrime legislation.Footnote ¹¹ The Netherlands was among the initiators of the Convention on Cybercrime, adopted by the Council of Europe in 2001, which includes provisions that relate to the processing of police data.Footnote ¹² This Convention regulates, among other things, the protection of personal data and international cooperation, including the exchange of personal data in criminal law cases between authorities of different countries. Also, the Netherlands ratified a series of legal instruments that aim to advance the cooperation and sharing of information between Member States, such as the Prüm TreatyFootnote ¹³ (for exchanging DNA data, fingerprints, and traffic data), the Schengen Information SystemFootnote ¹⁴ (for international criminal investigation information), the Visa Information SystemFootnote ¹⁵ (for visa data, including biometrical data), and the Customs Information SystemFootnote ¹⁶ and EurodacFootnote ¹⁷ (for fingerprints of asylum seekers and stateless people). The institutional regulations for Europol, Eurosur, and Eurojust contain provisions for the exchange of criminal law information between Member States.

In short, the Netherlands appears to be among the first countries in the European Union to develop both privacy and data protection, and digital forensics and cybercrime legislation. This characteristic is relevant because if there is a disconnect between legal frameworks and actual practices with regard to data as evidence in criminal courts in a country that seems to be in the forefront of regulation, in this case the Netherlands, it may be expected that similar problems also exist in other EU Member States.

In the Netherlands, a founding member of the European Union and its predecessors, there has been an extensive debate in society and in politics on how to balance using data in a criminal law context and protecting the right to privacy.Footnote ¹⁸ This debate has influenced the legal frameworks that regulate the use of data in criminal law. There are competing legal frameworks regulating this area: on the one hand, criminal law, including both substantive and procedural criminal law, and, on the other hand, privacy law, more specifically data protection law. It is important to note that both legal frameworks provide rules for allowing and restricting the use of personal data in criminal law, as sometimes there is a misunderstanding that criminal law would only or mainly allow the collection and processing of data, whereas data protection law would only or mainly restrict such data collection and processing.

The focus of this chapter is the discrepancy between legal frameworks and actual practices. First, the relevant legal frameworks for processing data in Dutch criminal courts are analyzed, i.e., Dutch criminal procedure law and EU data protection law). After this legal analysis, current court practices are examined, mainly by looking at typical case law and current developments in society and technology.

This chapter is structured as follows. Section II provides a brief general introduction to Dutch criminal procedure law. Section III provides a brief general introduction to EU data protection law and to some extent its implementation in Dutch data protection law, focusing on the GDPR and the LED respectively. Section IV investigates the actual use of evidence in Dutch criminal courts by focusing first on current court practices as reflected in case law, and second on current developments in society and technology. Section V compares current court practices with the developments in society and technology, in order to see whether there is a need to change court practices or the underlying legal frameworks.

II Criminal Procedure Law: The Example of the Netherlands

As the Netherlands is used as an example of national law in this chapter, some background information is provided regarding Dutch criminal law. The Dutch Code of Criminal Procedure (Dutch CCP)Footnote ¹⁹ dates back to 1926. Back then, the Code was characterized as “moderately accusatorial” since it introduced more rights for the defense than before that time.Footnote ²⁰ Today, however, the suspect remains to a large extent the object of investigation, rather than, e.g., the victim, which has become increasingly important in Dutch criminal law in recent decades.Footnote ²¹ This is especially the case in the stages of police investigation, before the start of the trial. Although over the years more possibilities for the defense to influence the earlier investigation were introduced, such as the right to contra-expertise during police investigation in Article 150b of the Dutch CCP, the defense and the prosecutor are far from equal parties. Basically, the room for maneuver for the defense largely depends on the prosecutor’s goodwill, as it is the prosecutor who leads the criminal investigation.Footnote ²² A more accurate description of Dutch criminal procedure would therefore be “moderately inquisitorial.”Footnote ²³

Fundamental to the position of the defense is the right to silence in Article 29 of the Dutch CCP. Rights and principles such as the privilege against self-incrimination, the equality of arms, and the presumption of innocence are not explicitly laid down in the Dutch CCP. They apply, however, directly to Dutch criminal procedure through Article 6 of the ECHR.

The Dutch CCP has been amended and supplemented many times since its creation in 1926. As a result, the Dutch CCP now looks more like a patchwork instead of structured and clear-cut Code. This is also one of the reasons that the legislator started the major, still-running project “Modernisation Criminal Procedure” (Modernisering Strafvordering) in 2014. This revision of legislation was not finished as of 2023, and it will take several more years before it is finished. The idea is to revise the Dutch CCP in order to make criminal procedure, among other things, more accessible and efficient.Footnote ²⁴ Another aim of the revision is to tackle one of the greater challenges criminal procedures face nowadays, those of keeping up with technological developments in criminal investigation practice and developing an overall framework for regulating criminal investigation in the digital era. The Dutch CCP is still very much an analog-style Code that regulates the searching of homes, the seizure of letters, wiretapping, the questioning of witnesses, etc. Various digital investigation methods can be conducted on the basis of existing powers, e.g., a computer that was seized in a home can be searched just like a diary or a pistol that was seized in a home,Footnote ²⁵ and several new digital investigation methods have been laid down in the Dutch CCP, e.g., the network search of Article 125j of the Dutch CCP or the hacking powers in Article 126nba of the Dutch CCP,Footnote ²⁶ but many digital methods are still unregulated. Awaiting legislation, some gaps have been filled provisionally by the Supreme Court, in cases where the defense questioned the legitimacy of certain methods. One important discussion concerns the legitimacy of searching a smartphone that was seized from a suspect after arrest. In 2017, the Supreme Court ruled that the general power of a policeman to “seize and search objects the suspect carries with him when arrested” in Articles 94 and 95 of the Dutch CCP can be the basis of a smartphone search under the condition that the infringement on the right to privacy remains limited.Footnote ²⁷ In cases where the infringement exceeds a limited search, such a search should be conducted or authorized by the public prosecutor. When it is foreseeable that the privacy-infringement will be “serious” (zeer ingrijpend), the investigatory judge needs to be involved.

The smartphone ruling of the Supreme Court needs to be understood from the perspective of the procedural legality principle that is laid down in Article 1 of the Dutch CCP. This article states that criminal procedure can only take place as foreseen by law,Footnote ²⁸ which means that the police cannot use investigation methods that infringe fundamental rights which are not explicitly grounded in a sufficiently detailed and explicit statutory investigation power. However, investigation methods that are not explicitly regulated in the Dutch CCP, like the seize and search powers in Articles 94 and 95 of the Dutch CCP mentioned above, and that only cause minor infringements, can be based on Article 3 of the Police Act.Footnote ²⁹ This Article contains the general description of the task carried out by the police: “it is the task of the police to maintain the legal order in accordance with the rules and under the subordination of the competent authority.”Footnote ³⁰ In case law, several digital investigation methods have been found to constitute only a minor infringement and therefore did not need to be explicitly regulated.Footnote ³¹ For example, sending stealth text messagesFootnote ³² to someone’s cell phone can in principle be based on the general police task description, except when this is done for such a period or with such frequency and intensity that a complete image is revealed of certain aspects of someone’s private life.Footnote ³³ The smartphone case, in which a very general power to seize was found to be a sufficient statutory basis for a limited smartphone search, builds on this settled case law.Footnote ³⁴ In its legislative draft on digital investigation, the “Modernisation” legislator has incorporated the so-called “pyramid-structure” of the smartphone case, i.e., within the categories of limited, more than limited, and serious intrusions. A larger privacy infringement demands a higher approval authority, so instead of the police, a prosecutor or investigatory judge is required. Also, limited intrusions do not have to be explicitly regulated, while more than limited and serious intrusions are in need of more detailed and stringent legislation. To distinguish between the different levels of privacy intrusion, the legislator uses the concept of “systematicness” (stelselmatigheid).Footnote ³⁵ This means that, e.g., a “forseeably systematic” computer or network search can be ordered by the public prosecutor, while a “foreseeably serious systematic” computer or network search also needs a warrant from the investigating judge.Footnote ³⁶ The same regime applies to research in open sources.Footnote ³⁷ The post-smartphone case law already demonstrates that the category of seriously systematic is almost non-existent in practice.Footnote ³⁸ Although the introduction of the pyramid structure is also based on the practical premise that the investigating judge should not be overburdened within the context of digital investigations, this does raise serious concerns about the level of legal protection.

III Dutch and EU Data Protection Law

IV Evidence in Dutch Criminal Law

V Conclusion

In this chapter, we focused on the increasing discrepancy between legal frameworks and actual practices regarding the use of data as evidence in criminal courts. The two legal frameworks under consideration are criminal law and data protection law. Since the EU harmonization of criminal law is very limited, we used the example of the Netherlands to further examine the use of data as evidence in criminal courts. Even though the Netherlands is a front runner in the areas of privacy and data protection law, as well as digital forensics and cybercrime, large parts of its criminal law were developed before digital evidence existed. Data protection law, which is more recent, is highly harmonized throughout the European Union via the GDPR and the LED.

The two major legal frameworks of criminal law and data protection law are not fully integrated and adjusted to each other. There seems to be a structural ambiguity here. When it comes to regulating data as evidence, these frameworks together need to cover three separate but intertwined activities: (1) collection of data; (2) processing and analysis of data, including storage, selecting, combining; and (3) evaluation of data.Footnote ⁹⁸ In the Netherlands, the Dutch CCP covers the collection and evaluation, while the processing is mainly the domain of the Wpg and Wjsg in accordance with the LED.

Based on the analysis of the existing legal frameworks, the actual use of data as evidence in criminal courts, and developments in society and technology, we have four major observations, regarding the final aspect of our research question: i.e., what is needed next. A first observation regarding regulation is that the existing legal frameworks in the Netherlands barely or not at all obstruct the collection of data for evidence. Hence, the legal frameworks essentially allow law enforcement agencies and public prosecutors to make use of the opportunities that data can offer as evidence in criminal courts. Although many digital investigation methods are not provided for in the Dutch CCP, and as a result, fundamental issues on privacy are debated, this seems to have few consequences for the legitimacy of data as evidence in specific cases. This is partly due to the fact that, in the Netherlands, illegally gathered evidence rarely leads to serious consequences. The Supreme Court case law thus reflects the importance given to crime fighting. Another explanation is that the debate on how to define and protect the right to digital privacy within criminal procedure is still in its infancy.

Our second observation is that regulation regarding collection via the Dutch CCP and regulation on processing and analysis via the Wpg and Wjsg is not integrated. As with other written law, these legal frameworks use different language and definitions, have different structures, and lack any cross-reference to one another. The Dutch CCP is not specifically aimed at what can be done with data once collected, but what can be done with data is also relevant for the evaluation of the extent of the privacy intrusion, and hence the design of the investigation powers. An integrated approach is also necessary for other reasons. Under data protection law, data subjects have a series of data subject rights they can invoke, such as the right to information, transparency, and access. These rights can be somewhat of a farce, as people may not know about them and how to invoke them and, if they do, they may be blocked in cases where a criminal investigation is still ongoing.Footnote ⁹⁹

Our third observation concerns the absence of regulation of automated data analysis during all stages in the criminal justice system, including the prevention, investigation, detection, or prosecution of criminal offenses, the use of data as evidence in criminal courts, and the execution of criminal penalties. Automated data analysis raises fundamental questions regarding the equality of arms, and because all parties should have access to all relevant data and be able to assess data selection, we would like to argue that introducing some additional provisions for regulating data analytics, subsequent to data collection, would be appropriate. We have not seen any similar provisions in the legislation of other EU Member States,Footnote ¹⁰⁰ but we did encounter an example of such a provision in the Dutch Intelligence Agencies Act (Wet Inlichtingen- en Veiligheidsdiensten).Footnote ¹⁰¹ Article 60 of this Act states that the Dutch intelligence agencies are empowered to perform automated data analytics on their own datasets and open sources. The data can be compared and used for profiling and pattern recognition. Since no similar provision exists in criminal law, it is unclear whether law enforcement agencies are allowed to do the same. We are not arguing that they should or should not be allowed to do this, but we would like to argue that there should be more clarity regarding this issue.

The absence of regulation of data analysis raises issues regarding privacy and data protection of the data subjects whose data is being processed, but it can also raise issues regarding equality of arms during litigation in courts. Normally, suspects have access to all evidence brought forward in their case, including any data underlying the evidence. In practice, defendants may only get what prosecutors grant them, and they may not be aware of what is missing. Furthermore, if data analysis is based on large amounts of data, and that data includes the data of others,Footnote ¹⁰² a suspect may not be granted access to it; the GDPR prevents this in order to protect privacy and personal data. As a result, a suspect may not have full transparency regarding the data on which the analysis was based and may be unable to reproduce the analysis.Footnote ¹⁰³ If the data analytics involve very sophisticated self-learning technology such as AI, the prosecutor may not even know how the data analysis took place.

Finally, as a fourth observation, what may also need further attention is the level of court expertise in dealing with digital data as evidence. Given the increasing importance of data as evidence in criminal courts, it is imperative that judges understand some of the basics of how data is collected and processed before it results in the evidence that is presented to them. In order to evaluate the reliability and strength of the data-evidence, they have to be very aware of any of the pitfalls and issues mentioned in the previous section. Judges should be able to contest different types of data brought forward as evidence, even if the data is not contested by any of the litigating parties. For this reason, further training in this area may be important, as well as procedural rules identifying the basis for judicial assessment of how data was seized.

In view of these observations, we conclude that, on the one hand, there are perhaps no major obstructions in the existing legal frameworks for the use of data as evidence in criminal courts, but that, on the other hand, much of this area is in practice still a work in progress. In order to find the right balance between the interests of law enforcement and the rights of subjects in criminal cases, further work is needed. Further work would include research, but obviously also the development of case law, as the balancing of interests approach is at the heart of what courts do, most notably supreme courts, and particularly in search and seizure jurisprudence. Since criminal law and data protection law are more or less separate legal frameworks, they need to be further aligned, not necessarily by adjusting the legislation, but at least in detailing the actual practices and policies of law enforcement agencies further. The absence of any regulation regarding automated data analysis is a major concern and may have considerable consequences for data subjects and their rights in criminal cases. We suggest that, after further research, regulation be considered. Regulation can be done via legislation, but perhaps also via policies. And, finally, further training of actors in courts may be required to make all of this work.

When looking at the developments in society and technology, we expect that the use of data as evidence in courts will significantly increase in the coming decades. This means that the issues identified in this chapter, such as limited effectiveness of data subject rights provided in the LED and issues regarding the principle of equality of arms during litigation, may become more pressing in the near future. It is therefore important to further prepare both courts and law enforcement agencies for these challenges, as suggested above.

However, having said this, we do not expect that the use of other types of evidence in criminal courts, such as statements from suspects, victims, or witnesses, will fall out of use. We think it is important to consider the use of digital evidence in criminal courts as an addition to the use of statements and other types of evidence, not as a replacement. Humans seek to understand evidence by means of stories, which means that regardless of its digital nature, data will always need to fit into a story – the stories of suspects, victims, and witnesses.Footnote ¹⁰⁴

I Introduction

A wide array of robot technologies now inhabit our life worlds – and their population grows every day. The extent, degree, and diversity of our interactions with these technologies serve daily notice that we now live in an unprecedented age of surveillance. Many of us carry with us personal tracking devices in the shape of cellular phones allowing service providers and “apps” to monitor our locations and movements. The GPS chips embedded in smart devices provide detailed location data to a host of third parties, including apps, social media companies, and public health agencies. Wearable devices monitor streams of biological data. The IoT is populated by a dizzying array of connected devices such as doorbells, smart speakers, household appliances, thermostats, and even hairbrushes, which have access to the most intimate, if often quotidian, details of our daily lives. And then there is the dense network of surveillance technologies such as networked cameras, license plate readers, and radio frequency identification (RFID) sensors deployed on terrestrial and airborne platforms, including autonomous drones, that document our comings and goings, engagements and activities, any time we venture into public. Increasingly, these systems are backed by AI technologies that monitor, analyze, and evaluate the streams of data produced as we move through physical and online worlds, many of which also have the capacity and authority to take action. What once was the stuff of dystopian fiction is now a lived reality.

Privacy scholars have quite reasonably raised concerns about threats to fundamental rights posed by robots. For example, Frank Pasquale has advanced a trenchant critique of black-box algorithms, which have displaced human agents in a variety of contexts.Footnote ¹ On the other hand, we readily invite robots into our lives to advance personal and social goals. Some play seemingly minor roles, such as autonomous vacuums and refrigerators capable of tracking their contents, determining when supplies are low, and submitting online orders. Others less so, such as fitness monitors that summon emergency medical personnel when they determine their human partners are in crisis, or mental wellness apps that utilize biometric data to recommend, guide, and monitor therapy.

Because they entail constant and intimate contact, these human–robot interactions challenge our conceptions of self, privacy, and society, stretching the capacities of our legal regimes to preserve autonomy, intimacy, and democratic governance. Prominent among these challenges are efforts to understand the role of constitutions as guarantors of rights and constraints on the exercise of power. In the United States, this is evident in conversations about the Fourth Amendment and technology.

The Fourth Amendment provides that: “The right of the people to be secure in their persons, houses, papers, and effects, against unreasonable searches and seizures, shall not be violated, and no warrants shall issue, but upon probable cause, supported by oath or affirmation, and particularly describing the place to be searched, and the persons or things to be seized.” Since the Supreme Court’s pivotal 1967 decision in Katz v. United States,Footnote ² the Fourth Amendment has been cast as a guarantor of privacy, which suggests that it might have a role to play in normalizing, protecting, and regulating the relationships among us, our technologies, corporations, and government agencies. Specifically, we might imagine the Fourth Amendment protecting us from threats to privacy posed by robots or securing our relationships with robots against threats of interference or exploitation. Unfortunately, doctrinal rules developed by the US Supreme Court have dramatically reduced the capacity of the Fourth Amendment to serve either role. Some of these rules have earned considerable attention, including the public observation doctrineFootnote ³ and the third party doctrine.Footnote ⁴ Others have so far avoided close scrutiny.

This chapter examines two doctrinal rules in this more obscure category that are particularly salient to robot–human interactions. The first is that privacy is a personal good, limiting standing to bring Fourth Amendment challenges to those who have suffered violations of their personal expectations of privacy. The second is that the Fourth Amendment can only reach the actions of state agents. This chapter will show that neither is required by the text or history of the Fourth Amendment. To the contrary, the text, history, and philosophical lineage of the Fourth Amendment favor a broader understanding of privacy as a public good that “shall be” secure against threats of intrusive surveillance and arbitrary power by both government and private actors, whether human or robotic. This reading should lead us to alter our understanding of a variety of Fourth Amendment doctrines,Footnote ⁵ including rules governing standing and the state agency requirement, thereby enhancing the potential of the Fourth Amendment to play a salutary role in efforts to understand, regulate, and even protect human–robot interactions.

Before turning to that work, it is worth pausing for a moment to wonder whether we would be better off abandoning the Fourth Amendment to these doctrinal rules and focusing instead on legislation or administrative regulation as a means to govern robot–human interactions. There are good reasons to doubt that we would be better off. Legislatures generally, and the US Congress in particular, have failed to take proactive, comprehensive action as new technologies emerge.Footnote ⁶ Instead, this is an area where Congress has tended to follow the courts’ lead. A good example is the Wire Tap Act,Footnote ⁷ passed in 1968 right after the Court’s landmark decision in Katz. Most important, however, is that accepting the degradation of any constitutional right out of deference to the political branches turns constitutional democracy on its head. The whole point of constitutional rights is to guarantee basic protections regardless of legislative sanction or inaction. At any rate, defending constitutional rights does not exclude legislative action. For all of these reasons, we should question the doctrinal rules that seem to limit the scope of constitutional rights rather than accepting them in the hope that legislatures or executive agencies will ride to the rescue.

II Fourth Amendment Standing

Established Fourth Amendment doctrine imagines that privacy is a personal good.Footnote ⁸ This received truth traces back to the Supreme Court’s 1967 opinion in Katz.Footnote ⁹ Confronted with the unregulated deployment and use of emerging surveillance technologies, including wiretaps and electronic eavesdropping devices, the Katz Court adopted a novel definition of “search” as a violation of a subjectively manifested “expectation of privacy … that society is prepared to recognize as ‘reasonable.’”Footnote ¹⁰ Applying this definition, the Katz Court held that eavesdropping on telephone conversations is a “search,” and therefore subject to Fourth Amendment regulation.

Once hailed as progressive, Katz’s revolutionary potential has been dramatically limited by its assumption that privacy is a personal good.Footnote ¹¹ This point is manifested most clearly in the Court’s decisions governing Fourth Amendment “standing.” “Standing” is a constitutional rule limiting the jurisdiction of US courts. To avail themselves of a court’s jurisdiction, Article III, section 2, of the US Constitution requires litigants to show that they have suffered a legally cognizable injury caused by the opposing party, and that the court can provide relief. Fourth Amendment standing shares a conceptual kinship with Article III standing, but is neither jurisdictional nor compelled by the text. It is, instead, derivative of the assumption in Katz that privacy is a personal good. Thus, a litigant must establish that his “own Fourth Amendment rights [were] infringed by the search and seizure which he seeks to challenge.”Footnote ¹²

Fourth Amendment standing doctrine hamstrings efforts to challenge overreaching and even illegal conduct. Consider United States v. Payner.Footnote ¹³ There, Internal Revenue Service (IRS) agents suspected that taxpayers were using a Bahamian bank to hide income and avoid paying federal taxes. Unable to confirm those suspicions, agents decided to steal records from Michael Wolstencroft, a bank employee. To facilitate their plan, agents hired private investigators Norman Casper and Sybol Kennedy. Kennedy established a relationship with Wolstencroft and arranged to go to dinner with him.Footnote ¹⁴ During their dinner date, Wolstencroft left his briefcase in Kennedy’s apartment. Casper retrieved the briefcase and delivered it to IRS agent Richard Jaffe. Caspar and Jaffe broke into the briefcase and copied its contents, including hundreds of pages of bank documents. They then replaced the documents, relocked the briefcase, and returned it to Kennedy’s apartment. This fantastic criminal conspiracy was carried out with the full knowledge and approval of supervisory agents at the IRS.

Among the stolen documents were records showing that Payner used Wolstencroft’s bank to hide income. Based on this evidence, Payner was charged with filing a false tax return. At trial, he objected to the introduction of the stolen documents on the grounds that they were the fruits of a conspiracy to violate the Fourth Amendment. District Judge John Manos granted Payner’s motion and condemned the government’s actions as “outrageous.”Footnote ¹⁵ In an effort to deter similar misconduct in the future, Judge Manos suppressed the stolen documents, concluding that “[i]t is imperative to signal all likeminded individuals that purposeful criminal acts on behalf of the Government will not be tolerated in this country and that such acts shall never be allowed to bear fruit.”Footnote ¹⁶ The government appealed to the Supreme Court.

Writing for the Court in Payner, Justice Lewis Powell acknowledged that the government intentionally engaged in illegal activity and did so on the assumption that it would not be held accountable. He also agreed that: “No court should condone the unconstitutional and possibly criminal behavior of those who planned and executed this ‘briefcase caper.’”Footnote ¹⁷ Nevertheless, Justice Powell held that Payner could not challenge the government’s illegal conduct because Payner’s personal “expectation of privacy” was not violated. The briefcase belonged to Wolstencroft. The documents belonged to the bank. Payner had no personal privacy interest in either. He therefore did not have “standing” to challenge the government’s illegal actions.

Payner shows how treating privacy as a personal good prevents many criminal litigants from challenging illegal searches and seizures. That may seem defensible in the context of a criminal case, where demonstrably guilty defendants seek to avoid responsibility by suppressing reliable evidence. But what about public-minded civil actions? Consistent with its English heritage, US law allows for civil actions seeking equitable relief in the form of declaratory judgments and injunctions. Given the different interests at stake in this context, and the decidedly public orientation of these actions, one might expect to see a more expansive approach to questions of standing when litigants bring Fourth Amendment suits designed to benefit “the people” by challenging the constitutionality of search and seizure practices and demanding reform. Unfortunately, doctrinal rules governing Fourth Amendment standing make it nearly impossible to pursue declaratory or injunctive relief in most circumstances.Footnote ¹⁸ The culprit, again, is the assumption that privacy is a personal good. Los Angeles v. LyonsFootnote ¹⁹ offers a vivid example.

Adolph Lyons sued the Los Angeles Police Department (LAPD) and the City of Los Angeles after officers put him in a chokehold during a traffic stop. The chokehold was applied with such intensity that Lyons lost consciousness and suffered damage to his larynx. Given that he was the person assaulted, Lyons clearly had standing to bring a civil action alleging violations of his Fourth Amendment rights. To his credit, however, Lyons was interested in more than personal compensation. He wanted to use his suit to compel the LAPD to modify its practices, policies, and training on the use of force. Those remedies would have benefited not just Lyons, but also the people of Los Angeles and the United States generally, enhancing the people’s collective security against unreasonable seizures. Unfortunately, the Court dismissed these equitable claims on the grounds that Lyons did not have standing.

In order to demonstrate standing to pursue injunctive relief, the Court held, Lyons would need to “establish a real and immediate threat that he would again be stopped for [a] traffic violation, or for any other offense, by an officer or officers who would illegally choke him into unconsciousness without any provocation or resistance on his part.”Footnote ²⁰ That is a virtually insurmountable burden, but it is entirely consistent with the Court’s assumption that privacy and Fourth Amendment rights are personal goods. No matter how public-minded he might be, or how important the legal questions presented, Lyons could not pursue judicial review of LAPD chokehold practices because he could not establish that his personal Fourth Amendment rights were in certain and immediate danger. It did not matter that the LAPD was still engaging in a pattern and practice of using dangerous, unjustified chokeholds, jeopardizing the Fourth Amendment rights of the people of Los Angeles as a whole. Those practices, and the threats they posed, did not violate Lyons’ personal interests, so he was powerless to demand change.

Both the assumption that privacy is a personal good and the derivative rules governing Fourth Amendment standing have consequences for robot–human interactions. For example, they can be deployed to insulate from judicial review the kinds of electronic data-gathering that are essential for robotic actors and easily conducted by robots. A ready example is Clapper v. Amnesty International.Footnote ²¹ There, a group of attorneys, journalists, and activists challenged the constitutionality of a provision of the FISA Amendments Act, 50 USC §1881a, granting broad authority for government agents to surveil the communications of non-US persons located abroad. The plaintiffs argued that this authority imperiled their abilities to maintain the confidence of their sources and clients, compromising their important work. While admitting that plaintiffs’ concerns were theoretically valid, the Supreme Court held that they did not have standing to challenge the law precisely because their fears were theoretical. In order to establish standing, the plaintiffs needed to demonstrate that their communications had actually been intercepted or were certain to be intercepted pursuant to authority granted by §1881a. As a result, the authority granted by §1881a remained in place, condemning “the people,” individually and collectively, to a state of persistent insecurity in their electronic communications against both human and robot actors.

Courts have also wielded rules governing Fourth Amendment standing to limit the ability of service providers to protect their customers’ privacy interests. For example, in California Bankers Association v. Shultz, the Supreme Court concluded that banks do not have standing to raise Fourth Amendment claims on their customers’ behalf when contesting a subpoena for financial records.Footnote ²² In Ellwest Stereo Theaters, Inc. v. Wenner, the Ninth Circuit Court of Appeals held that an adult entertainment operator had “no standing to assert the fourth amendment rights of his customers.”Footnote ²³ In a 2012 case where investigators subpoenaed Twitter for messages posted by political protestors along with location data, a New York trial court decided that Twitter did not have standing to object on Fourth Amendment grounds.Footnote ²⁴ In 2017, a federal court in Seattle found that Microsoft Corporation did not have Fourth Amendment standing to challenge the gag order provisions of 18 USC §2705(b), which the government routinely invoked when compelling providers of data, internet, and communication services to disclose information relating to their customers.Footnote ²⁵

We are likely to see more of these kinds of decisions in coming years as courts continue to wrestle with new and emerging technologies. Consider, as an example, Amazon’s Ring.

Ring is an internet-connected doorbell equipped with cameras, microphones, and speakers that allows owners to monitor activity around their front doors through a smartphone, tablet, or computer, whether they are inside their homes or in another time zone. Ring is capable of coordinating with other smart devices to provide users with access and control over many aspects of their home environments. There is also a Ring device for automobiles. Although Ring does not make independent intelligent choices or perform tasks based on environmental stimuli, it represents the kinds of technologies that inhabit the IoT, which includes a rapidly rising population of robots. Some of these robots gather stimuli directly through their onboard sensors. Others draw on sensorial inputs from other devices, such as Ring. Either way, devices like Ring represent a critical point of engagement for robots and humans as we grant intimate access to our lives and the world outside our front doors in order to obtain the convenience and benefits of robotic collaborators. As recent experiences with Ring show, that access is ripe for exploitation.

In August 2019, journalists revealed that Amazon had coordinated with hundreds of law enforcement agencies to allow them access to video images and other information gathered by Ring without seeking or securing a court order or the explicit permission of owners.Footnote ²⁶ It is hard to imagine a starker threat to the security of the people guaranteed by the Fourth Amendment than a program granting law enforcement access to our home environments. But who has standing to challenge this program? Criminals prosecuted using evidence from a Ring doorbell do not. That is the lesson from Payner. Owners of Ring doorbells do not, unless they can show that their personal devices have been exploited. That is the lesson from Clapper. But even a Ring owner who can show that her device was exploited cannot challenge the program writ large or demand programmatic reform. That is the lesson from Lyons. As a result, the Fourth Amendment appears to be functionally powerless, both to protect these sites of human–robot interaction,Footnote ²⁷ and to protect the public from robotic exploitation via these kinds of devices and the data they generate. As an example of technologies in this latter category, consider facial recognition, which is capable of conducting the kinds of independent analysis once the sole province of carbon-based agents.Footnote ²⁸

Rules governing Fourth Amendment standing are not the only culprits in the apparent inability of the Fourth Amendment to regulate robot–human interactions. As the next section shows, the state agency requirement also limits the role of the Fourth Amendment in protecting and regulating many robot–human interactions.

III The State Agency Requirement

Conventional doctrine holds that the Fourth Amendment binds state agents, not private actors.Footnote ²⁹ This state agency requirement limits the capacity of the Fourth Amendment to regulate and protect many human–robot interactions. Justice Samuel Alito recently explained why:

Many, if not most, of the robot–human interactions that challenge our conceptions of privacy, fracture social norms, destabilize our institutions, and that will most likely play central roles in our lives, are produced, deployed, and controlled by private companies. Smart speakers and other IoT devices, wearable technologies, and the myriad software applications that animate phones, tablets, and televisions are all operated by private enterprises. Some of these corporations have more immediate effects on our lives than government entities, and pose greater threats to our privacy, autonomy, and democratic institutions than government entities, but they stand immune from constitutional constraint because they are not state agents. As a consequence, the Fourth Amendment appears unable “to protect [the public] from this looming threat to their privacy.”Footnote ³¹

Here again, Ring provides a good example. Ring is part of a larger ecosystem of connected devices designed, sold, and supported by Amazon. In addition to Ring, many folks have other Amazon products in their homes, including Alexa-enabled devices, which are equipped with microphones and voice recognition technologies. These devices allow users to play music, operate televisions, order goods and services, make phone calls, and even adjust the lighting using voice commands. This ecosystem is increasingly open to devices capable of making independent choices. These are all wonderful human–robot interactions, but they come with the cost of allowing Amazon and its affiliates access to our homes and lives. By virtue of the state agency requirement, that relationship stands outside of Fourth Amendment regulation. Amazon, directly or through its robot intermediaries, is at liberty to threaten the security of the people in their persons and homes without fear of constitutional constraint so long as they do not directly coordinate with government agencies.

Must it be this way? Or does the Fourth Amendment have more to say about robot–human interactions than is suggested by rules governing standing and the state agency requirement? As the next sections argue, the text and history of the Fourth Amendment suggest that it does.

IV Challenging Fourth Amendment Standing

The Fourth Amendment undeniably protects collective interests and recognizes that privacy is a public rather than an exclusively private good. That is evident in the text, which uses the phrase “the people” instead of “persons.”Footnote ³² This choice was deliberate.Footnote ³³ Those who drafted the Fourth Amendment had competing models to choose from, as represented in various state constitutions, some of which employed “persons”Footnote ³⁴ and others “the people.”Footnote ³⁵ The drafters demonstrated awareness of these alternatives by guaranteeing Fifth Amendment protections to “persons”Footnote ³⁶ and Sixth Amendment rights to “the accused.”Footnote ³⁷ By choosing “the people,” the First Congress aligned the Fourth Amendment with political rights protected elsewhere in the Constitution,Footnote ³⁸ such as the First Amendment right to assemble and petition the governmentFootnote ³⁹ and the Article I right of the people to elect their representatives.Footnote ⁴⁰ That makes sense in light of contemporaneous experiences with general warrants and writs of assistance, which showed how search and seizure powers could be weaponized to silence political speech. As we shall see, those cases contributed to founding-era concerns that general warrants and writs of assistance threatened the collective security of the people, not just those who were actually the subject of searches and seizures, because the very existence of broad, indiscriminate licenses to search and seize threatened the security of the people as a whole.Footnote ⁴¹

The text of the Fourth Amendment reflects founding-era understandings that security against arbitrary searches and seizures was an essential feature of democratic society. The founders understood how searches and seizures could be used to oppress thought and speech. But they also understood the idea, well-established since the time of the ancients, that security in our persons, houses, papers, and effects is essential to processes of ethical, moral, and intellectual development, which in turn are essential to the formation and sustenance of citizens capable of performing the duties of democratic government.Footnote ⁴² This is privacy as a public good. The Fourth Amendment guarantees that public good by securing space for liberty, autonomy, civil society, and democracy against threats of oppressive scrutiny.

The Supreme Court is not completely blind to the collective interests at stake in the Fourth Amendment. Consider, as an example, its exclusionary rule jurisprudence. Most Fourth Amendment claims arise in the context of criminal trials where the remedy sought is exclusion of illegally seized evidence.Footnote ⁴³ The idea that illegally seized evidence should be excluded at trial is not derived from the text or history of the Fourth Amendment.Footnote ⁴⁴ In fact, nineteenth-century jurists rejected the idea.Footnote ⁴⁵ The exclusionary rule is, instead, a prudential doctrine justified solely by its capacity to prevent Fourth Amendment violationsFootnote ⁴⁶ by deterring police officers from violating the Fourth Amendment in the future.Footnote ⁴⁷ Although illegal evidence is excluded in the cases of particular defendants, there is no individual right to exclude evidence seized in violation of the Fourth Amendment.Footnote ⁴⁸ To the contrary, the Court has made clear that admitting evidence seized in violation of the Fourth Amendment “works no new Fourth Amendment wrong.”Footnote ⁴⁹ In making this prudential case for the exclusionary rule on general deterrence grounds, the Court recognizes that there is more at stake in a particular search or seizure than the personal privacy of a specific person.

The Court’s awareness of the collective interests at stake in Fourth Amendment cases is not limited to its exclusionary rule jurisprudence. For example, in Johnson v. United States, decided in 1948, the Court noted that “[t]he right of officers to thrust themselves into a home is also a grave concern, not only to the individual, but to a society which chooses to dwell in reasonable security and freedom from surveillance.”Footnote ⁵⁰ Similarly, in United States v. Di Re, the Court concluded that “the forefathers, after consulting the lessons of history, designed our Constitution to place obstacles in the way of a too permeating police surveillance, which they seemed to think was a greater danger to a free people than the escape of some criminals from punishment.”Footnote ⁵¹ Of course, these sentiments were issued before Katz, which shifted the focus to individual interests.

Importantly, however, Katz did not close the door, and there is some evidence that the Supreme Court may be ready to rethink rules governing Fourth Amendment standing in light of new challenges posed by emerging technologies. The strongest evidence comes from the Court’s decision in Carpenter v. United States.Footnote ⁵² There, the Court was asked whether the Fourth Amendment regulates governmental access to cell site location information (CSLI). CSLI has been a boon to law enforcement. It can be used to track suspects’ past movements and to establish their proximity to crimes. That is precisely what investigators did in Carpenter. Based on information from a co-conspirator, they knew that Carpenter was involved in a string of armed robberies. In order to corroborate that information, they obtained several months of CSLI for Carpenter’s phone, establishing his proximity to several robberies. At trial, Carpenter objected to the admission of this evidence on Fourth Amendment grounds.

In light of the Court’s views on standing and the state agency requirement, there was good reason to think that the government would prevail. After all, it was Carpenter’s cell phone company who, of its own accord, tracked his phone and stored his location information. It certainly did not appear to be acting as a state agent. Moreover, the information was recorded in the company’s business records. If Payner did not have standing to challenge the search of banking records, then why would Carpenter have standing to challenge the search of cellular service records? Despite these challenges, the Supreme Court held that the “location information obtained from Carpenter’s wireless carriers was the product of a search.”Footnote ⁵³ In doing so, the Court seemed to return to the pre-Katz era:

This reasoning marks a potential broadening of the Court’s approach to Fourth Amendment questions. Along the way, the Court seemed to recognize the important collective dimensions of the Fourth Amendment.Footnote ⁵⁵

The majority opinion in Carpenter does not directly address the question of Fourth Amendment standing. Nevertheless, Justices Anthony Kennedy and Clarence Thomas make clear that something potentially revolutionary is afoot in their dissenting opinions. For his part, Justice Kennedy reminds us that the Court’s precedents “placed necessary limits on the ability of individuals to assert Fourth Amendment interests in property to which they lack a requisite connection.”Footnote ⁵⁶ “Fourth Amendment rights, after all, are personal,” he continues, “[t]he Amendment protects ‘[t]he right of the people to be secure in their … persons, houses, papers, and effects’ – not the persons, houses, papers, and effects of others.” In the case of the business records at issue in Carpenter, Justice Kennedy concluded that they belonged to the cellular service provider “plain and simple.” Consequently, Carpenter, like Payner, “could not assert a reasonable expectation of privacy in the records.” Justice Thomas was even more pointed in his criticism, lambasting the majority for endorsing the idea that individuals can “have Fourth Amendment rights in someone else’s property.”Footnote ⁵⁷

The Carpenter majority offers no direct response to these charges, but there are hints consistent with the arguments sounding in the collective rights reading of the Fourth Amendment advanced in this chapter. For example, the Court recognizes that allowing government agents unfettered access to CSLI implicates general, collective interests rather than the specific interests of an individual. As Chief Justice John Roberts, writing for the majority, points out, cellular phones are ubiquitous, to the point that there are more cellular service accounts with US carriers than there are people. Furthermore, most people “compulsively carry cell phones with them all the time … beyond public thoroughfares and into private residences, doctor’s offices, political headquarters, and other potentially revealing locales.”Footnote ⁵⁸ From these facts, the majority concludes that granting unfettered governmental access to CSLI would facilitate programs of “near perfect surveillance, as if [the Government] had attached an ankle monitor to the phone’s user.”Footnote ⁵⁹ “Only the few without cell phones could escape this tireless and absolute surveillance.”Footnote ⁶⁰ This exhibits a keen awareness that the real party of interest in the case was “the people” as a whole. At stake was “the tracking of not only Carpenter’s location but also everyone else’s, not for a short period, but for years and years.”Footnote ⁶¹ Denying customers’ standing to challenge government access to those records would leave the people insecure against threats of broad and indiscriminate surveillance – exactly the kind of “permeating police surveillance” the Fourth Amendment was designed to prevent.Footnote ⁶²

Recognizing the collective dimensions of the Fourth Amendment provides good grounds for reconsidering rules governing Fourth Amendment standing. As the founders saw it, any instance of unreasonable search and seizure in essence proposed a rule, and the Fourth Amendment prohibits endorsement of any rule that threatens the security of the people as a whole.Footnote ⁶³ It follows that anyone competent to do so ought to be able to challenge a proposed rule and the practice or policy it recommends. To be sure, a citizen challenging search and seizure practices should be limited in terms of the remedy she can seek. Actions at law seeking compensation should be limited to individuals who have suffered a direct, compensable harm. On the other hand, anyone competent to do so should have standing to bring actions seeking equitable relief in the form of declaratory judgments condemning search and seizure practices or injunctions regulating future conduct. Neither should we require the kind of surety of future personal impact reflected in the Court’s decisions in Lyons and Clapper. The founding generation recognized that the very existence of licenses granting unfettered discretion to search and seize threaten the security of the people as a whole. Why, then, would we not permit a competent representative of “the people” to challenge a statute, policy, or practice that, by its very existence, leaves each of us and all of us to live in fear of unreasonable searches and seizures?

Expanding the scope of Fourth Amendment standing would enhance human–robot interactions by allowing competent persons and groups to challenge efforts to exploit those interactions. It would likewise enhance our security against robotic surveillants. It would allow the activist groups like those who brought suit in Clapper to challenge legislation granting broad access to electronic communications and other data sources likely to play a role on robot–human interactions. It would allow technology companies to challenge government demands for the fruits and artifacts of our engagements with technologies. It would also license competent individuals and organizations to seek declaratory and injunctive relief when companies and government agencies exploit our relationships with robots and other technologies or seek to deploy robotic monitors. There is no doubt that this expanded access to the courts would enhance the security, integrity, and value of our interactions with a wide range of technologies that inhabit our daily lives, both directly and indirectly, by increasing pressure on the political branches to act.

V Reconsidering the State Agency Requirement

Contemporary doctrine holds that the Fourth Amendment applies only to state agents, and primarily the police. A closer look at the text and history of the Fourth Amendment suggests that it is not, and was not conceived to be, so narrow in scope.

To start, there is the simple fact that the police as we know them today did not exist in eighteenth-century America. That was not for lack of models or imagination. By the late eighteenth century, uniformed, paramilitary law enforcement agencies with general authority to investigate crimes were familiar in continental Europe. But England had rejected efforts to adopt that model, at least in part because members of the nobility feared privacy intrusions by civil servants. When Sir Robert Peel was able to pass the Metropolitan Police Act in 1829, establishing the Metropolitan Police Force, the “Peelers” (later “Bobbies”) were limited to maintaining the peace and did not have authority to investigate crimes. America was a decade behind England, with police forces making their first appearances in Boston (1838) and New York (1845). It was not until the late nineteenth century that professionalized, paramilitary police forces with full authority to investigate crime became a familiar feature of American society. By then, the Fourth Amendment was a venerable centenarian.

By dint of this historical fact, we know that the Fourth Amendment was not drafted or adopted with police officers as its sole or even primary antagonists. The text reflects this, making no mention of government agents of any stripe. Who then, was its target? The historical record suggests that it was overstepping civil functionaries, including constables, administrative officials, tax collectors, and their agents, as well as private persons. This is evidenced by the complicated role of warrants in eighteenth-century common law.

Contemporary Fourth Amendment wisdom holds that the warrant requirement plays a critical prospective remedial role, guarding the security of citizens against threats of unreasonable search and seizure by interposing detached and neutral magistrates between citizens and law enforcement.Footnote ⁶⁴ Among others, Laura Donohue has made a persuasive case that the “unreasonable searches” targeted by the Fourth Amendment were searches conducted in the absence of a warrant conforming to the probable cause, particularity, oath, and return requirements described in the warrant clause.Footnote ⁶⁵ But, as Akhil Amar has pointed out, the eighteenth-century history of warrants is somewhat more complicated.Footnote ⁶⁶ Some of those complications highlight the role of private persons in conducting searches and seizures.

In a world before professional, paramilitary police forces, private individuals bore significant law enforcement responsibilities. In his commentaries, Blackstone recognized the right of private persons to effect arrests on their own initiative or in response to a hue and cry.Footnote ⁶⁷ Searches and seizures in support of criminal investigations often were initiated by civilians who might go to a justice of the peace to swear-out a complaint against a suspected thief or assailant.Footnote ⁶⁸ So, too, a plaintiff in a civil action could swear-out a warrant to detain a potential defendant.Footnote ⁶⁹ A justice of the peace would, in turn, exercise his authority through functionaries, such as constables, who, as William Stuntz has noted, were “more like private citizens than like a modern-day police officer,”Footnote ⁷⁰ or even civilian complainants themselves, by issuing a warrant authorizing those persons to conduct a search or seizure.Footnote ⁷¹ These private actors could conduct searches purely on their own authority as well, but in doing so would risk exposing themselves to claims in trespass.Footnote ⁷² Warrants provided immunity against these actions.

Searches and seizures performed by minor functionaries and civilians raised significant concerns in eighteenth-century England because they threatened established social hierarchies by licensing civil servants to invade the privacy of the nobility. Those same worries underwrote resistance to professional police forces and founding-era critiques of general warrants and writs of assistance.Footnote ⁷³ Unlike the particularized warrants issued by judicial officers based on probable cause imagined in the warrant clause, general warrants and writs of assistance provided broad, unfettered authority for bearers to search wherever they wanted, for whatever reason, with complete immunity from civil liability. These instruments were reviled by our eighteenth-century forebears because they invited arbitrary abuses of power.Footnote ⁷⁴ But those threats did not come exclusively from agents of the state or only in the context of criminal actions. To the contrary, one of the most pernicious qualities of general warrants and writs of assistance was that they allowed for the delegation of search and seizure authority to minor functionaries and private persons. This is evident in the signal eighteenth-century cases challenging general warrants and writs of assistance.

The philosophical lineage of the Fourth Amendment traces to three eighteenth-century cases involving general warrants and writs of assistance that “were not only well known to the men who wrote and ratified the Bill of Rights, but famous through the colonial population.”Footnote ⁷⁵ The first two, Wilkes v. WoodFootnote ⁷⁶ and Entick v. Carrington,Footnote ⁷⁷ dealt with efforts to persecute English pamphleteers responsible for writing and printing publications critical of King George III and his policies. In support of cynical efforts to silence these critics, one of the king’s secretaries of state, Lord Halifax, issued general warrants licensing his “messengers” to search homes and businesses and to seize private papers. After their premises were searched and their papers seized, Wilkes and Entick sued Halifax and his agents in trespass, winning large jury awards. The defendants claimed immunity, citing the general warrants issued by Halifax. In several sweeping decisions written in soaring prose, Chief Judge Pratt – later Lord Camden – rejected those efforts, holding that general warrants were contrary to the common law.Footnote ⁷⁸

The third case providing historical grounding for the Fourth Amendment is Paxton’s Case.Footnote ⁷⁹ This was one among a group of suits brought by colonial merchants challenging the use of writs of assistance to enforce British customs laws in the American colonies. The colonists were ably represented by former Advocate General of the Admiralty James Otis, who left his post in protest when asked to defend writs of assistance. In an hours-long oration, Otis condemned writs of assistance as “the worst instrument of arbitrary power, the most destructive of English liberty and the fundamental principles of law that ever was found in an English law book.”Footnote ⁸⁰ He ultimately lost the case; but colonial fury over the abuse of search and seizure powers played a critical role in fomenting the American Revolution.Footnote ⁸¹

Outrage over general warrants and writs of assistance was evident during the American constitutional movement.Footnote ⁸² Courts condemned them;Footnote ⁸³ state constitutions banned them,Footnote ⁸⁴ and states cited the absence of a federal prohibition on general warrants as grounds for reservation during the ratification debates.Footnote ⁸⁵ In order to quiet these concerns, proponents of the Constitution agreed that the First Congress would draft and pass an amendment guaranteeing security from threats posed by unfettered search and seizure powers. The Fourth Amendment fulfills that promise.

All of this goes to show that we can look to founding-era experiences with, and objections to, general warrants and writs of assistance to inform our understandings of the Fourth Amendment. That record shows that the Fourth Amendment should not be read as applying exclusively to government officials. In their critiques of general warrants and writs of assistance, founding-era courts and commentators often highlighted the fact that they provided for the delegation of search and seizure powers to civilian functionaries. For example, the court in Wilkes argued that: “If such a power [to issue general warrants] is truly invested in a secretary of state, and he can delegate this power, it certainly may affect the person and property of every main this kingdom, and is totally subversive of the liberty of the subject.”Footnote ⁸⁶ James Otis railed that “by this writ [of assistance], not only deputies, etc., but even their menial servants, are allowed to lord it over us.”Footnote ⁸⁷ “It is a power,” he continued, “that places the liberty of every man in the hands of every petty officer.” “What is this,” he lamented, “but to have the curse of Canaan with a witness on us; to be the servant of servants, the most despicable of God’s creation?” The extent of that servitude, he explained, was virtually without limit, so that “Customhouse officers [and] [t]heir menial servants may enter, may break locks, bars, and everything in their way; and whether they break through malice or revenge, no man, no court, can inquire.” Because a writ of assistance “is directed to every subject in the king’s dominions,” he concluded: “Everyone with this writ, may be a tyrant.”

To be sure, many of the antagonists in these cases were state agents, if only of minor rank, or were acting at the direction of state agents. But the existence of general warrants and writs of assistance allowed both private citizens and government officials to threaten home and hearth. Otis explained why in his oration, quoting language common to writs of assistance that allowed “any person or persons authorized,”Footnote ⁸⁸ including “all officers and Subjects,” to conduct searches and seizures.Footnote ⁸⁹ That inclusion of “persons” and “Subjects” reflected the fact that writs of assistance and general warrants were issued not just in cases of customs and tax enforcement, but also to assist private litigants in civil actionsFootnote ⁹⁰ or even to vindicate private animosities. Otis explained the consequences: “What a scene does this open! Every man prompted by revenge, ill humor, or wantonness, to inspect the inside of his neighbor’s house, may get a writ of assistance. Others will ask it from self-defense; one arbitrary exertion will provoke another, until society be involved in tumult and blood.”Footnote ⁹¹

Anticipating a charge of dramatization, Otis offered this anecdote:Footnote ⁹²

So, at the heart of this speech marking the birth of the American Revolution, we see Otis decrying general warrants and writs of assistance because they protected private lawlessness. That is hard to square with the contemporary state agency requirement.

The facts in Wilkes and Entick provide additional evidence of the potential for general warrants and writs of assistance to vindicate private interests and facilitate abuses of power. The searches in these cases aimed to discover evidence of libel against the king. In fact, the court in Entick characterized the effort as “the first instance of an attempt to prove a modern practice of a private office to make and execute warrants to enter a man’s house, search for and take away all his books and paper in the first instance ….”Footnote ⁹³ The Entick Court went on to suggest that allowing for the issuance of general warrants in search of libels would pose a threat to the security of everyone in their homes because simple possession of potentially libelous publications was so common.Footnote ⁹⁴

So, neither the text nor history of the Fourth Amendment appear to support a state agency requirement, at least not in its current form. That is evidenced by the fact that a strict state agency requirement appears to exclude from Fourth Amendment regulation some of the searches and seizures cited as bêtes noires in the general warrants and writs of assistance cases. Certainly, nothing in the text suggests that state agents are the only ones capable of threatening the security of the people against unreasonable searches and seizures. Moreover, eighteenth-century criticisms of search and seizure powers indicate that the founding generation was concerned about arbitrary searches performed by a range of actors. Given that history, there is good reason to conclude that the Fourth Amendment governs the conduct of private entities to the extent they pose a threat to collective interests, including privacy as a public good. Fortunately, the Court appears to be developing some new sympathies that line up with these ancient truths.

In addition to sparking a potential revolution in the rules governing Fourth Amendment standing, the Carpenter Court also appears to have introduced some complications to the state agency requirement. To start, the Court is never clear about when, exactly, the search occurred in Carpenter and who did it. At one point, the Court states that the “Government’s acquisition of the cell-site records was a search within the meaning of the Fourth Amendment.”Footnote ⁹⁵ That would be in keeping with the state agency requirement. Elsewhere, the Court holds that the “location information obtained from Carpenter’s wireless carriers was the product of a search,”Footnote ⁹⁶ suggesting that his cellular service provider performed the search when it gathered, aggregated, and stored the CSLI. That is intriguing in the present context.

The Carpenter Court does not explain its suggestion that cellular service providers engage in a “search” for purposes of the Fourth Amendment when they create CSLI records. This is an omission that Justice Alito, writing in dissent, finds worrisome, pointing out that: “The Fourth Amendment … does not apply to private actors.” Again, the Court offers no direct response, which may leave us to wonder whether its suggestion that gathering CSLI is a search was a slip of the pen. There are good reasons for thinking this is not the case. Foremost, Carpenter is a landmark decision, and Chief Justice Roberts has a well-deserved reputation for care and precision in his writing. Then there is the fact that what cellular service providers do when gathering CSLI can quite naturally be described as a “search.” After all, they are looking for and trying to find an “effect” (the phone) and, by extension, a “person” (the user).Footnote ⁹⁷ By contrast, it is hard to describe the simple act of acquiring records as a “search,” although looking through or otherwise analyzing them certainly is. And then there is the fact that the acquisition was done by the familiar process of subpoena. As Justice Samuel Alito points out at length in his dissenting opinion, treating acquisition of documents by subpoena as a “search” would bring a whole host of familiar discovery processes within the compass of the Fourth Amendment.Footnote ⁹⁸ By contrast, treating the aggregation of CSLI as the search would leave that doctrine untouched. For all these reasons, the best, most coherent, and least disruptive option available to the Court may have been holding that the cellular service provider conducted the search at issue in Carpenter.

Expanding the scope of Fourth Amendment regulations to searches conducted by private actors would provide invaluable protections for human–robot interactions and protections from robot surveillants. As Justice Alito points out in his Carpenter dissent, many of the most significant contemporary threats to privacy come from technology companies and parties who have access to us and our lives through our robot collaborators or deploy and use robots as part of their businesses. This gives them extraordinary power. We have certainly seen the potential these companies hold to manipulate, influence, and disrupt civil society and democratic institutions – just consider the autonomous decisions made by social media algorithms when curating content. In many ways, these companies and their robots are more powerful than states and exercise greater degrees of control. There can be no doubt that holding them to the basic standards of reasonableness commanded by the Fourth Amendment would substantially enhance individual and collective security, both in our engagements with robots and against searches performed by robots.

VI Conclusion

This chapter has shown that the Fourth Amendment’s capacity to fulfill its promise is limited by two established doctrines, individual standing and the state agency requirement. Together, these rules limit the ability of the Fourth Amendment to normalize, protect, and regulate human–robot interactions. Fortunately, the text and history of the Fourth Amendment provide grounds for a broader reading that recognizes collective interests, guarding privacy as a public good against threats posed by both state and private agents. More fortunately still, the modern Supreme Court has suggested that it may be willing to reconsider its views on standing and state action as it struggles to contend with new challenges posed by robot–human interactions. As they move forward, the Justices would be well-advised to look backward, drawing insight and wisdom from the text and animating history of the Fourth Amendment.