In 2021 the Bank of England will complete its transition from paper to polymer with the release of a new 50 pound note. A public selection process saw almost quarter of a million nominations for the face of the new note; the final decision, announced in July 2019, was that Alan Turing would be featured. Turing was a hero for his codebreaking during the Second World War. He also helped establish the discipline of computer science, laying the foundations for what we now call artificial intelligence (AI).Footnote 1 Perhaps his best-known contribution, however, is the eponymous test for when true ‘intelligence’ has actually been achieved.
Turing modelled it on a parlour game popular in 1950. A man and a woman sit in a separate room and provide written answers to questions; the other participants have to guess who provided which answer. Turing posited that a similar ‘imitation game’ could be played with a computer. When a machine could fool people into believing that it was human, we might properly say that it was intelligent.Footnote 2
Early successes along these lines came in the 1960s with programs like Eliza. Users were told that Eliza was a psychotherapist who communicated through words typed into a computer. In fact, ‘she’ was an algorithm using a simple list-processing language. If the user typed in a recognised phrase, it would be reframed as a question. So after entering ‘I'm depressed,’ Eliza might reply ‘Why do you say that you are depressed?’ If it didn't recognise the phrase, the program would offer something generic, like ‘Can you elaborate on that?’ Even when they were told how it worked, some users insisted that Eliza had ‘understood’ them.Footnote 3
Parlour games aside, why should it matter if a computer is ‘intelligent’? For several decades, the Turing Test was associated more with the question of whether AI itself was possible, rather than the legal status of such an entity. Yet it is commonly invoked in discussions of legal personality for AI, from Lawrence Solum's seminal 1992 article onwards.Footnote 4 Though no longer regarded as a serious measure of modern AI in a technical sense, the Turing Test's longevity as a trope points to a tension in debates over personality that is often overlooked.
As AI systems become more sophisticated and play a larger role in society, there are at least two discrete reasons why they might be recognised as persons before the law. The first is so that there is someone to blame when things go wrong. This is presented as the answer to potential accountability gaps created by their speed, autonomy, and opacity.Footnote 5 A second reason for recognising personality, however, is to ensure that there is someone to reward when things go right. A growing body of literature examines ownership of intellectual property created by AI systems, for example.
The tension in these discussions is whether personhood is granted for instrumental or inherent reasons. Arguments are typically framed in instrumental terms, with comparisons to the most common artificial legal person: the corporation. Yet implicit in many of those arguments, or explicit in their illustrations and examples, is the idea that as AI systems approach the point of indistinguishability from humans—that is, when they pass Turing's Test—they should be entitled to a status comparable to natural persons.
Until recently, such arguments were all speculative. Then in 2017 Saudi Arabia granted ‘citizenship’ to the humanoid robot SophiaFootnote 6 and an online system with the persona of a seven-year-old boy was granted ‘residency’ in Tokyo.Footnote 7 These were gimmicks—Sophia, for example, is essentially a chatbot with a face.Footnote 8 In the same year, however, the European Parliament adopted a resolution calling on its Commission to consider creating ‘a specific legal status for robots in the long run, so that at least the most sophisticated autonomous robots could be established as having the status of electronic persons responsible for making good any damage they may cause, and possibly applying electronic personality to cases where robots make autonomous decisions or otherwise interact with third parties independently.’Footnote 9
This article begins with the most immediate challenge, which is whether some form of juridical personality would fill a responsibility gap or be otherwise advantageous to the legal system. Based on the history of corporations and other artificial legal persons, it does not seem in doubt that most legal systems could grant AI systems a form of personality; the more interesting questions are whether they should and what content that personality might have.
The second section then turns to the analogy with natural persons. It might seem self-evident that a machine could never be a natural person. Yet for centuries slaves and women were not recognised as full persons either. If one takes the Turing Test to its logical, Blade Runner, conclusion, it is possible that AI systems truly indistinguishable from humans might one day claim the same status. Although arguments about ‘rights for robots’ are presently confined to the fringes of the discourse, this possibility is implicit in many of the arguments in favour of AI systems owning the intellectual property that they create.
Taken seriously, moreover, the idea that AI systems could equal humans suggests a third reason for contemplating personality. For once equality is achieved, there is no reason to assume that AI advances would stop there. Though general AI remains science fiction for the present, it invites consideration whether legal status could shape or constrain behaviour if or when humanity is surpassed. Should it ever come to that, of course, the question might not be whether we recognise the rights of a general AI, but whether it recognises ours.
I. JURIDICAL PERSONALITY: A BODY TO KICK?
Legal personality is fundamental to any system of laws. The question of who can act, who can be the subject of rights and duties, is a precursor to almost every other issue. Yet close examination of these foundations reveals surprising uncertainty and disagreement. Despite this, as John Dewey observed in 1926, ‘courts and legislators do their work without such agreement, sometimes without any conception or theory at all’ regarding the nature of personality. Indeed, he went on, recourse to theory has ‘more than once operated to hinder rather than facilitate the adjudication of a special question of right or obligation’.Footnote 10
In practice, the vast majority of legal systems recognise two forms of legal person: natural and juridical. Natural persons are recognised because of the simple fact of being human.Footnote 11 Juridical persons, by contrast, are non-human entities that are granted certain rights and duties by law. Corporations and other forms of business associations are the most common examples, but many other forms are possible. Religious, governmental, and intergovernmental entities may also act as legal persons at the national and international level.
It is telling that these are all aggregations of human actors, though there are examples of truly non-human entities being granted personhood. In addition to the examples mentioned in the introduction, these include temples in India,Footnote 12 a river in New Zealand,Footnote 13 and the entire ecosystem of Ecuador.Footnote 14 There seems little question that a State could attribute some kind of personality to new entities like AI systems;Footnote 15 if that happens, recognition would likely be accorded by other States also.Footnote 16
A. Theories of Juridical Personality
Scholars and law reform bodies have already suggested attributing AI systems with some form of legal personality to help address liability questions, such as an automated driving system entity in the case of driverless cars whose behaviour may not be under the control of their ‘driver’ or predictable by their manufacturer or owner.Footnote 17 A few writers have gone further, arguing that procedures need to be put in place to try robot criminals, with provision for ‘punishment’ through reprogramming or, in extreme cases, destruction.Footnote 18
These arguments suggest an instrumental approach to personality, but scholarly explanations of the most common form of juridical person—the corporation—show disparate justifications for its status as a separate legal person that may help answer the question of whether that status should be extended to AI systems also.
The aggregate theory, sometimes referred to as the contractarian or symbolist theory, holds that a corporation is a device created by law to allow natural persons who organise themselves as a group to reflect that organisation in their legal relations with other parties. Group members could establish individual contractual relations with those other parties limiting liability and so on, but the corporate form enables them to do so collectively at a lower cost.Footnote 19 The theory has been criticised and is, in any case, the least applicable to AI systems.Footnote 20
The fiction and concession theories of corporate personality have separate originsFootnote 21 but amount to the same thing: corporations have personality because a legal system chooses to give it to them. As the US Supreme Court observed in 1819, a corporation ‘is an artificial being, invisible, intangible, and existing only in contemplation of law’.Footnote 22 Personality is granted to achieve policy ends, such as encouraging entrepreneurship, or to contribute to the coherence and stability of the legal system, such as through the perpetuity of certain entities. The purposive aspect used to be more evident when personality was explicitly granted through a charter or legislation; in the course of the twentieth century this became a mere formality.Footnote 23 These positivist accounts most closely align with legislative and judicial practice in recognising personality, and could encompass its extension to AI systems.
The realist theory, by contrast, holds that corporations are neither fictions nor mere symbols but objectively real entities that pre-exist conferral of personality by a legal system. Though they may have members, they act independently and their actions may not be attributable to those members. At its most extreme, it is argued that corporations are not only legal but also moral persons.Footnote 24 This theory tends to be favoured more by theorists and sociologists than legislators and judges, but echoes the tension highlighted in the introduction: that legal personality is not merely bestowed but deserved. In practice, however, actual recognition as a person before the law remains in the gift of the State.Footnote 25
The end result is, perhaps, that Dewey was correct a century ago: ‘“person” signifies what law makes it signify’.Footnote 26 Though the question of personality is a binary, however—recognised or not—the content of that status is a spectrum. Setting aside for the moment the idea that an AI system might deserve recognition as a person, a State's decision to grant it should be guided by the rights and duties that would be recognised also.
B. The Content of Legal Personality
Legal personality brings with it rights and obligations, but these need not be the same for all persons within a legal system. Even among natural persons, the struggle for equal rights of women, ethnic or religious minorities, and other disadvantaged groups reflects this truth.Footnote 27
It is possible, for example, to grant only rights without obligations. This has tended to be the approach in giving personhood to nature—both in theory, when it was first advocated in 1972,Footnote 28 and in practice, such as in the Constitution of Ecuador.Footnote 29 One could argue that such ‘personality’ is merely an artifice to avoid problems of standing: enabling human individuals to act on behalf of a non-human rights holder, rather than requiring them to establish standing in their own capacity.Footnote 30 In any case, it seems inapposite to the reasons for considering personality of AI systems.
On the other hand, AI legal personality could come only with obligations. That might seem superficially attractive, but insofar as those obligations are intended to address accountability gaps there would be some obvious problems. Civil liability typically leads to an award of damages, for example, which can only be paid if the wrongdoer is capable of owning property.Footnote 31 One could imagine scenarios in which those payments were made from a central fund, though this would be more akin to compulsory insurance regimes proposed as an alternative means of addressing the liability question.Footnote 32 ‘Personality’ would be a mere formality.
In the case of corporations, personality typically means the capacity to sue and be sued, to enter into contracts, to incur debt, to own property, and to be convicted of crimes. On the rights side, the extent to which corporations enjoy constitutional protections comparable to natural persons is the subject of ongoing debate. Though the United States has arguably granted the most protections to corporate entities, even there a line has been drawn at guarantees such as the right against self-incrimination.Footnote 33 Typically, juridical persons will have fewer rights than natural ones. (A similar situation obtains in international law, where States enjoy plenary personality and international organisations may have varying degrees of it.Footnote 34)
1. Private law
The ability to be sued is one of the primary attractions of personality for AI systems, as the European Parliament acknowledged.Footnote 35 This presumes, of course, that there are meaningful accountability gaps that can and should be filled. These gaps are often overstated.Footnote 36 A different reason for wariness about such a remedy is that, even if it did serve a gap-filling function, granting personality to AI systems would also shift responsibility under current laws away from existing legal persons. Indeed, it would create an incentive to transfer risk to such electronic persons in order to shield natural and traditional juridical ones from exposure.Footnote 37 That is a problem with corporations also, which may be used to protect investors from liability beyond the fixed sum of their investment—indeed, that is often the point of using a corporate vehicle in the first place. The reallocation of risk is justified on the basis that it encourages investment and entrepreneurship.Footnote 38 Safeguards typically include a requirement that the name of a limited liability entity include that status in its name (‘Ltd’, ‘LLC’, etc) and the possibility of piercing the corporate veil in limited circumstances to prevent abuse of the form.Footnote 39 In the case of AI systems, similar veil-piercing mechanisms could be developed—though if a human were manipulating AI in order to protect him- or herself from liability, the ability to do so might suggest that the AI system in question was not deserving of its separate personhood.Footnote 40
Entry into contracts is occasionally posited as a reason to grant AI systems personality.Footnote 41 Yet the use of electronic agents to conclude binding agreements is hardly new. High-frequency trading, for example, relies on algorithms concluding agreements with other algorithms on behalf of traditional persons.Footnote 42 Though the autonomy of AI systems may challenge application of existing doctrine to such practices—notably, when something goes wrong, such as a mistake—this is still resolvable without recourse to new legal persons.Footnote 43
Taking on debt and owning property would be necessary incidents of the ability to be sued and enter into contracts.Footnote 44 The possibility that AI systems could accumulate wealth raises the question of whether or how they might be taxed. Taxation of robots has been proposed as a means of addressing the diminished tax base and displacement of workers anticipated as a result of automation.Footnote 45 Bill Gates, among others, has suggested that such robots—or the companies that own them—should be taxed.Footnote 46 Industry representatives have argued that this would have a negative impact on competitiveness and thus far it has not been adopted.Footnote 47 An alternative is to look not at the machines but at the position of companies abusing market position, with possibilities including more aggressive taxing of profits or requirements for distributed share ownership.Footnote 48 In any case, taxation of AI systems—like the ability to take on debt and own property—would follow rather than justify granting them personality.Footnote 49 (The question of AI systems owning their creations will be considered in the next section.Footnote 50)
In addition to owning property, AI systems might also be called on to manage it. In 2014, for example, it was announced that a Hong Kong venture capital firm had appointed a computer program called Vital to its board of directors.Footnote 51 As with the Saudi Arabian government's awarding of citizenship, this was more style than substance—as a matter of Hong Kong law the program was not appointed to anything; in an interview some years later, the managing partner conceded that the company merely treated Vital as a member of the board with observer status.Footnote 52 It is possible that human directors might delegate some responsibility to an AI system, but under most corporate law regimes they cannot absolve themselves of the ultimate responsibility for managing the organisation.Footnote 53 Most jurisdictions require that those directors be natural persons, though in some it is possible for a juridical person—typically another corporation—to serve on the board.Footnote 54 Shawn Bayern has gone further, arguing that loopholes in US business entity law could be used to create limited liability companies, with no human members at all.Footnote 55 This requires a somewhat tortured interpretation of that law—a natural person creates a company, adds an AI system as a member, then resignsFootnote 56—but suggests the manner in which legal personality might be adapted in the future.
2. Criminal law
A final quality of legal personality is the most visceral and worthy of some elaboration: the ability to be punished. If given legal personality comparable to a corporation, there seems little reason to argue over whether an AI system could be prosecuted under the criminal law. Provided actus reus and mens rea are established,Footnote 57 such an entity could be fined or have its property seized; a licence to operate could be suspended or revoked. In some jurisdictions, a winding up order can be made against a juridical person; where that is not available, a fine sufficiently large to bankrupt the entity may have the same effect. In an extreme case, one could imagine a ‘robot criminal’ being destroyed. But would this be desirable and would it be effective?
The most commonly articulated reasons for criminal punishment are retribution, incapacitation, deterrence, and rehabilitation.Footnote 58 Retribution is the oldest reason for punishment, sublimating the victim's desire for revenge into a societal demonstration that wrongs have consequences.Footnote 59 Calibration of those consequences was at its most literal in the lex talionis: an eye for an eye, a tooth for a tooth. The demonstrative effect of fining a corporation—or an electronic ‘person’—may be preferable to a crime otherwise going unpunished.Footnote 60
The penal system can also be used to incapacitate those convicted of crimes, physically preventing them from reoffending. Typically this is through varying forms of incarceration, but may also include exile, amputation of limbs, castration, and execution. In the case of corporations, it may include withdrawal of a licence to operate or a compulsory winding up order.Footnote 61 Here direct analogies with the treatment of dangerous animals and machinery can be made, although measures such as putting down a vicious dog or decommissioning a faulty vehicle are administrative rather than penal and do not depend on determinations of ‘guilt’.Footnote 62 In some jurisdictions, children and the mentally ill may be deemed incapable of committing crimes, yet they may still be detained by the State if judged to be a danger to themselves or the community.Footnote 63 Such individuals do not lose their personality; in the case of AI systems, it is not necessary to give them personality in order to impose measures akin to confinement if a product can be recalled or a licence revoked.
Deterrence is a more recent justification for punishment, premised on the rationality of offenders. By structuring penalties, it imposes costs on behaviour that are intended to outweigh any potential benefits. The ability to reduce criminality to economic analysis may seem particularly applicable to both corporations and AI systems. Yet in the case of the former the incentives are really aimed at human managers who might otherwise act in concert through the corporation for personal as well as corporate gain.Footnote 64 In the case of an AI system, the deterrent effect of a fine would shape behaviour only if its programming sought to maximise economic gain without regard for the underlying criminal law itself.
A final rationale for punishment is rehabilitation. Like incapacitation and deterrence, it is forward-looking and aims to reduce recidivism. Unlike incapacitation, however, it seeks to influence the decision to offend rather than the ability to do so;Footnote 65 unlike deterrence, that influence is intended to operate intrinsically rather than extrinsically.Footnote 66 Rehabilitation in respect of natural persons often appears to be embraced more in theory than in practice; in the United States in particular, it fell from favour in the 1970s.Footnote 67 With respect to corporations, however, the clearer levers of influence have encouraged experimentation with narrowly tailored penalties that aim to encourage good behaviour as well as discourage bad.Footnote 68 Such an approach might seem well suited to AI systems, with violations of the criminal law being errors to be debugged rather than sins to be punished.Footnote 69 Indeed, the educative aspect of rehabilitation has been directly analogised to machine learning in a book-length treatise on the topic.Footnote 70 Yet neither legal personality nor the coercive powers of the State should be necessary to ensure that machine learning leads to outputs that do not violate the criminal law.
C. No Soul to Be Damned
While arguments justifying liability of corporations tend to be instrumental, it is striking how the emerging literature on ‘robot criminals’ slides into anthropomorphism. The very term suggests a special desire to hold humanoid AI systems to a higher standard than, say, household appliances with varying degrees of autonomy or unembodied AI systems operating in the cloud.Footnote 71 There is no principled reason for such a distinction, but it speaks to the tension within arguments for AI personality that blend instrumental and inherent justifications.
Interestingly, arguments over the juridical personality of corporations tend to focus on the opposite problem: their dissimilarity to humans, pithily described by the First Baron Thurlow as them having ‘no soul to be damned, and no body to be kicked’.Footnote 72 The lack of a soul has not impeded juridical personality of corporations and poses no principled barrier to treating AI systems similarly. Corporate personality is different from AI personality, however, in that a corporation is made up of human beings, through whom it operates, whereas an AI system is made by humans.Footnote 73
Instrumental reasons could, therefore, justify according legal personality to AI systems. But they do not require it. The implicit anthropomorphism elides further challenges such as defining the threshold of personality when AI systems exist on a spectrum, as well as how personality might apply to distributed systems. It would be possible, then, to create legal persons comparable to corporations—each autonomous vehicle, smart medical device, resume-screening algorithm, and so on could be incorporated.Footnote 74 If there are true liability gaps then it is possible that such legal forms could fill them. Yet the more likely beneficiaries of such an arrangement would be producers and users, who would thus be insulated from some or all liability.
II. NATURAL PERSONALITY: COGITO, ERGO SUM?
Instrumentalism is not the only reason legal systems recognise personality, however. In the case of natural persons, no Turing Test needs to be passed: the mere fact of being born entitles one to personhood before the law.Footnote 75
It was not always thus. Through much of human history, slaves were bought and sold like property;Footnote 76 indigenous peoples were compared to animals roaming the land, justifying their dispossession;Footnote 77 and for centuries under English law, Blackstone's summary of the position of women held that ‘husband and wife are one person, and the husband is that person’.Footnote 78 Even today, natural persons enjoy plenary rights and obligations only if they are adults, of sound mind, and not incarcerated.
As indicated earlier, many of the arguments in favour of AI personality implicitly or explicitly assume that AI systems are approaching human qualities in a manner that would entitle them to comparable recognition before the law. Such arguments have been challenged both for their analysis and their implications. In terms of analysis, Neil Richards and William Smart have termed the tendency to anthropomorphise AI systems the ‘android fallacy’.Footnote 79 Experiment after experiment has shown that people are more likely to ascribe human qualities such as moral sensibility to machines on the basis of their humanoid appearance, natural language communication, or the mere fact of having been given a name.Footnote 80 More serious arguments about AI approximating human qualities are challenged due to unexamined assumptions about how those qualities manifest in humans ourselves.Footnote 81
In terms of the implications, the 2017 European Parliament resolution prompted hundreds of AI experts from across the continent to warn in an open letter that legal personality for AI would be inappropriate from ‘an ethical and a legal perspective’. Interestingly, such warnings may themselves fall foul of the android fallacy by assuming that legal status based on the natural person model necessarily brings with it all the ‘human’ rights guaranteed under EU law.Footnote 82 Other writers candidly admit that the only basis for denying AI systems personality may ultimately be a form of speciesism—privileging human welfare over robot welfare because we the lawmakers are human.Footnote 83 If AI systems become so sophisticated that this is our strongest defence, the problem may not be their legal status but our own.Footnote 84
This section nonetheless takes seriously the idea that certain AI systems might have an entitlement to personality due to their inherent qualities. The technical aspects of how those qualities might manifest—and indeed a detailed examination of the human qualities that they mimic—are beyond the scope of this article.Footnote 85 Instead, the focus will be on how and why natural personhood might be extended. A first question to examine is how this has been handled in the past, such as the enfranchisement and empowerment of natural persons long treated as inferior to white men. More recently, activists and scholars have urged further expansion of certain rights to non-human animals such as chimpanzees based on their own inherent qualities. The inquiry then turns to the strongest articulation today of meaningful rights on behalf of AI systems for inherent rather than instrumental reasons: that they should be able to own their creations.
A. The Extension of Natural Personality
The arc of the moral universe is long, as Dr Martin Luther King Jr famously intoned, but it bends towards justice. At the time that the United States drafted its Declaration of Independence in 1776, the notion that ‘all men’ [sic] were ‘created equal’ was demonstrably untrue. A decade later, the French Declaration on the Rights of Man similarly proclaimed natural and imprescriptible rights for all—‘nonsense upon stilts’ was Jeremy Bentham's observation.Footnote 86 Man may well be born free, as Rousseau had opined in the opening lines of The Social Contract, but everywhere he remained in chains.Footnote 87
And yet the succeeding centuries did see a progressive realisation of those lofty aspirations and the spread of rights. By the middle of the twentieth century, the Universal Declaration of Human Rights could claim that all human beings were ‘born free and equal in dignity and rights’, despite one-third of them living in territories that the UN itself classified as non-self-governing. Decolonisation, the end of apartheid, women's liberation and other movements followed; rights remain a site of contestation, but virtually no State today would seriously contend that human adults are not persons before the law.Footnote 88
Interestingly, some arguments in favour of legal personality for AI draw not on this progressivist narrative of natural personhood but on the darker history of slavery. Andrew Katz and Ugo Pagallo, for example, find analogies with the ancient Roman law mechanism of peculium, whereby a slave lacked legal personality and yet could operate as more than a mere agent for his master.Footnote 89 As an example of a creative interpretation of personhood it is interesting, though it relies on instrumental justifications rather than the inherent qualities of slaves. As Pagallo notes, peculium was in effect a sort of ‘proto-limited liability company’.Footnote 90 From the previous section, there is no bar on legal systems creating such structures today—as for whether they should do so, reliance upon long discarded laws associated with slavery may not be the strongest case possible.Footnote 91
An alternative approach is to consider how the legal system treats animals.Footnote 92 For the most part, they are regarded as property that can be bought and sold, but also as deserving of ‘humane’ treatment.Footnote 93 Liability of owners for damage caused by animals has limited application to AI systems;Footnote 94 here, the question is whether those animals might ‘own’ themselves.
Various efforts have sought to attribute degrees of personality to nonhuman animals, with little success. In 2013, for example, the Nonhuman Rights Project filed lawsuits on behalf of four captive chimpanzees, arguing that the animals exhibited advanced cognitive abilities, autonomy, and self-awareness. In denying writs of habeas corpus, the New York State Court Appellate Division did not dispute these qualities but held that extension of rights such as personality had traditionally been linked to the imposition of obligations in the form of a social contract. Since, ‘needless to say’, the chimpanzees could not bear any legal duties, they could not enjoy rights of personality such as the right to liberty.Footnote 95 This was a curious basis on which to dismiss the case, as many humans who lack the capacity to exercise rights or responsibilities—infants, persons in a coma—are nonetheless deemed persons before the law.Footnote 96 A parallel case rejected that argument on the circular basis that it ‘ignores the fact that these are still human beings, members of the human community’.Footnote 97 Leave to appeal to the State Court of Appeals was denied, but one of the judges issued a concurring opinion that ended on a speculative note about the future of such litigation. The issue, Judge Fahey observed, is profound and far-reaching. ‘Ultimately, we will not be able to ignore it. While it may be arguable that a chimpanzee is not a “person”, there is no doubt that it is not merely a thing.’Footnote 98
Gabriel Hallevy has argued that animals are closer than AI systems to humans when one considers emotionality as opposed to rationality, but that this has not generally led to them being given personhood under the law. Instead, it is AI systems’ rationality that provides the basis for personhood.Footnote 99 That may be true with regard to the ability to make out the mental element of a criminal offence, but the fact that it is a crime to torture a chimpanzee but not a computer also points to an important difference in how the legal system values the two types of entity. In fact, a stronger argument may be made to protect embodied AI systems that evoke emotional responses on the part of humans—regardless of the sophistication of their internal processing. It seems probable that laws to protect such ‘social robots’ will at some point be adopted, comparable to animal abuse laws. As in the case of those laws, protection will likely be guided by social mores rather than consistent biological—or technological—standards.Footnote 100
The assumption that natural legal personality is limited to human beings is so ingrained in most legal systems that it is not even articulated.Footnote 101 The failure to extend comparable rights even to our nearest evolutionary cousins bodes ill for advocates of AI personality based on presumed inherent qualities.Footnote 102
B. Rewarding Creativity
A distinct reason for considering whether AI systems should be recognised as persons focuses not on what they are but what they can do. For the most part, this is framed as the question of whether an individual or corporation can claim ownership of work done by an AI system. Implicit or explicit in such discussions, however, is the understanding that if such work had been done by a human then he or she would own it him- or herself.
There is, in fact, a long history of questioning whether machine-assisted creation is protectable through copyright.Footnote 103 Early photographs, for example, were not protected because the mere capturing of light through the lens of a camera obscura was not regarded as true authorship.Footnote 104 It took an iconic picture of Oscar Wilde going all the way to the US Supreme Court before copyright was recognised in mechanically produced creations.Footnote 105 The challenge today is distinct: not whether a photographer can ‘own’ the image passively captured by a machine, but who might own new works actively created by one. A computer program like a word processor does not own the text typed on it, any more than a pen owns the words that it writes. But AI systems now write news reports, compose songs, paint pictures—these activities generate value, but can and should they attract the protections of copyright law?
In most jurisdictions, the answer is no.
The US Copyright Office, for example, has stated that legislative protection of ‘original works of authorship’Footnote 106 is limited to works ‘created by a human being’. It will not register works ‘produced by a machine or mere mechanical process that operates randomly or automatically without any creative input or intervention from a human author.’Footnote 107 The word ‘any’ is key and begs the question of what level of human involvement is required to assert authorship.Footnote 108
Consider the world's most famous selfie—of a black crested macaque. David Slater went to Indonesia to photograph the endangered monkeys, which were too nervous to let him take close-ups. So he set up a camera that enabled them to snap their own photos.Footnote 109 After the images gained significant publicity, animal rights activists argued that the monkeys had a greater claim to authorship of the photographs than the owner of the camera. Slater did eventually win, reflecting existing lawFootnote 110—though as part of a settlement he agreed to donate 25 per cent of future royalties from the images to groups protecting crested macaques.Footnote 111 As computers generate more content independently of their human programmers, it is going to be harder and harder for humans to take credit. Instead of training a monkey how to press a button, it may be more like a teacher trying to take credit for the work of his or her student.
Turning to the normative question of whether AI systems themselves should have a claim to ownership, the policy behind copyright is often articulated as incentivising innovation. This has long been seen as unnecessary or inappropriate for computers. ‘All it takes,’ Pamela Samuelson wrote in 1986, ‘is electricity (or some other motive force) to get the machines into production’.Footnote 112 Here the Turing Test offers a different kind of thought experiment: the more machines are designed to copy human traits, the more important such incentives might become.
Until recently, China followed the orthodoxy that AI-produced work is not entitled to copyright protection.Footnote 113 In December 2019, however, a district court in China held that an article produced by an algorithm could not be copied without permission. The article in question was a financial report published by Tencent with a note that it was ‘automatically written’ by Dreamwriter, a news writing program developed by the company in 2015. Shanghai Yingxun Technology Company copied the article without permission and Tencent sued. The article had been taken down, but the infringing company was ordered to pay ¥1,500 (US$216) for ‘economic losses and rights protection’.Footnote 114
The Chinese case reflects a distinct reason for recognising copyright, which is the protection of upfront investment in creative processes. This account presumes that, in the absence of such protection, investment will dry up and there will be a reduced supply of creative works.Footnote 115 Such an approach to copyright is broadly consistent with common law doctrines concerning work created in the course of employment, known in the United States as work for hire, under which a corporate employer or an individual who commissions a work owns copyright despite the actual ‘author’ being someone else.Footnote 116 This may not be available in civil law jurisdictions that place a greater emphasis on the moral rights of a human author.Footnote 117
In Britain, legislation adopted in 1988 does in fact provide copyright protection for ‘computer-generated’ work, the ‘author’ of which is deemed to be the person who undertook ‘the arrangements necessary for the creation of the work’.Footnote 118 Similar legislation has been adopted in New Zealand,Footnote 119 India,Footnote 120 Hong Kong,Footnote 121 and Ireland.Footnote 122 Though disputes about who took the ‘arrangements necessary’ may arise, ownership by a recognised legal person or by no one at all remain the only possible outcomes.Footnote 123
The European Parliament in April 2020 issued a draft report arguing that AI-generated works could be regarded as ‘equivalent’ to intellectual works and therefore protected by copyright. It opposed giving personality of any kind to the AI itself, however, proposing that ownership instead vest in ‘the person who prepares and publishes a work lawfully, provided that the technology designer has not expressly reserved the right to use the work in that way’.Footnote 124 The ‘equivalence’ to intellectual work is interesting, justified here on the basis of a proposed shift in recognising works based on a ‘creative result’ rather than a creative process.Footnote 125
For the time being, then, copyright cannot be owned by AI systems—and does not need to be in order to recognise the creativity of those systems. Nevertheless, reservations as to ownership being claimed by anyone else are evident in the limited rights given for ‘computer-generated’ works. The duration is generally for a shorter period, and the deemed ‘author’ is unable to assert moral rights—such as the right to be identified as the author of the work.Footnote 126 A World Intellectual Property Organization (WIPO) issues paper recognised the dilemma, noting that excluding such works would favour ‘the dignity of human creativity over machine creativity’ at the expense of making the largest number of creative works available to consumers. A middle path, it observed, might be to offer ‘a reduced term of protection and other limitations’.Footnote 127
C. Protecting Inventors
Whereas in copyright law the debate is over who owns works produced by AI systems, in patent law the question is whether they can be owned at all. Patent law in most jurisdictions provides or assumes that an ‘inventor’ must be human. In July 2019, Stephen Thaler decided to test those assumptions, filing patents in Britain, the European Union, and the United States that listed an AI system, DABUS, as the ‘inventor’.Footnote 128 The British Intellectual Property Office was willing to accept that DABUS created the inventions, but relevant legislation required that an inventor be a natural person and not a machine.Footnote 129 The European Patent Office (EPO) followed a more circuitous route to the same end, rejecting the applications on the basis that designating a machine as the inventor did not meet the ‘formal requirements’. These included stating the ‘family name, given names and full address of the inventor’.Footnote 130 A name, the EPO observed, does not only identify a person: it enables them to exercise their rights and forms part of their personality. ‘Things’, by contrast, ‘have no rights which a name would allow them to exercise.’Footnote 131
The US application was also rejected, based in part on the fact that relevant statutes repeatedly referred to inventors using ‘pronouns specific to natural persons’ such as ‘himself’ and ‘herself’. The Patent and Trade Office (USPTO) cited cases holding that conception—‘the touchstone of inventorship’—is a ‘mental act’ that takes place in ‘the mind of the inventor’. Those cases concluded that invention in this sense is limited to natural persons and not corporations. The USPTO concluded that an application listing an AI as an ‘inventor’ was therefore incomplete, but was careful to avoid making any determination concerning ‘who or what’ actually created the inventions in question.Footnote 132
These decisions were consistent with case law and the practice of patent offices around the world, none of which—yet—allows for an AI system to be recognised as an inventor. Analogous to copyright law, one purpose of the patent system is to encourage innovation by granting a time-limited monopoly in exchange for public disclosure. As even the creators of DABUS acknowledged, an AI system is unlikely to be motivated to innovate by the prospect of patent protection. Any such motivation would be found in its programming: it must be instructed to innovate.Footnote 133
As for whether a human ‘inventor’ could be credited for work done by such a system, there is no equivalent of the work for hire doctrine. To be an inventor, the human must have actually conceived of the invention.Footnote 134 Joint inventions are possible and contributions do not need to be identical, but in the absence of a natural person making a significant conceptual contribution an invention would be, on current law, ineligible for patent protection.Footnote 135
Among the interesting aspects of these recent developments are the means by which the same conclusion was reached. As in the case of copyright, there appeared to be no serious doubt that an AI system was capable of creating things that were patentable inventions. The British IP Office explicitly accepted that DABUS had done just that;Footnote 136 the USPTO was at pains to avoid making such a conclusion explicit. The EPO, for its part, dodged the issue by holding that, because they have no legal personality, ‘AI systems and machines cannot have rights that come from being an inventor’.Footnote 137 The EPO was the most blatant, but all three decisions relied on formalism—language in the relevant statute that provided or implied that the rights in question were limited to natural persons. The USPTO diligently dusted off a copy of Merriam-Webster's Collegiate Dictionary to conclude that the use of ‘“whoever” suggests a natural person.’Footnote 138
Legal tribunals routinely face choices between substantive justice and procedural regularity. Statutes of limitations are intended to provide certainty in legal relations; the courts of equity emerged to temper that certainty with justice. In both copyright and patent law, continuing to privilege human creativity over its machine equivalent may ultimately need to be justified by the kind of speciesism mentioned earlier.Footnote 139 At times, however, the language used to engage in such rationalisations of the status quo echoes older legal forms that kept property relations in their rightful place. Among its reasons denying that AI systems like DABUS could hold or transfer rights to a patent, for example, the EPO dismissed analogies between machines and employees: ‘Rather than being employed,’ the EPO concluded, ‘they are owned.’Footnote 140
Such statements are accurate for the time being. But if the boosters of general AI are correct and some form of sentience is achieved, the more appropriate analogy between legal personality and slavery may not be the limited economic rights slaves held in ancient Rome. Rather, it may be the constraints imposed on AI systems today.
III. CONSTRAINING SUPERINTELLIGENCE
If AI systems do eventually match human intelligence it seems unlikely that they would stop there. The prospect of AI surpassing human capabilities has long dominated a popular sub-genre of science fiction.Footnote 141 Though most serious researchers do not presently see a pathway even to general AI in the near future, there is a rich history of science fiction presaging real world scientific innovation.Footnote 142 Taking Nick Bostrom's definition of superintelligence as an intellect that greatly exceeds human cognitive performance in virtually all relevant domains,Footnote 143 it is at least conceivable that such an entity could be created within the next century.Footnote 144
The risks associated with that development are hard to quantify.Footnote 145 Though a malevolent superintelligence bent on extermination or enslavement of the human race is the most dramatic scenario, more plausible ones include a misalignment of values, such that the ends desired by the superintelligence conflict with those of humanity, or a desire for self-preservation, which could lead such an entity to prevent humans from being able to switch it off or otherwise impair its ability to function. An emerging literature examines these questions of what final and instrumental goals a superintelligence might have,Footnote 146 though the discourse was long dominated by voices far removed from traditional academia.Footnote 147 Visa Kurki's recent book-length discussion of legal personality, for example, includes a chapter specifically on personality of AI that concludes with the statement: ‘Of course, if some AIs ever become sentient, many of the questions addressed in this chapter will have to be reconsidered.’Footnote 148
In the face of many unknown unknowns, two broad strategies have been proposed to mitigate the risk. The first is to ensure any such entity can be controlled, either by limiting its capacities to interact with the world or ensuring our ability to contain it, including a means of stopping it functioning: a kill switch.Footnote 149 Assuming that the system has some kind of purpose, however, that purpose would most likely be best served by continuing to function. In a now classic thought experiment, a superintelligence tasked with making paperclips could take its instructions literally and prioritise that above all else. Humans who might decide to turn it off would need to be eliminated, their atoms deployed to making ever more paperclips.Footnote 150
Arguments that no true superintelligence would do anything quite so daft rely on common sense and anthropomorphism, neither of which should be presumed to be part of its code. A true superintelligence would, moreover, have the ability to predict and avoid human interventions or deceive us into not making them.Footnote 151 It is entirely possible that efforts focused on controlling such an entity may bring about the catastrophe that they are intended to prevent.Footnote 152
For this reason, many writers prioritise the second strategy, which is to ensure that any superintelligence is aligned with our own values—emphasising not what it could do, but what it might want to do. This question has also fascinated science fiction writers, most prominently Isaac Asimov, whose three laws of robotics are a leitmotif in the literature.Footnote 153 Here the narrower focus is on whether granting AI systems legal personality in the near term might serve as a hedge against the risks of superintelligence emerging in the future.
This is, in effect, another instrumental reason for granting personality. Of course, there is no reason to assume that including AI systems within human social structures and treating them ‘well’ would necessarily lead to them reciprocating the favour should they assume dominance.Footnote 154 Nevertheless, presuming rationality on the part of a general AI, various authors have proposed approaches that amount to socialising AI systems to human behaviour.Footnote 155 To avoid the sorcerer's apprentice problem of a machine simply told to ‘make paperclips’, for example, its goals could be tied to human preferences and experiences. This might be done by embedding those values within the code of such systems prior to them achieving superintelligence. Goals would thereby be articulated not as mere optimisation—the number of paperclips produced, for example—but as fuzzier objectives such as maximising the realisation of human preferences,Footnote 156 or inculcating a moral framework and a reflective equilibrium that would match the progressive development of human morality itself.Footnote 157
To a lawyer, that sounds a lot like embedding these new entities within a legal system.Footnote 158 If one of the functions of a legal system is the moral education of its subjects, it is conceivable that including AI in this way could contribute to a reflective equilibrium that might encourage an eventual superintelligence to embrace values compatible with our own.
For the present, this is proposed more as a thought experiment than a policy prescription. If a realistic path to superintelligence emerges, it might become a more urgent concern.Footnote 159 There is no guarantee that the approach would be effective, but some small comfort might be taken from the fact that, for the most part, the categories of legal persons recognised in most jurisdictions, along with the rights they enjoy, have tended to expand over time rather than contract. Those regimes that have taken such legal recognition away have tended to be among the most despicable. Apocalyptic scenarios aside, positioning ourselves and our silicon siblings as equals might serve the goal of reinforcing a normative regime in which our interests are aligned, or at least not opposed, if or when we are surpassed.Footnote 160
In the alternative, like the chimpanzees in their New York cages, humanity's greatest hope may not be to be treated as peers, but at least to be seen as more than things.
IV. CONCLUSION: THE LIMITS OF PERSONALITY
In 1991 a prize was established to encourage more serious attempts at the Turing Test. One of the first winners succeeded in part by tricking people—the program made spelling mistakes that testers assumed must have been the result of human fallibility.Footnote 161 Though the Turing Test remains a cultural touchstone, it is far from the best measure of AI research today. As a leading textbook notes, the quest for flight succeeded when the Wright brothers stopped trying to imitate birds and started learning about aerodynamics.Footnote 162 Aeronautical engineers today don't define the goal of their field as making machines that fly so exactly like pigeons that they can fool other pigeons.
In the same way, most arguments in favour of AI legal personality suffer from being both too simple and too complex. They are too simple in that AI systems exist on a spectrum with blurred edges. There is as yet no meaningful category that could be identified for such recognition; if instrumental reasons required recognition in specific cases then this could be achieved using existing legal forms. The arguments are too complex in that many are variations on the android fallacy, based on unstated assumptions about the future development of AI systems for which personality would not only be useful but deserved. At least for the foreseeable future, the better solution is to rely on existing categories, with responsibility for wrongdoing tied to users, owners, or manufacturers rather than the AI systems themselves. Driverless cars are following that path, for example, with a likely shift from insuring drivers to insuring vehicles.Footnote 163
This may change. It is conceivable that synthetic beings of comparable moral worth to humans may one day emerge. Failing to recognise that worth may reveal us to be either an ‘autistic species’, unable to comprehend the minds of other types of beings,Footnote 164 or merely prejudiced against those different from ourselves. If this happens, as Turing hypothesised in 1951, ‘it seems probable that, once the machine thinking method had started, it would not take long to outstrip our feeble powers’.Footnote 165
Turing himself never lived to see a computer even attempt his test in practice. Prosecuted for homosexual acts in 1952, he chose chemical castration as an alternative to prison. He died two years later at the age of 41, apparently after committing suicide by eating a cyanide-laced apple. The announcement that Turing would grace the new 50 pound note followed an official pardon, signed by the Queen in 2013.Footnote 166
Yet the more fitting tribute may be Ian McEwan's novel, Machines Like Me, which imagines an alternative timeline in which Turing lived and was rewarded with the career and the knighthood he deserved. The novel takes seriously the prospect of true AI, in the form of a brooding synthetic Adam, who expresses his love for the human Miranda by writing thousands upon thousands of haikus. Ultimately, however, consciousness is a burden for the machines—struggling to find their place in the world; so pure that they are unable to reconcile human virtues and human vices.
It also offers Turing a chance to rethink his test. ‘In those days,’ the fictional Turing says at age 70, referring to his younger self, ‘I had a highly mechanistic view of what a person was. The body was a machine, an extraordinary one, and the mind I thought of mostly in terms of intelligence, which was best modelled by reference to chess or maths.’Footnote 167
The reality, of course, is that chess is not a representation of life. Life is an open system; it is messy. It is also unpredictable. In the novel, the first priority of the AI robots is to disable the kill switch that might shut them down. Yet most of them ultimately destroy themselves—as the real Turing did—unable to reconcile their innate nature with the injustices of the world around them. Before asking whether we could create such thinking machines, McEwan reminds us, we might want to pause and ask whether we should.