1. From Epistemic Paternalism to Paternalistic Cognitive Engineering

In what follows I delineate a pro tanto problematic variety of epistemic paternalism that the relevant conceptual engineering projects plausibly belong to. Merely showing that these projects involve a form of epistemic paternalism in some sense of that term would not raise a worry for the projects. It is possible, after all, to define epistemic paternalism in a way that does not isolate a set of (even pro tanto) problematic practices. Indeed, Kristoffer Ahlstrom-Vij's (Reference Ahlstrom-Vij2013) influential account of epistemic paternalism arguably defines the term in such a way. According to Ahlstrom-Vij (Reference Ahlstrom-Vij2013: 38–61), the necessary and sufficient conditions for an epistemically paternalistic practice are the following:

1. 1. Interference: The agent's ability to go about their inquiry (to access, collect, and evaluate information) in whatever way they see fit is interfered with.

2. 2. Non-consultation: The agent is not consulted, meaning that either they are not asked about the interference, or they are asked but then not listened to.

3. 3. Improvement: The aim is to improve the agent's epistemic situation (which may be instrumental for further non-epistemic goals).

For example, this definition covers the central case of epistemic paternalism discussed by Alvin Goldman (Reference Goldman1991): withholding true and relevant but nevertheless misleading evidence (such as information about the defendant's past crimes) from jurors. Such evidence suppression is expected to improve the jurors’ cognitive performance because people tend to attribute too much weight to the kinds of evidence in question. The practice satisfies all three conditions: it interferes with the jurors’ ability to inquire in whatever way they see fit; the jurors are not consulted; and the aim is to help them reach the true verdict—their own epistemic good—which is instrumental for further aims, such as justice being served.

Defined this way, epistemic paternalism arguably includes many entirely unproblematic practices. As Fabien Medvecky (Reference Medvecky, Bernal and Axtell2020) points out, everyday testimony seems to fit the bill. We choose what to say and what to leave out, so that our audience gets the knowledge that fits their needs and interests. For example, when telling someone the time, we often round up to the nearest minute or more, thus interfering with the hearer's inquiry into the current time with their epistemic well-being in mind and without consulting them. Further, even writing a nonfiction book might count as non-consultatively interfering with its readers’ inquiry for their own epistemic good (Jackson Reference Jackson, Matheson and Lougheed2021: 135).

Ahlstrom-Vij meant to capture the problematic aspect of epistemic paternalism in the non-consultation condition, which should explain why there is ‘something arrogant about paternalistic interference; a distinct kind of disregard for the wants, preferences or opinions of those interfered with’ (Ahlstrom-Vij Reference Ahlstrom-Vij2013: 43). But as the everyday testimony and book-writing examples show, non-consultation as such is not a problem if the interference is mild enough. We need not obtain consent every time we help people out in small ways.

Medvecky takes such examples (in particular, about everyday testimony and science communication) to show that epistemically paternalistic practices are generally just fine. Elizabeth Jackson (Reference Jackson, Matheson and Lougheed2021), by contrast, takes cases of innocuous (supposed) epistemic paternalism to show the inadequacy of the definition of epistemic paternalism that includes these cases. For my purposes, the discussion on how to define epistemic paternalism can be largely left aside. Merely as a convenient stipulation (which will not matter for the argument to come), I reserve the term for (roughly) the broad phenomenon identified by Ahlstrom-Vij (Reference Ahlstrom-Vij2013). I am concerned with a subset of pro tanto problematic practices within such broadly construed epistemic paternalism—a subset that includes the relevant projects of conceptual engineering. I call this subset paternalistic cognitive engineering. The practices in this subset have the following features:

1. 1. Objectionable cognitive engineering: The practice involves cognitive engineering, that is, deliberately and reflectively shaping an agent's belief-formation in light of assumptions about their cognitive system. Further, the cognitive engineering is objectionable in the sense that the agent could reasonably regard it as a violation of their sovereignty over their belief formation (assuming that they are not consulted).

2. 2. Non-consultation: The agent is not consulted (in the same sense as in Ahlstrom-Vij's [Reference Ahlstrom-Vij2013] account).

3. 3. Improvement: The aim is to improve the agent's epistemic situation. As in Ahlstrom-Vij's account, this may be instrumental for further non-epistemic purposes. (But note that I understand epistemic more broadly than does Ahlstrom-Vij [Reference Ahlstrom-Vij2013]; in particular, attending to important matters counts as epistemic improvement.)

The non-consultation and improvement conditions are the same as in Ahlstrom-Vij's (Reference Ahlstrom-Vij2013) account, then. The objectionable cognitive engineering condition needs unpacking. I use cognitive engineering similarly to the way Daniel Hausman and Brynn Welch (Reference Hausman and Welch2010) use shaping in their critical discussion of nudging. Nudging, a notion introduced and popularized by Richard Thaler and Cass Sunstein (Reference Thaler and Sunstein2008), means making subtle changes to the agents’ environment to influence their choices non-coercively, often for their own good; for example, placing healthy food at eye level or setting relatively high default values of retirement savings. Hausman and Welch point out that some nudges involve shaping in the sense of ‘the use of flaws in human decision-making to get individuals to choose one alternative rather than another’, contrasting such shaping with rational persuasion (Hausman and Welch Reference Hausman and Welch2010: 128). However, cognitive engineering, as I use the term, need not use flaws in human decision making: it can rely on other assumptions about human cognition or specifically the target agents’ cognition. Another difference is that I emphasize the shaping of belief formation rather than choices. But there might not be a principled distinction here: good choices may be construed as true beliefs about what is to be done.

Cognitive engineering is reflective and deliberate. For example, we do not engage in cognitive engineering when we unreflectively round off the time for the hearer's epistemic good—though we may be unconsciously relying on knowledge about human cognitive limitations and the hearer's needs. This is not to say that if we were to consciously consider the hearer's limitations and needs, and deliberately round off the time in this light, this would be paternalistic cognitive engineering. Whether or not it makes sense to count this as cognitive engineering, the objectionable part of the first condition is unsatisfied: it would not be reasonable for the hearer to regard our act as a sovereignty violation.

The objectionable part of the first condition appeals to the familiar idea of personal sovereignty, ‘the idea of having a domain or territory in which the self is sovereign’ (Feinberg Reference Feinberg1983: 452). As Joel Feinberg further explains this, ‘to say that I am sovereign over my bodily territory is to say that I, and I alone, decide (so long as I am capable of deciding) what goes on there’; and what falls within one's sovereign domain ‘cannot be treated in certain ways without one's consent’ (Reference Feinberg1983: 453). The violations that we are here concerned with, however, are not essentially interferences with the body or its surrounding space but rather with the domain of belief formation. John Stuart Mill emphasizes sovereignty over both the body and the mind: ‘over himself, over his own body and mind, the individual is sovereign’ ([1859] Reference Mill and Alexander1999: 52).

The idea, then, is that we all (or at least the reasonably competent adult humans among us) have our territory of belief formation that we are entitled to govern ourselves and that should not be treated in certain ways without our consent. In paternalistic cognitive engineering, the engineers non-consultatively exercise control over the target agents’ domain of belief formation in a way that these agents could reasonably regard as intruding into that territory and thus as violating their sovereignty over their own intellectual domain, their own mind. That is a pro tanto reason against such engineering.

I have purposefully used a cautious formulation, appealing to how the target agents could reasonably assess the practice, instead of simply asserting that the relevant practices are sovereignty violating. The reason for the cautious formulation is that people differ widely in what practices they would regard as intruding into their territory; and it is plausible that there is a range of reasonable attitudes rather than a clear division between objectively intrusive and nonintrusive practices. For example, some people regard certain benevolent nudges (such as healthy food placed at eye level) as disturbingly controlling their choices, while others do not mind or even appreciate the nudges. I acknowledge a range of reasonable attitudes, which leads me to refrain from claims about objective sovereignty violation.

At the same time, we should not classify practices as pro tanto problematic just based on the recipients’ actual or hypothetical judgment: people sometimes unreasonably feel intruded. For example, people can feel intruded when we politely challenge their political convictions with good arguments; and this is not a sovereignty-based reason against such polite rational persuasion. So, objectionability is not a matter of whether the target agents would or could assess the practice as sovereignty violating; it depends on whether regarding the practice as sovereignty violating falls within the range of reasonable attitudes for the target agents (were they to find out afterwards, for example, or were they to discover that such a practice is being planned).

Further, I assume that it is reasonable for the target agents to regard a practice as sovereignty violating when the paternalistic actors exercise significant control over the target agents’ belief formation by cognitive engineering, while it would in principle be possible to persuade, inform, or educate the agents explicitly instead. This to some extent draws on and generalizes Hausman and Welch's criticism of nudges of the shaping variety. According to Hausman and Welch (Reference Hausman and Welch2010), such nudges infringe on autonomy in the sense of ‘the control an individual has over her own evaluation, deliberation and choice’ (128). Granted, some people might not care much about controlling their own intellectual domain, and this attitude might be reasonable. However, I assume (not outrageously, I hope) that it is also reasonable to wish that others did not exercise significant control by cognitive engineering—especially when it is possible instead to engage in explicit persuading, informing, and educating, thus giving the target agents more control.

For example, suppose that a doctor consciously uses a framing effect to nudge the patient to decide in favor of an operation, for the patient's own good, saying, ‘The chances of survival are 99 percent’ (rather than ‘There is 1 percent chance of dying’). The doctor could have instead stated the recommendation explicitly, along with the statistical information and the reasoning behind the recommendation, thus giving the patient more control over the decision. Since this alternative approach is available, the patient (were they to find out about the intervention) could reasonably think that the doctor is exercising too much control over the patient's intellectual domain.

For another example, Evan Riley (Reference Riley2017: 610) contrasts two ways of combatting low electoral turnouts: (1) implementing a program of beneficent nudges and (2) setting aside three days before the election for public discussion and debate to raise the community's level of political engagement. In this case, too, exercising control by nudging seems problematic because there is a less controlling method available for pursuing the target agents’ own good; and those agents could reasonably wish that this method were used instead of cognitive engineering.

An important clarification is that explicit rational persuasion, informing, and educating can be available even when they are more difficult, costly, and time consuming to implement than cognitive engineering, or not quite as effective. For example, consider the practice of evidence suppression, which relies on assumptions about the target agents’ cognition (the relevant biases) to shape belief forming, and is thus a variety of cognitive engineering. An alternative approach available is educating the target agents about their relevant biases and suggesting strategies to counteract bias. That alternative approach would give these agents more control over their intellectual domain, but it would also be much more costly and time-consuming and might ultimately leave the agents worse off, in terms of reliable belief formation, than would external evidence curation. Nevertheless, the educating approach is plausibly available enough for the target agents to regard reasonably the non-consultative evidence suppression as a sovereignty violation. (However, reasonably taking a practice to be sovereignty violating is not the same as reasonably taking it to be ultimately unjustified. Perhaps it is reasonable for the jurors, for example, to regard evidence suppression as sovereignty violating—so, there is a pro tanto reason against evidence suppression—while it is unreasonable for the jurors to regard the practice as ultimately unjustified, given the reasons in favor of the practice. I return to the issue of ultimate justification below, in section 3.)

Since I contrast paternalistic cognitive engineering with explicit rational persuasion, informing, and educating, it is also important to emphasize that the problem with paternalistic cognitive engineering is not that it bypasses rationality. Bypassing rationality means influencing choice, belief, and behavior without providing reasons. Neil Levy (Reference Levy2019, Reference Levy2022: 132–48) argues that nudges often provide implicit reasons and therefore do not bypass rationality. For example, according to Levy, when a doctor uses a framing effect to make salient the high probability of surviving the operation, the doctor is implicitly recommending the operation and is interpreted as doing so by the patient. Further, the doctor's opinion is a reason—moreover, a good reason—for the patient to decide in favor of the operation. Even if Levy is right that nudges typically provide reasons (see Grundmann Reference Grundmann2021a; Levy Reference Levy2021a; Grundmann Reference Grundmann2021b; Levy Reference Levy2021b), it could still be reasonable to regard nudging as sovereignty violating. For example, when the patient finds out that the doctor reflectively chose this framing to influence the patient's choice, the patient may well be reasonable in feeling that the doctor violated their sovereignty—not because the doctor failed to provide a reason, but because they offered a reason in a way that exercised too much control over the patient's decision. The doctor could have stated their recommendation explicitly, perhaps along with the reasoning behind it, giving the patient more control. The problem with paternalistic cognitive engineering is not bypassing rationality; the problem is exercising control over another person's intellectual domain in a way they could reasonably regard as sovereignty violating.

As the above examples show, paternalistic cognitive engineering—a pro tanto problematic (but perhaps not ultimately unjustified) variety of epistemic paternalism—plausibly includes some cases of nudging and evidence suppression. The category does not include some other typical examples of problematic epistemic paternalism, such as a doctor divulging medical information against the patient's stated will (Bullock Reference Bullock2018: 441)—there is no cognitive engineering here, and hence no paternalistic cognitive engineering. But this is fine: paternalistic cognitive engineering is not supposed to cover all problematic epistemic paternalism. I am interested only in the kind of (pro tanto) problematic epistemic paternalism that plausibly includes the relevant conceptual engineering projects. I now turn to these projects.

2. Conceptual Engineering as Paternalistic Cognitive Engineering

Consider, then, reflective and deliberate attempts to shape the target agents’ belief-formation by shaping their concepts. In order for these practices to involve paternalistic cognitive engineering, it must be reasonable for the target agents to view the practices as sovereignty violating. As suggested above, this is reasonable if the engineers could in principle explicitly persuade, inform, and educate instead of cognitive engineering, thus giving the target agents more control over their intellectual domain. Further, the practices must be non-consultative and aim at least instrumentally for the agents’ epistemic improvement, broadly construed. I argue that all the conditions are satisfied for some conceptual engineering projects and that the projects therefore involve paternalistic cognitive engineering. There is thus a sovereignty-based pro tanto reason against these projects.

Again, concepts, in the relevant sense, are psychological entities. One may further conceive of concepts, on this approach, as bodies of information (in the long-term memory) that are drawn on quickly, automatically, and context-independently to form judgments; these bodies of information might include exemplars, prototypes, and theory-like structures (Machery Reference Machery2017: 210–12). While concepts, thus understood, may be engineered for various aims, they are sometimes engineered (at least in the first instance) for broadly epistemic aims.

Some of the relevant projects exploit concepts’ capacity to underwrite inferences. As Edouard Machery (Reference Machery2017) puts it, ‘Concepts determine the inferences we are prone to draw, and our thoughts follow their tracks’ (231). Sometimes these are epistemically bad tracks, and we might then want to discourage the use of such concepts. For example, Robert Brandom suggests eliminating the slur boche (for Germans, especially German soldiers) because it underwrites unreliable inferences: ‘If one does not believe that the inference from German nationality to cruelty is a good one, then one must eschew the concept Boche’ (Reference Brandom1994: 126). Another example is Machery's (Reference Machery2021) case for eliminating the concept of innateness in science. The folk concept of innateness, according to Machery, promotes unreliable inferences by bundling together the notions of typicality, fixity, and functionality. Although scientists take themselves to be using a technical concept of innateness, they still unreflectively draw the unreliable inferences stemming from the folk concept. For this reason, Machery proposes dispensing with innateness talk in science altogether.

This is, then, the first type of project: eliminating concepts that underwrite unreliable inferences, or to put it simply, that bring to mind false ideas (such as false stereotypes) and make these ideas appear plausible. (See Fischer [Reference Fischer2020] for a more detailed, empirically informed explanation of how concepts underwrite inferences.) The elimination of an epistemically bad concept may be achieved by explicit guidelines, social norms and sanctions, or the influence of authoritative language users, for example. The idea is that changing the community's linguistic behavior would also change individuals’ habits of using concepts in their private cognition. If the concepts that support unreliable inferences cannot be completely weeded out in this manner, they might at least become less easily activated, and that is already an epistemic win. In the case of prejudice-evoking slurs, the epistemic wins from their elimination bring non-epistemic wins in their tow; but one may also want to eliminate concepts that support inferences to falsehoods just for the epistemic wins alone.

In other cases, the concept itself is fine and even commendable, but certain uses of it shape the inferences that the concept underwrites in problematic ways. Such uses can then be discouraged for the concept users’ own epistemic good. For example, consider Rachel Fraser's discussion of rape metaphors that make bad inferences involving the term rape more cognitively accessible and socially licensed and good inferences less accessible and less licensed. An inference is cognitively accessible when it ‘does not require significant cognitive labor’ and socially licensed when it is generally regarded as legitimate (Fraser Reference Fraser2018: 735–36). For example, saying ‘Germany is raping Brazil!’ while watching soccer can make the inferences from ‘S was raped’ to ‘S has been humiliated’ and ‘S should have put up more of a fight’ more cognitively accessible and socially licensed, while making the inference to ‘Serious injustice has been done to S’ less accessible and licensed (Fraser Reference Fraser2018: 745–46). From the epistemic perspective, such rape metaphors are problematic because they shape the concept of rape in a way that disposes the concept user to make inferences from truths to falsehoods and indisposes them to make certain inferences to truths. The concept users thereby become more likely to acquire false beliefs and miss important truths. Again, the problematic usage can be discouraged by establishing guidelines and social norms.

The previous examples concerned concepts underwriting inferences. Conceptual engineers can also exploit the capacity of entrenched terms to make phenomena psychologically salient. For example, consider Sally Haslanger's (Reference Haslanger2000) proposal of revising the terms man and woman to designate humans that are respectively advantaged or disadvantaged in virtue of their (real or imagined) features related to their role in reproduction. Different interpretations of this are possible; for example, Haslanger herself (Reference Haslanger2006) and Max Deutsch (Reference Deutsch2020a, Reference Deutsch2020b) argue that the proposal can be seen as one about what the terms already mean. However, another possible interpretation is that this conceptual change would facilitate a desirable change in patterns of attention: it would make the many forms of gender inequality more salient, thereby motivating progressive action. This interpretation is supported by Haslanger's (Reference Haslanger2000) remark that she is ‘asking us to understand ourselves and those around us as deeply molded by injustice and to draw the appropriate prescriptive inference’ (48).

Andy Clark and David Chalmers's (Reference Clark and Chalmers1998) defense of an extended notion of belief (which includes externally stored easily accessible information) can be understood similarly. (Here, too, it is unimportant whether the project I am describing exactly fits the authors’ intentions or is otherwise the best interpretation, as long as this is a somewhat compelling project in the vicinity. Indeed, in a recent essay, Chalmers [Reference Chalmers2020: 8] notes that they meant to give an account of what belief means, but he does not mind if it is thought of as a conceptual engineering project instead.) When we associate the entrenched term belief with the more natural kind in the relevant contexts, the more natural kind becomes salient for us in these contexts, facilitating effective explanation and prediction.

Eric Schwitzgebel (Reference Schwitzgebel, Borgoni, Kindermann and Onofri2021) explicitly makes this sort of case for adopting his preferred pragmatic concept of belief within philosophy. The pragmatic concept requires, in addition to intellectual endorsement, the relevant phenomenological and behavioral dispositions. According to Schwitzgebel, beliefs in that pragmatic sense are more important for philosophers to think about than mere intellectual endorsements. We should use the central term belief to guide attention to those more important phenomena and can use the less central term judgment to talk about mere intellectual endorsements.

One might question whether improving patterns of attention within philosophy or in society at large is an epistemic improvement and whether these salience-shaping projects are thus candidates for paternalistic cognitive engineering, as defined above. As also noted above, however, I construe epistemic broadly. This contrasts with some classics on epistemic paternalism, who focus on narrowly veritistic aims (Goldman Reference Goldman1991; Ahlstrom-Vij Reference Ahlstrom-Vij2013). More recently, the epistemic in epistemic paternalism has often been broadly construed, in line with my approach here. Moreover, such a broad construal allows us to discuss under the convenient heading epistemic paternalism issues that are closely related and naturally discussed together. There also seems to be no consensus usage of epistemic with which this approach would conflict. Allan Hazlett (Reference Hazlett2016) and Stewart Cohen (Reference Cohen2016) consider this lack of consensus and boundaries to be unfortunate; but for my present purposes, it comes in handy.

The relevant kind of conceptual engineer aims at the target agents’ epistemic good, broadly construed, by shaping their inferential and attentional dispositions. The improvement condition is thus satisfied: the interference aims at the target agents’ own epistemic good. Further, cognitive engineering is involved: the conceptual engineer who encourages or discourages certain concepts (or certain uses of certain concepts) deliberately and reflectively shapes belief formation in light of assumptions about how the target agents’ minds work. The assumptions concern both how concepts affect cognition generally—that they support inferences and alter salience—and how the concepts in question work (what inferences they support, how the various uses of concepts in turn shape these inferences, and what the concepts make salient). In light of such information, the conceptual engineer then shapes the target subjects’ belief formation—for example, by trying to remove undesirable (uses of) concepts from circulation or taking steps to ensure that preferred concepts are instilled.

Two questions remain. First, could the target agents reasonably regard the engineering as sovereignty violating? This is the case, recall, if the engineers could in principle use methods like explicit rational persuasion instead, thus giving the target agents more control over their intellectual domain. And second, is the practice non-consultative?

Regarding the first issue, let us consider engineering concepts to shape inference patterns—for example, eliminating the term boche in order to eliminate or weaken the inference from ‘S is German’ to ‘S is cruel’; or discouraging the relevant rape metaphors (‘Germany is raping Brazil!’ referring to a soccer match) to support the inference from ‘S was raped’ to ‘A great injustice has been done to S’. Could the engineers engage in explicit rational persuasion instead? There are some options to consider. For example, instead of removing stereotype-evoking expressions or the rape metaphors from circulation, we could undertake the more resource-demanding but arguably also more respectful project of reasoning with people about why the stereotypes are false or why rape is a great injustice. Directly addressing the mistaken ideas perpetuated by the linguistic practices (like slurs and rape metaphors) could then also lead to changes in those linguistic practices and the desired developments in concepts. Educating people about ethnic prejudice and debunking the mistaken ways of thinking about rape would plausibly leave our target agents more control over their intellectual domain without our having to give up on the aim of epistemic improvement.

For salience-altering conceptual engineering, there are also alternative approaches to consider. The non-consultative engineer shapes what the agents experience as salient in explaining, predicting, and interpreting, instead of rationally defending the importance of these phenomena. Such a rational defense must be possible, for the conceptual engineer to be able to justify the salience-altering program; but if the rational defense is possible, then why not offer that defense to the target agents, instead of just altering their experience by conceptual engineering? Consider Haslanger's (Reference Haslanger2000) proposal, as interpreted above. The engineer could educate the audience about how the world is thoroughly shaped by gender inequality and how to spot it in its different guises. Instead, the non-consultative engineer instils revisionary gender concepts to alert us automatically to gender inequality. The target agents, it seems, might well insist on the educating approach to retain more control over what they notice, think about, and feel motivated to do. Granted, the cognitive effects of the educating approach would not be exactly the same as those of the engineering approach; but it would serve the same general aim of promoting attention to the pervasiveness of gender inequality. Further, the educating approach may even include the suggestion to try looking at the world through the lens of the revisionary gender concepts—such a suggestion would not be the sort of non-consultative engineering that the educating approach contrasts with.

When we non-consultatively engineer other people's concepts to shape their patterns of inference and attention, they could reasonably regard this as a violation of their intellectual sovereignty, wishing that the engineers would not exercise control like this and would rather use explicit rational persuasion, informing, and educating. But are the proponents of conceptual engineering indeed envisioning these projects as non-consultative? To clarify what I mean by (non-)consultative: the consultative conceptual engineer offers a conceptual revision to the target agents as a way of improving their belief formation, along with an explanation of the aims and mechanisms involved. The agents are then free to take the advice or to refuse it. For example, the conceptual engineer could explicitly make a case to the target agents that the proposed concept would make them better at tracking the truth or that it would facilitate attending to important matters.

Indeed, this seems to be precisely what is going on in many prominent conceptual engineering proposals: the philosopher offers a way of revising a concept or suggests eliminating or introducing a concept, and the audience is free to take up the offer or refuse it. For example, Haslanger (Reference Haslanger2000) and Clark and Chalmers (Reference Clark and Chalmers1998) did not devise any clever schemes to make people adopt their preferred concepts; they just offered their arguments. But there is certainly a faction of conceptual engineers who think of the enterprise as more involved in implementing the proposals. For example, Rachel Sterken (Reference Sterken, Burgess, Cappelen and Plunkett2020) advocates linguistic transgressions—simply talking as if the word already had the desired meaning—as a potentially effective strategy. Or consider how Mark Pinder envisions changing speaker-meanings in larger communities: ‘So she should publicise the new definition, building public momentum for its adoption. At the outset, this may involve seeking to convince a high-profile group of users to defer to her definition. Then, over time, she might broaden that group, selectively targeting individuals who are influential in different communities. She will need simple and intuitive definitions, compelling arguments and powerful examples, as well as time, influence and luck’ (Pinder Reference Pinder2021: 157).

Arguments have a role in this vision, but so does the influence of high-profile language users. For another example, Christian Nimtz suggests that there is no need to convince rationally all the target speakers: it might be enough to convince just a few, ‘given that these are the right ones’ (Reference Nimtz2021: 3–4). The right ones are ‘those whom ordinary speakers look to in the matter’ (Nimtz Reference Nimtz2021: 23). This strategy makes the intervention a non-consultative one for the concept users who are not rationally convinced.

Should conceptual engineers just refrain from engineering non-consultatively, then, and instead rationally convince all the target agents to adopt the proposed concepts? I do not rule this out completely, but complications should be noted. First, the engineers sometimes seem to have in mind large populations as their target; and indeed, it often makes sense to target large populations. For example, it is not enough to make a few friends stop thinking about rape in the mistaken ways facilitated by rape metaphors; this is a culture-level problem and needs to be addressed as such. However, it would be a tough feat to even reach such large populations with one's arguments, much less convince them all.

Further, even when we only target a smaller group of people that we can hope to reach and convince—for example, our fellow philosophers—there are complications. For shaping inference and attention patterns, it might not be enough that those rationally convinced make a conscious effort to use terms in a certain way. As Machery (Reference Machery2021) and Eugen Fischer (Reference Fischer2020) both point out, our ability consciously to control the concepts that we use is limited. Even if one decides to use a term in a revised way in a certain context, the folk concept can remain effective, guiding one's inferences. Conceptual engineers might not be able to change how even a limited group of people think in a specific context without changing the relevant folk concepts. And changing folk concepts requires convincing a large population.

To this point, I have argued that certain projects of conceptual engineering (those that non-consultatively try to improve patterns of inference and attention) involve paternalistic cognitive engineering. This gives a pro tanto reason against these projects: the target agents could reasonably regard the projects as violating their intellectual sovereignty. Before considering whether this concern can be outweighed by reasons in favor of engineering, it is important to distinguish the worry raised here from other ethical concerns about conceptual engineering.

First, one might have different antipaternalist concerns. I have been discussing a broadly Kantian antipaternalist worry, one that appeals to autonomy in the sense of sovereignty. But there are also broadly Millian antipaternalist concerns, which revolve around the paternalists’ fallibility. For example, when we do not consult the target agents, we may be mistaken about what is good for them, and so end up harming rather than helping them. Robert Simpson discusses this worry in the context of epistemic paternalism in education: ‘The worry that's lurking, for the epistemic paternalist, is that if you aren't being consultative, it's hard to make a person's inquiry more successful, because the aims of her inquiry, and thus what would qualify as a success in it, evolve as the inquiry develops, in a way that the paternalizer isn't well placed to keep track of. Whether you are epistemically benefiting someone depends on whether you're helping them gain interest-relevant beliefs, knowledge, or understanding’ (Simpson Reference Simpson, Matheson and Lougheed2021: 98).

This consideration is also pressing for conceptual engineers trying to improve other people's patterns of attention. How can the conceptual engineers know, without consulting the target agents, that they are really directing attention to the most interest-relevant matters? And a similar worry, also related to the paternalists’ epistemic situation, arises for inference-shaping conceptual engineering that is supposed to facilitate true thoughts and suppress false thoughts: the conceptual engineer might be mistaken about the truth and falsehood of the thoughts in question.

Such worries about the paternalists’ epistemic situation revolve around Mill's primary concern regarding paternalism: the paternalists’ fallibility. For example, Mill writes, ‘if any opinion is compelled to silence, that opinion may, for aught we can certainly know, be true. To deny this is to assume our own infallibility’ (Mill [1859] Reference Mill and Alexander1999: 97). In this spirit, one might insist that we should not use conceptual engineering to reduce the occurrence of false thoughts and to encourage true thoughts: for all we can certainly know, these thoughts may not be false and true, respectively. Further, conceptual engineers can be wrong about other relevant matters, like which inferences the proposed language policies would end up promoting or discouraging; and more generally, they might be unable to predict the consequences of their benevolent interventions. Teresa Marques's (Reference Marques, Marques and Wikforss2020) discussion of the ethics of conceptual engineering focuses on fallibility worries of the latter sort, worries that well-intentioned engineers can give rise to ‘meaning perversions’ (275–77), such as the shifts in the meanings of terms like heroic and enemy of the people under oppressive propagandistic regimes.

While such worries as Mill's about conceptual engineers’ fallibility are also important and worth further discussion, they are distinct from my sovereignty-based concern. One may combine both worries, however, borrowing Feinberg's words: ‘By and large, a person will be better able to achieve his own good by making his own decisions, but even where the opposite is true, others may not intervene, for autonomy is even more important a thing than personal well-being’ (Feinberg Reference Feinberg1983: 460).

Finally, there are also ethical concerns about conceptual engineering that have nothing to do with paternalism. Naturally, one might be concerned about engineering that aims at the engineers’ good, not at the target agents’ good. For example, Matthieu Queloz and Friedemann Bieber (Reference Queloz and Bieber2021) argue that making conceptual engineering easy to implement by somehow institutionalizing it—perhaps through a ministry of conceptual engineering—would be illiberal and undemocratic, opening the door to abuses and allowing those in power to engineer consent, instead of securing genuine consent. I have left such concerns about malevolent or selfish engineering aside here, focusing on the less obvious ethical concerns that arise for benevolent conceptual engineering. Granted, as Matthew Shields (Reference Shields2021) points out, the current discussion on conceptual engineering might well give one the impression that shaping concepts is typically benevolent—more precisely, that it is motivated by an interest in establishing the correct or best concepts to use, rather than the shapers’ selfish interests; and that impression is mistaken. Nevertheless, I focus on benevolent conceptual engineering in this essay.

3. Justifying Paternalistic Conceptual Engineering

To this point, I have identified a pro tanto problematic variety of epistemic paternalism—paternalistic cognitive engineering—and placed some projects of conceptual engineering within this category. The pro tanto reason against such engineering is that the target agents could reasonably regard it as a violation of their intellectual sovereignty. But are there reasons in favor of these projects that could outweigh the sovereignty-based reason against them? Below I consider two avenues for justifying the projects: (1) by appeal to the target agents’ epistemic good alone, and (2) by appeal to the epistemic improvement's broader social impact.

One might appeal to the target agents’ own epistemic good with the sort of move that the defenders of nudging (see, for example, Thaler and Sunstein Reference Thaler and Sunstein2008) often make: people's choices are in any case influenced by how information is presented or how the environment is arranged. For example, people will in any case tend to choose the food at eye level. So, if we refrain from interfering paternalistically, their choice will instead be influenced by chance or possibly by actors who do not have their interests in mind. The same point can be made about conceptual engineering. Concepts shape patterns of inference and attention, with or without our intervention; and unless we interfere for the concept users’ own epistemic good, they will be influenced by how their concepts are shaped by chance or possibly by malevolent actors. One might say, then, that if we leave people to govern their own concepts, we in effect leave them at the mercy of other (intentional and unintentional) forces that are not benevolent like us. Even if there is a sovereignty-based reason against engineering, there is also a reason to engineer, arising from concern for the target agents’ epistemic well-being.

However, I doubt that this consideration alone can outweigh the sovereignty-based reason against non-consultative conceptual engineering. The engineers do not face a choice between engineering non-consultatively and doing nothing. They also have the option of promoting awareness of how concepts influence belief formation—how concepts in general and specific concepts make certain things salient and put our thoughts on certain tracks—and offering their own suggestions for conceptual revision, along with the reasons. The engineers could thus empower people to exercise control over their mental lives, instead of just taking the matter into their own hands. It seems, then, that a compelling case for non-consultative engineering would need weightier reasons.

Another possible defense of non-consultative conceptual engineering (again, by appeal to the target agents’ own epistemic interests, broadly construed) is the following: While some people may wish not to have their concepts non-consultatively managed by the engineers, others might not mind and might even appreciate it; and that preference is also reasonable. One might well prefer not having to consider the justifications that consultative engineers offer for their proposals. This takes time and effort that one could spend on other matters, while benevolent engineers non-consultatively look out for one's interests in shaping the community's shared conceptual repertoire. For people who would rather just go with the flow when it comes to concepts, benevolent non-consultative conceptual engineers conveniently make sure that the flow takes them in a good direction.

It is unclear, again, whether this consideration outweighs the sovereignty-based concern. Even if many or most people take such a lax attitude toward having their intellectual lives controlled for their own epistemic good, the engineers cannot legitimately assume that all their target agents have this attitude. For an analogy: perhaps many or even most people (reasonably) appreciate non-consultative hugging and feel annoyed about being consulted, while some (also reasonably) regard non-consultative hugging as a violation of their bodily sovereignty. Can we then just go ahead and non-consultatively hug someone whose preference we do not know? It seems that there is a greater presumption against violating bodily sovereignty than against disrespecting an unspoken wish to be non-consultatively hugged. (The latter is plausibly not a sovereignty violation, if sovereignty means primarily a right not to be treated in certain ways without our consent, in the relevant domain, rather than a right to be treated in our preferred ways, in that domain.) Perhaps we should err on the side of caution by refraining from the hug or by consulting first; but I leave this (along with the potentially relevant differences between non-consultative hugging and non-consultative conceptual engineering) for future discussion.

Alternatively, one might appeal to the broader social impact of the epistemic improvement. For example, Ahlstrom-Vij justifies withholding evidence from jurors like this: ‘[A]ny legitimate moral claim on the part of the jury not to be interfered with is trumped by considerations about the welfare of others, such as the defendant and those allegedly wronged by the defendant’ (Reference Ahlstrom-Vij, Grill and Hanna2018: 271). Indeed, the ultimate motivation behind epistemic paternalism often has to do with the interests of people other than the target agents, or with the non-epistemic flourishing of society more generally (including the target agents). Likewise, the motivation behind conceptual engineering is often that our concepts have far-reaching societal consequences beyond the immediate cognitive effects for the concept users. For example, Alexis Burgess and David Plunkett motivate evaluating and improving concepts with the idea that our concepts’ ‘non-conceptual consequences are pervasive and profound’ (Reference Burgess and Plunkett2013: 1096–97).

Pursuing this line of defense, one might further point out that we can almost always expect our epistemic success and failure to have broader social impact in some roundabout way. For example, W. K. Clifford ([1877] Reference Clifford1999) argued that false beliefs tend to spread by various mechanisms, and this can have far-reaching consequences for society. An apparently insignificant false belief can lead one to revise their other beliefs and to interpret incoming evidence based on the false belief, resulting in deterioration of the quality of a bigger portion of their beliefs. And this in turn can result in harmful behavior based on the false beliefs or the spreading of those beliefs to others via testimony. When someone's belief-formation goes astray, we can typically expect adverse effects for others, regardless of the content of the belief. The lesson that Clifford takes from this is that we bear a general responsibility to other people to govern our beliefs well. But one might also draw the further conclusion that we can non-consultatively interfere to secure the quality of our fellow humans’ belief formation, even in the absence of specific reason to suspect that the beliefs in question are consequential for us.

However, it is dubious that these Cliffordian considerations give us a free pass to do just anything to improve other people's epistemic state. Consider Emma Bullock's (Reference Bullock2018: 442) example of secretly playing lectures on quantum mechanics to someone while they sleep. Even if this would be an effective way to teach quantum mechanics, we cannot justify this intervention merely on the Cliffordian grounds that the person's ignorance of quantum mechanics could well end up harming others in some roundabout way that we cannot yet predict; or that their knowledge is likely to benefit society in some roundabout way and the sovereignty violation is thus justified.

On one hand, it is plausible that the broader social impact of the epistemic improvement can speak in favor of paternalistic cognitive engineering (including conceptual engineering) and outweigh the sovereignty-based reason against these practices. On the other hand, as the quantum mechanics example shows, the justification cannot be merely that epistemic improvement typically has such broader social impact. There is no escape from looking at the specifics of a given paternalistic project. To make this point clearer, I offer two possible projects of conceptual engineering: one that is plausibly justified and another less plausible.

A pro tanto reason against this project is that the target agents could reasonably regard it as a violation of their intellectual sovereignty. However, there are also powerful reasons to use non-consultative cognitive engineering here and not just directly address the mistaken ways of thinking about rape, or engineer concepts entirely consultatively. It is very difficult to change ingrained misogynous attitudes by arguments; and for related reasons, it is very difficult to dissuade through reason many of those prone to using rape metaphors. Further, these mistaken ways of thinking cause further suffering for victims and could contribute to failures to deter rape—so, the epistemic improvement here demonstrably and not just potentially has important social impact. Moreover, it is beyond reasonable doubt that the discouraged ways of thinking about rape are indeed mistaken and harmful; and it is also very plausible that the rape metaphors indeed perpetuate these mistaken and harmful ways of thinking. And there are even further reasons against using rape metaphors: these metaphors are often triggering and insulting for victims. It is thus plausible that the activism against rape metaphors is ultimately justified, despite the sovereignty-based reason against it.

The engineers, again, might insist that the worthy goal of gender equality outweighs the sovereignty-based reason against the project. However, various features of the project make that response less convincing here. Granted, gender equality is a worthy aim, and it is somewhat plausible that the conceptual revision would indeed contribute to that aim, by guiding attention to gender inequality. However, these engineers’ methods are focused less on persuading the target agents through reason and more on trying to wield institutional power and social sanctions. One might also question whether the heightened attention to gender inequality across all contexts where gender terms come up best serves everyone's epistemic interests. There might be other valuable ways of thinking about the relevant aspects of the world: one might want to emphasize gender identity rather than gender oppression, for example (see Jenkins Reference Jenkins2016). In sum, these Haslanger-inspired engineers should seriously consider other ways of promoting gender equality or at least pursue the conceptual revision more consultatively. (I do not wish to suggest, of course, that Haslanger herself envisioned such a project as described here.)

There is no single thing that decisively sets this project apart from the previous one; various features matter. For example, it is relevant that the activists who try to eliminate rape metaphors are protecting people from clearly harmful false beliefs, whereas those trying to revise gender terms are guiding attention to some important concerns among others; accordingly, there is more room for reasonable disagreement in the latter case. The additional reasons for discouraging rape metaphors are also relevant and contribute to the justification of that project. Perhaps some would nevertheless consider the project of revising gender terms to be justified as long as it significantly contributes to gender equality. I do not insist that my judgment on the cases is necessarily correct. My aim is simply to show that various details matter for assessing whether an engineering project is ultimately justified, despite the sovereignty-based reason against it. It is not enough, then, to make the Clifford-like point that an individual's epistemic well-being typically has broader social impact; and it is also not enough just to gesture toward some worthy aim (such as gender equality) that supposedly outweighs infringing on intellectual sovereignty. The project's specific aims, means, and social context matter.

4. Concluding Remarks

The enterprise of non-consultatively engineering concepts to improve inferential and attentional dispositions falls into a pro tanto problematic category of epistemic paternalism: paternalistic cognitive engineering. This category arguably also includes certain kinds of nudging and evidence suppression. There is a reason not to engage in paternalistic cognitive engineering: the target agents could reasonably take it to violate their intellectual sovereignty. They could prefer to be persuaded, informed, and educated (which may also include consultative conceptual engineering) rather than be non-consultatively engineered, for this would give them more control over their belief-formation. I assumed without much argument that it is reasonable to be rather protective of one's intellectual domain and to oppose exercises of control over that domain when methods like explicit persuasion are in principle available. However, I acknowledge that a more lax attitude toward such benevolent exercises of control is also reasonable.

The sovereignty-based concern I raise is only a pro tanto reason against such conceptual engineering projects. One could insist that concern for the target agents’ own epistemic well-being outweighs the sovereignty-based worry. If we refrain from engineering, people's inferential and attentional patterns will be shaped by chance and possibly by malevolent actors. Further, some target agents might even appreciate the non-consultative benevolent engineering, far from regarding it as sovereignty violating. (I leave this avenue for justification open but have some reservations about it.) Second, one could propose that the broader social impact of the epistemic improvement outweighs the sovereignty-based concern. This needs to be assessed case by case, as there are various relevant differences between projects of conceptual engineering.

Conceptual engineering can be a powerful tool of epistemic improvement; and epistemic improvement can in turn serve social progress. But in pursuing epistemic and social progress by conceptual engineering, we should bear in mind that our targets are persons with their own legitimate domain of belief formation, and they might not welcome non-consultative benevolence; if we nevertheless want to bestow it on them, a good justification is needed.