1. Manipulation in Philosophy and Politics

People influence each other every day. Some forms of influence, such as reasoned persuasion, are morally permissible. Others, such as bargaining, are morally acceptable, unless unequal bargaining is exploitative. Other forms of influence, such as coercion, are seen as morally unacceptable in most circumstances. Generally, manipulation is a type of influence where the target is moved to action without overt threats or coercion, but that falls short of persuasion. There are many accounts in philosophy and political science that attempt to identify the phenomenon and provide necessary and sufficient conditions for manipulation. Some, though not all, rely upon an overly idealized conception of human rationality and deliberation. In this section, we note some of the shortcomings of existing models, arguing that we need a realistic, applicable and more precisely specified model that can be applied to a range of cases in politics and public policy, including nudges.

There is no universally agreed-upon definition of manipulation. Noggle (Reference Noggle2022), for example, identifies three distinct conceptions of manipulation, alongside a series of hybrid or disjunctive views aiming to bridge the gaps between them. Some scholars even argue that manipulation is too multifaceted or vague to be rigorously defined (Ackerman Reference Ackerman1995: 337–338; Wilkinson Reference Wilkinson2013).Footnote ² The most prominent accounts tend to fall into one of three categories: (1) manipulation as bypassing rational deliberation, (2) manipulation as a form of trickery and (3) manipulation as a form of non-coercive pressure (Noggle Reference Noggle2022).

Manipulation as a form of influence bypassing or subverting rational deliberation (Gorin’s (Reference Gorin2014a) ‘bypassing or subverting view’, henceforth BSV) is especially prominent in debates about whether nudges are manipulative. Critics claim some nudges operate ‘behind the back of choosers’ (Waldron Reference Waldron2014), ‘circumvent deliberative faculties’ (Grüne-Yanoff Reference Grüne-Yanoff2012), ‘undermine that individual’s control over her own deliberation’ (Hausman and Welch Reference Hausman and Welch2010), or otherwise ‘pervert’ people’s reasoning capacities (Wilkinson Reference Wilkinson2013). Even nudge defenders concede that a nudge can be manipulative ‘to the extent that it does not sufficiently engage or appeal to their capacity for reflection and deliberation’ (Sunstein Reference Sunstein2016: 216, emphasis removed), and that ‘[m]anipulation occurs when one influences another by bypassing their capacity for reason, either by exploiting nonrational elements of psychological makeup or by influencing choices in a way that is not obvious to the subject’ (Blumenthal-Barby and Burroughs Reference Blumenthal-Barby and Burroughs2012: 5).

Nevertheless, BSV accounts offer no consensus on the threshold for such bypassing or subversion. As Noggle (Reference Noggle2022) argues, too high a threshold – where one would have to bypass one’s rational capacities fully through (unrealistically) effective subliminal advertising or brainwashing – would rule out numerous behaviours commonly considered manipulative. On the other hand, too low a threshold – where any non-rational influence can count as manipulative – would lead to overinclusive accounts, condemning uses of non-verbal or emotional cues not commonly considered manipulative. Another problem with BSV accounts is that neither bypassing nor subversion is necessary for manipulation to occur across a range of plausible BSV accounts; Gorin (Reference Gorin2014a) offers a series of counterexamples to show that one can manipulate a target while simultaneously engaging their rational faculties.

Given that defining the right threshold will always be controversial, it would be advantageous for an account of manipulation to avoid requiring an account of how rational capacities should be engaged. Much of what we believe and do is based on testimony, hearsay or habit; but that does not mean we have been manipulated (Dowding Reference Dowding2016). We can be influenced by poor reasons or irrational drives without being manipulated. We might act in a deliberately reasoned manner yet be manipulated – if we have been lied to, for example. An account that does not rely on an account of rational autonomy is an advantage so long as it captures the examples that BSV accounts desire.

The second most common account of manipulation views it as a type of trickery (Mills Reference Mills1995; Noggle Reference Noggle1996; Barnhill Reference Barnhill, Coons and Weber2014; Hanna Reference Hanna2015). Noggle describes manipulation as a more general form of deception that ‘tricks the target into adopting a faulty mental state’ and can influence any range of mental states, leading to false beliefs, but also to faulty desires or inappropriate emotions (Noggle Reference Noggle2020: 243).

The third account is manipulation as a form of positive or negative pressure that does not rise to the level of coercion (Baron Reference Baron2003; Wood Reference Wood, Coons and Weber2014; Noggle Reference Noggle2020). Examples include emotional blackmail, nagging, flattery and the ‘charm offensive’. Noggle (Reference Noggle2020) unifies the second and third account by arguing that both induce the target to make a mistake or engage in flawed reasoning. The deceptive manipulator takes advantage of cognitive biases to influence the target’s decision. The pressure manipulator takes advantage of akratic tendencies, such as the propensity to overvalue short-term obvious costs and benefits over long-term less obvious ones. In all such cases, the manipulator intends the target to fall below certain norms that ought to govern the formation of one’s beliefs, desires or emotions (see also Noggle Reference Noggle1996, Reference Noggle2022).

While there is much to commend in such accounts, some aspects are currently underspecified in ways that leave open how they can be applied to cases. First, one must determine which norms ought to govern non-manipulative influence. Second, one needs to determine whose perspective on the relevant norms ought to take priority. Noggle’s account is open-ended about whether the manipulator intends to get the target to fall below the manipulator’s norms, the target’s norms, or the prevailing social or moral norms. While these tasks are often feasible, a more parsimonious set of conditions is preferable. Our account provides two such conditions.

A fourth account of manipulation is especially prominent in political science. While a subset of the manipulation-as-deception camp, political theorists present it independently with specifically political applications. It derives originally from Goodin (Reference Goodin1980), who describes manipulation as involving ‘unknown interference’ (that is, a covert and deceptive exercise of power over the target) and ‘unwelcome interference’ (that is, power exercised in ways contrary to the known or assumed will of the target). Whitfield’s (Reference Whitfield2022) recent account shares these key features, while aiming to expand the scope of manipulation beyond direct manipulation in interpersonal cases. He offers the following definition:

Our account is largely compatible with this definition, while bringing out important features of such attempts. It does so without relying on what Klenk (Reference Klenk2022) calls the ‘covertness thesis’ which equates manipulation to ‘hidden influence’ (Klenk Reference Klenk2022: 86). To be sure, manipulatory acts themselves need not be hidden, only (some of) the intentions of the agent.Footnote ⁴ There are many examples where the manipulatory act is not itself covert (Barnhill Reference Barnhill, Coons and Weber2014; Gorin Reference Gorin, Coons and Weber2014b; Klenk Reference Klenk2022), but the agent wishes to conceal some key motivating aspects of their intentions. The conditions we develop rule out these forms of concealment without requiring a further discussion of manipulative versus non-manipulative methods à la Whitfield. We should also note that while unseen influence can be manipulative, this does not mean that if the target uncovers the attempted manipulation, the attempt was therefore non-manipulative. As we argue in section 3, disclosure and hypothetical disclosure by the agent can affect the degree of manipulation involved (regardless of the method deployed) in ways that discovery or hypothetical discovery by the target cannot. Whitfield’s definition thus needs weakening to admit A’s methods aiming to render their intentions obscure or deniable.

For Dowding (Reference Dowding2016, Reference Dowding2018), what matters for distinguishing persuasion from manipulation is how the agent attempts to influence the target. Persuasion occurs when reasons for some proposition P endorsed by the agent are offered to the target, and the agent displays deliberative honesty. That is, the agent is herself motivated by or identifies with those reasons and does not deliberately hide contrary considerations. In discussion, the agent is willing to change her views.Footnote ⁵

So far, we have suggested that we need an account of manipulation that (1) is realistic (that is, does not rely on an idealized account of rational deliberation), (2) is sufficiently precise to apply across a range of cases in politics and public policy, and (3) pays attention to the intentions of the putative manipulator rather than just the target of the manipulation. Reviews of the literature suggest this last point has been relatively neglected (Engelen and Nys Reference Engelen and Nys2020; Noggle Reference Noggle2022), although some accounts aim to redress this (Mills Reference Mills1995; Baron Reference Baron, Coons and Weber2014; Gorin Reference Gorin, Coons and Weber2014b; Klenk Reference Klenk2022). The account which centrally features the agent’s intentions is Dowding’s (Reference Dowding2016, Reference Dowding2018) – although he defines persuasion rather than manipulation. He provides two ‘reliability’ conditions that are jointly sufficient for non-coercive persuasion: (1) common reasons (roughly, that the agent persuades the listener using reasons that she herself holds for her beliefs) and (2) the agent’s intention is to learn the truth together (or otherwise come to an agreement regarding the truth). Unfortunately, Dowding’s account suffers some limitations.

First, it is narrowly specified to cases of interpersonal communication in the context of democratic deliberation. The intention condition, in particular, is too narrow to be applicable in cases where a speaker aims to convince an audience or where a government agent aims to influence the behaviour of the citizenry through public policy. In sections 2 and 3, we will show how suitably modified versions of Dowding’s original conditions allow us to distinguish between manipulative and non-manipulative influence in a range of contexts within politics and public policy – deliberative as well as non-deliberative.

Second, the original conditions are too demanding. For example, reasons that an agent finds plausible or persuasive, but that are not shared with the target, will fail to pass the non-manipulation threshold. According to Whitfield (Reference Whitfield2021), Dowding’s account rules out non-manipulative cases of ‘conjectural reasoning’ (that is, cases of persuasion through reasons that the target accepts, but the agent does not). This is partly because Dowding’s conditions are meant to be sufficient rather than necessary for non-coercive persuasion. In this paper, we modify and expand the common-reason condition to account for a broader range of non-manipulative cases, including cases of conjectural reasoning allowing us to provide necessary as well as sufficient conditions, further improving the original account.

Finally, our account moves beyond the original Dowding conditions to introduce a way in which we can make judgements about the degree of manipulation, helping assess its normative status across different situations labelled manipulative. This allows us to develop a more general account of manipulation that can speak to the concerns of multiple literatures, including the ethics and politics of nudge, the philosophical literature on interpersonal manipulation, and the role of manipulation within politics.

2. Conditions for Non-manipulation: Communicative Influence

In the following sections, we introduce two conditions to distinguish persuasion from manipulative influence. We begin by looking at communicative influence, where an agent i attempts to influence a target agent j by providing evident reasons for j to believe P or choose option x over alternatives. We then proceed (in section 3) to investigate situational influence, where an agent i arranges the situation that j finds herself in, so as to get j to believe P or choose x. These forms of situational influence are often undisclosed and require separate analysis.

Communicative influence involves evident communication between the agent i and the target j. This communication could be verbal and bidirectional, or indirect and involve a recorded message sent to j by i. In addition to words and arguments, such communications sometimes involve images, sounds and emotional signals, where those signals are evident (such as traffic lights).

We should clarify that, when we say that communicative influence involves i providing reasons to j, we rely on a minimal conception of reasons. A reason is any explicit or implicit statement that can be taken by j to cause, explain or justify belief in P or choice x. These do not have to be morally good, philosophically rigorous, or prudentially sound reasons. They simply have to influence j to assent to P or to choose x, even if the epistemic practice leading j is flawed or inadequate. Biased evidence, the status of i, or even feelings described by i can count as reasons.

Reasons are often thought of as conscious decisions or beliefs. Certainly, those accounts suggesting manipulation involves bypassing rational deliberation seem to suggest so. However, as we have already pointed out, much of our behaviour is habitual (though still propositionally justifiable). On this externalist account we view any affect as a potential reason, even when people are not consciously motivated. We can (sometimes) infer that j is choosing x for reason E,Footnote ⁶ but that doesn’t mean that j is necessarily aware that E is the reason for their choice. In a famous study, Californian residents were exposed to one of five possible door hangers inviting them to conserve energy (Nolan et al. Reference Nolan, Schultz, Cialdini, Goldstein and Griskevicius2008). The control provided (truthful) information about how one could conserve energy. The four treatment door hangers included messages about how conserving energy: (1) protects the environment; (2) benefits future generations; (3) saves money; and (4) is already done by most of one’s neighbours. The largest effect on behaviour, measured by household electricity use before and after the intervention, came from the social norm treatment (4). Yet individuals claimed the habits of their neighbours were the least important reason, showing a mismatch between stated reasons and behaviourally revealed reasons. Our externalist account will identify both stated and revealed reasons as reasons, even if the agent may only consciously deliberate about the first category.

Drawing on Dowding (Reference Dowding2016), we give two conditions that separate non-manipulative from manipulative influence. These conditions build on the paradigmatic understanding of interpersonal persuasion, where an individual i offers suitable reasons to a target j in a truthful manner in order to influence j’s beliefs, actions or choices in ways that align with i’s own beliefs, actions and choices. By contrast, a paradigmatic form of manipulative communication involves individual i offering any reason to a target j that will influence their beliefs, actions or choices, irrespective of whether the reasons given are shared, truthfully presented or plausible to i. In this case, i’s goal is only to successfully influence j’s behaviour by whatever means necessary. The underlying intuition is that while i causes j to believe P, where i and j share the same common cause, E, leading them to believe P, then i has not manipulated j. Where i has used any means whatsoever to lead j to believe P, then this is not persuasion, but some form of manipulation. Common cause by E is weakened (see below) to suitable reasons. We thus redefine non-manipulative persuasion, recognizing that manipulation comes in various forms.

We specify the two reliability conditions for non-manipulative influence: Suitable Reasons (SR) and Testimonial Honesty (TH) conditions. These conditions are weaker than Dowding’s original conditions, which were constructed in the context of deliberation and persuasion in order to reach common agreement. Political and public policy communications are not always so deliberative in form.

When we say that certain reasons are ‘suitable’, we mean that these reasons support the claim that the agent wants to make in the opinion of the agent i. These do not have to be objectively true or objectively fitting reasons, but merely reasons that the agent finds suitable as grounds for a given proposition P or as reasons to choose x. They do not need to support P on their own (that is, they need not be decisive reasons) and may require significant supplementary or background assumptions to do so. They need not even constitute the most important reasons why the agent could come to believe P, though we will subsequently include a separate condition about truthfulness and relevance in the opinion of the agent.

Ideally, both i and j believe that the evidence that constitutes the reasons E establish a suitable reason to assent to P (even if i or j does not assent to P for other contrary reasons). These grounds E might constitute content-dependent reasons for assenting to P. Proposition P does not have to be true, nor even to logically follow from or be demonstrated by E. It must only be the case that both i and j recognize that E roughly leads to a belief in P. For example, i might convince j that vaccination reduces the risk of becoming infected with COVID-19 on the grounds that i doesn’t have any vaccinated friends who have tested positive for COVID-19. Such anecdotal evidence based on i’s availability heuristic would not be sufficient to demonstrate the truth of P, but if both i and j believe E and take it to be a reason for believing P, then i has persuaded rather than manipulated j. In this example, i finds reasons E to be suitable reasons to vaccinate, but she might have other reasons – such as some specific medical condition – not to get vaccinated herself. If, on the contrary, i was arguing that j should get vaccinated based on E, but i herself found E not to be a suitable reason, failing to pass the level of scrutiny that i generally uses when making decisions that affect her (or her family), or because it is based on known false data, then i is attempting to manipulate.

The condition also does not rule out real-world circumstances where both i and j come to believe P based on content-independent reasons E (that is accepting the judgement of another without fully knowing their reasons for their judgement), because of testimony by a third-party k whom they deem trustworthy on the subject. Neither does it rule out j believing P for reasons E because j sees i as a trustworthy source on the subject, provided the second condition below is simultaneously met. Following the example above, i might convince j to believe P on the basis of E (for example, testimony on vaccine efficacy from the World Health Organization or national public health officials in their country). Or i might be regarded by j as having more authority on the subject and therefore j adopts the reasons E given by i. (This claim to authority might be well grounded – i is an infectious diseases specialist – or poorly grounded – i took a biology course in college. Provided i is not deceiving j about their credentials or qualifications in ways ruled out by our testimonial honesty condition, the SR condition can be met.)

We can see in other examples that belief in P will often have opposite consequences for individual choice. Telling two people that the beverage before them is unsweetened tea can lead one to drink and another to reject it. Such true communication is non-manipulative, even if it predictably leads different people to opposite (often highly predictable) choices. In other words, SR does not require that i and j come to the same conclusions given P, or that they make the same choices based on P. This applies to many communicative nudges through which i informs j about the behaviour of others through proposition P (for example, about how much energy j’s neighbours consume, how often j’s fellow citizens vote in elections, or how much people j’s age drink in an average week). As long as i communicates suitable reasons for P, the nudge is not manipulative. Many people respond to information about the behaviour of others by choosing to conform to the descriptive norm. But others will not. Not only that, but a given individual could oscillate from conformist to contrarian, depending on the context and their mood on a particular occasion.

The revised SR condition responds to Whitfield’s (Reference Whitfield2021) concern about ‘conjectural reasons’. Reasons can be suitable even if they are not decisive in the particular case. For example, i might convince j that choosing a beer from a local microbrewery is morally preferable to choosing one from a larger manufacturer even if i is a recovering alcoholic who does not drink or a teetotaller who chooses not to. This means that i can appeal to suitable reasons even in cases where i does not meet the supplementary conditions for reason E to be decisive in her own case. This gives us a different condition for when conjectural reasons are manipulative compared with Whitfield (Reference Whitfield2021). The decisive factor is not whether i is covert or open about being opposed to drinking alcohol. The decisive factor is whether i considers a given reason to be suitable. (Lying and claiming to drink when one does not is manipulative, according to the second condition below, as would failure to disclose when the issue is particularly relevant; but one is not under a general obligation to disclose one’s AA membership when discussing alcohol.)

TH focuses on the intentions of i in persuading j of P on grounds E. So long as i is aiming to provide relevant and truthful information (even if i holds false or irrelevant beliefs that i wrongly considers to be true), the TH condition can be met. We would not expect i to disclose any and all information to j. Not only would this be unrealistic, but it might even be counterproductive, since information overload can lead j to make worse decisions than following the presentation of less information that is pertinent to the question at hand. Instead, we would expect i to use common sense as to which information is relevant. One helpful heuristic would be to ask oneself, ‘What kind of information would I want in order to make an informed decision in this matter?’

Statistical information is particularly problematic in this regard. The deliberate misuse of statistical visual representation is well known (see, for example, Tufte Reference Tufte1997: 55−72; Calzon Reference Calzon2021). However, using correct but irrelevant statistics is also problematic. Stating the increased risks of heart disease or dying from an activity, without providing the baseline risk, is ubiquitous in newspaper reports of healthy eating and drinking, for example. Nonetheless, statisticians often disagree on what the relevant statistics are. What matters in our account is the integrity and the beliefs of those providing the information.

The two conditions are jointly necessary and sufficient for influence to be non-manipulative. Their role is to identify unambiguous ways in which tools of persuasion can be used without manipulating the target.

We can see how TH can hold, yet manipulation still occurs by breaking SR, in the imaginary case of Lucrative Suicide (Gorin Reference Gorin2014a). In this case, agent i is aiming to persuade target j to commit suicide by using information, that i considers truthful and relevant, about the non-existence of God. In this case, the reasons presented by i for the proposition P (there is no God) are reasons that i finds suitable for determining the truth of P. However, i himself does not think that these reasons are suitable as reasons to commit suicide. So i is persuading j to choose x (that is, to commit suicide) on the basis of reason E (the case for there being no God), which violates the SR condition.Footnote ⁹

We can also construct examples where the SR condition holds, but the TH condition is violated. These include cases where i considers E to be a suitable reason for believing P or choosing x if E was true, but at the same time believes E to be false. For example, i might be selling ‘junk food’, claiming it is nutritionally valuable when it is not. Although i might believe that E is a suitable reason for choosing their product if true, i can still violate TH if they do not believe that E is true or do not believe that E is relevant for the majority of customers (for example, for those who are not endurance athletes and therefore do not need to ‘carbo-load’).

Real-world examples may only approximate our two conditions in the sense that most real interactions might involve some form of manipulation. We do not consider this as problematic for our account. While we have defined what non-manipulative communication is, beyond the ‘on/off switch’ manipulation comes in degrees. Some forms of communication are more manipulative than others. After all, many of our actions, including persuasion, emanate from multiple reasons. Ceteris paribus, the more manipulative a given instance, the more normatively problematic it should appear to be. We have provided ideal conditions for non-manipulation, and as far as the real world approximates those conditions, we do not consider the interaction to be manipulative. As the interaction departs further and further from these conditions, we will consider it to be manipulative; however, some forms of manipulation might seem rather mild, warranting little concern in either personal or public policy contexts. The further away from our conditions the interaction is, the more manipulative it is and the more concerned we should be about it. Finally, though, we note that while manipulation at any level can be considered pro tanto wrong, in some contexts even severe manipulation might be normatively justified. In the final section, we introduce several considerations that might mitigate the normative concerns about manipulation in particular cases. Here it suffices to say that the more distant a communicative influence from our non-manipulation conditions, the more it is manipulative.

Our account provides a more precise understanding of manipulation in politics and public policy. In a recent US example, Fox News commentators, such as Tucker Carlson, compared vaccine mandates to apartheid, and undermined their efficacy on an almost daily basis. Yet the company itself mandated that all unvaccinated employees (comprising less than 10% of the total workforce) must undergo daily covid tests – a regime far stricter than that suggested by the Biden administration and subject to daily attacks from Fox broadcasters (Bauder Reference Bauder2021; Ecarma Reference Ecarma2021). This seems clearly to be manipulation. First, Carlson seems to be violating TH by failing to disclose relevant information such as the stricter vaccination mandates of the Fox organization and the high vaccination rate among Fox employees. Second, he and others seem to be violating SR by presenting reasons against COVID vaccination that their revealed choices suggest they do not believe. Of course, further information is required for a final determination in this (or any real-world case), including further information about the putative manipulators’ intentions and beliefs.

Another example concerns elected representatives manipulating the public with regard to climate change. The Center for American Progress (CAP) identifies ‘climate deniers’ as those who deny there is any climate change, or deny it is caused by human activity, or deny scientific consensus on the issue (Drennan and Hardin Reference Drennan and Hardin2021). The CAP classification is based on interviews or official or informal communications about climate change from each official. To be examples of manipulation, these communications would have to violate either the SR or the TH condition. An SR violation could be a politician who denies global warming because E (for example, there was some unusually cold weather last winter), but who knows that E is a bad reason for denying climate change. A TH violation could be a politician claiming there is no scientific consensus on global warming, which he knows to be false or misleading.

It seems that at least some politicians are violating either SR or TH, or both. Former Senator Richard G. Lugar claims that some of his Republican colleagues are not expressing their true beliefs on climate change: ‘So even if they privately believe we ought to do something about it, they’re reticent, especially with the Republican president taking the views he is now taking’ (Davenport and Lipton Reference Davenport and Lipton2017). Energy reporter Anthony Adragna (Reference Adragna2014) conducted interviews with dozens of former senior congressional aides, nongovernmental organizations, lobbyists and others:

Adragna lists economic crisis, the limited popularity of policies addressing climate change among Republican voters, the Tea Party’s influence over Republican primaries and the hostility of environmental groups towards Republican candidates as reasons for violating TH.

Some of these politicians, of course, might be truthfully conveying reasons that persuaded them to deny climate change and might themselves have been manipulated by others. Climate misinformation is intentionally spread through organizations funded by oil companies. Keith McCoy, a former top Exxon lobbyist, acknowledges Exxon’s funding of climate denial in a secret recording:

The extent of such manipulation needs more investigation, but our account points to the key conditions for assessing it: (1) Do elected officials find their own arguments suitable for denying climate change? (2) Are they conveying truthful and relevant information, at least to the best of their knowledge? A negative answer to either question means the politician is manipulating the public. Such manipulation matters because many people are influenced by politicians, news organizations and social media posts.

We can apply the same conditions to public policy examples. UK civil servant Nick Down was responsible for chasing down £600 million in unpaid personal income taxes and approached David Halpern’s Behavioural Insights Team (BIT) for help (Halpern Reference Halpern2015: Ch. 5). The tax office subsequently changed the text of the letter sent to taxpayers with overdue tax liabilities to include the truthful information that ‘nine out of ten taxpayers pay on time’. This nudge increased the payment rate by 4.5%. Although multiple agents were involved in designing and implementing the nudge, how far it may be deemed manipulative can be assessed by examining the intentions of key decision-makers, particularly Nick Down and the more senior civil servants involved. To see whether the nudge meets the SR condition, one should investigate whether Down considers that nine out of ten taxpayers paying taxes on time is a suitable reason for paying one’s taxes. We could infer the truth of the SR condition based on Down’s published statements, responses in interviews and his own behaviour when presented with such social proof. To see whether the nudge meets the TH condition, we should see if the information is true and believed by Down and assess its relevance. Although we would require further evidence for a definitive assessment, we find it plausible that the nudge can meet both SR and TH conditions.

Another communicative nudge example comes from Australia, where the government Behavioural Economics team designed an intervention to reduce online gambling. In a field experiment, they tested whether showing the gambler a statement summarizing their gambling history and transactions would reduce how much they gamble. The statement and the visual elements highlighting losses visually in red is presented in Figure 1. Online gamblers who were shown the statement ended up reducing the amount they bet by 7.6% (Commonwealth of Australia 2020).

Figure 1. Australian gambling nudge source: Behavioral economics team of the Australia Government, “Better Choices: Applying Behavioural Insights to Online Wagering” (December 2020), Figure 4, p.16 : https://behaviouraleconomics.pmc.gov.au/projects/applying-behavioural-insights-onlinewagering.

Once again, the nudge can meet the two conditions for non-manipulation, provided that the government agents approving its use (1) find the information provided to be a suitable reason to reduce one’s online gambling (SR condition) and (2) have selected information that is true and relevant to the gambler (that is, that the account balance actually reflects the individual’s gambling history) (TH). Again, both conditions could plausibly be met upon further empirical investigation.

A final communicative nudge example comes from the US Office of Evaluation Sciences (2018). The intervention involved mailing letters to seniors over 65 years of age, informing them of their potential eligibility for a Supplemental Security Income (SSI) managed by the Social Security Office. The letters informed seniors of their eligibility, mentioned the maximum benefit that they could be entitled to and noted that the application process is simple (see Figure 2) (Hemmeter et al. Reference Hemmeter, Safran and Wilson2020). As a result, SSI awards increased by up to 340%. Once again, assessing the degree of manipulation requires evaluating whether the reasons given meet the SR and TH conditions; in this case, too, it seems the conditions are met.

Figure 2. Supplementary Security Income Nudge Source: US Office of Evaluation Sciences, “Increasing SSI uptake among a potentially eligible population” https://oes.gsa.gov/projects/increasing-ssi-uptake/, based on Jeffrey Hemmeter, John Phillips, Elana Safran, and Nicholas Wilson, “Communicating Program Eligibility: A Supplemental Security Income (SSI) Field Experiment” (November 2020), p. 36.

We have discussed the conditions to attend to in determining whether a given communication is manipulative. Of course, real-world examples of communication involve the exchange of multiple propositions, some of which will meet the relevant conditions and some of which may not, and we may wish to focus on the preponderance and strength of one rather than another. Regardless of the complexities of any given case, however, a proper investigation of such communications should focus on the reasons given and the intentions of the presenter.

While these nudge examples from public policy are arguably not manipulative by our criteria, that does not, of course, mean others may not be manipulative, and we consider more examples below. Now, however, we turn to situational influence, which supplies the examples that are usually provided for manipulative nudges.Footnote ¹⁰

3. Conditions for Non-manipulation: Situational Influence

We can deploy the two conditions to draw distinctions between persuasion and manipulation. However, many forms of influence in politics and public policy do not provide evident reasons, but instead operate indirectly by influencing the choice context of targets. Many of the best-known nudges are situational. Take the following four paradigmatic cases:

• Opt-out Organ Donation. All citizens are considered organ donors upon their death unless they explicitly opt out of this status.

• Cafeteria. Healthy products in a cafeteria are displayed at eye level to make them more salient to consumers buying lunch.

• Slow down. Trees on the side of the road are planted closer together around sharper turns to get drivers to instinctively slow down.

• Fly in the urinal. A fly painted in the urinal helps men aim and keeps the bathroom cleaner.

Situational influence primarily relies on changes to the default rules or arrangement of alternatives to make certain choices more salient. They are not explicitly communicative between i and j because no overt reasons are provided to the target of the nudge. In some of these cases, the reason for the choice architecture might be considered evident. These days the fly in the urinal is well known and people understand its purpose (though that was not the case when it was first introduced in the Netherlands in the mid 1990s). Drivers are now well aware of lines painted closer together on roads to encourage them to slow down but might be less aware of the effects of trees. Most people do not understand the effects of placing items at eye level until it is pointed out to them. All these examples are designed to change behaviour and the reasons why the targets’ behaviour changes may not be evident to the targets themselves. In Organ Donor, an individual may never learn about this policy, particularly as it only affects them after their death.

The reasons behind other situational nudges are even less evident. Here are some noteworthy political examples:

• Ballot Order. The names of one party’s candidates are listed before the others. Voters with limited information tend to vote for the names listed higher up the ballot.

• Strategic Agenda-setting. The chair of a legislative session decides not to bring an agenda item to a vote since she expects the vote to go against her preference. She postpones the vote until a more favourable opportunity.

• Gerrymandering. Members of party i redraw the district map in a way that maximizes the number of seats won by their party given likely voting patterns from previous elections. They are not directly influencing any individual’s vote, but rather influence the vote count.

Ballot order is designed to change the behaviour of some voters. For most voters, the choice architecture is not evident; they have no idea the ordering of candidates is deliberate.Footnote ¹¹ In the other two cases, as we noted above, the intention of the agent is not to change behaviour, but to change the outcome given expected behaviour. Some may notice the changes to the voting rules, but without necessarily understanding their effects on the collective decision.

In general, situational influence is manipulative, even if not all situational influence is equally manipulative. Without evident reasons, situational nudges fail the TH condition of our account. Because no reasons are disclosed to the target, the agent i cannot intend to provide relevant and truthful information to j. However, situational influences can be made non-manipulative by adding a communicative element. If the explicit or implicit reasons are made public, then we can assess the reasons in terms of the TH and SR conditions. For example, a letter could be sent to citizens informing them of their enrolment as organ donors and justifying the change in default rules. (The government of Singapore, in fact, sends such letters to all citizens and permanent residents over the age of 21 (Government of Singapore 2013). One can interpret this move as making the recommendation sufficiently discernible to transform the situational influence into a communicative one.)

Even with situational influence, there are cases where the reasons are (or eventually become) sufficiently clear and the two conditions are met. Then we might say there is no manipulation. In these cases, the influence can be read as a recommendation (Levy Reference Levy2019). The clearest examples of defaults that entail recommendations come from online purchases where one is often asked ‘Would you like to protect your purchase with supplementary insurance/warranty etc.?’ and one of the choices (usually ‘yes’) is pre-selected, coloured green, or otherwise explicitly endorsed. In these cases, the default clearly comes with a discernible recommendation. To determine whether the influence is manipulative, we need to investigate whether the agent himself endorses the recommendation. For supplementary insurance, the answer is often no. In the organ case, if it is made clear that organ donation is recommended as the default and if the government agents endorse the recommendation, then we are dealing with non-manipulative influence. If the café makes clear that healthy options are put at eye level, then there is no manipulation. If, however, the café does not make clear that healthy options are at eye level or the organ-donation default is undisclosed, then the nudge is manipulatory. But how manipulatory?

Levy (Reference Levy2019) suggests that all situational nudges, including ballot order, can be seen as providing reasons. On his account, all external reasons, no matter that they are non-conscious or exist due to our social psychology or evolutionarily developed neural routines, constitute reasons. In his argument, such nudges therefore do not ‘bypass reason’. Of course, BSV proponents will be unimpressed by this argument since they mean that ‘conscious deliberative reason is bypassed’. Similarly, applying our conditions requires the presence of evident reasons that are given by the agent to the target (even if the reasons are external reasons). However, one way of closing the gap between the desideratum of the BSV proponents and our account in terms of external reasons is to consider how mild or strong manipulation is in such nudge cases. We cash this in term of a hypothetical disclosure test.

In our account, what matters for whether an act is one of manipulation is the intentions of the agent. If their evident communicative recommendation is based on suitable reasons and testimonial honesty, then it is not manipulative. If their non-evident situational recommendation is based on suitable reasons and testimonial honesty, then it is only mildly manipulative. Furthermore, the attitude of the target to those reasons can also be considered a test of how mild the manipulation is. We can consider these two aspects by way of a hypothetical disclosure test.

The hypothetical disclosure test is designed to expose the reasons why situational recommendation was made in relation to what the agent considers to be suitable reasons. We define such a test as the statement that could be made by i in accordance with the TH condition. In other words, the hypothetical disclosure statement represents what the agent i would say if she intended to provide relevant and truthful information to j. The extent of the manipulation depends on what we might call the ‘distance’ between the reasons in the hypothetical disclosure and reasons acceptable to i. To be sure, we are using the term ‘distance’ metaphorically. Measuring ‘distance’ precisely is difficult, so we rely on a psychological shortcut. We can think of this in terms of how unabashed the agent would be to disclose their honest and full reasons for the nudge.Footnote ¹² For the healthy food example, the disclosure would read: ‘We have placed healthy food at eye level as you are more likely to choose food at eye level and we wish to encourage healthy eating’. For ‘slotting’ behaviour in supermarkets: ‘since you are more likely to choose products at eye level, we place products at eye level according to the highest bidders’. The degree of the agent’s embarrassment at making such statements is one test of how manipulative the nudge is.Footnote ¹³

The second way of thinking about the hypothetical disclosure test is what effect it would have on the target if the disclosure were to be made. Would the target consider themselves to be manipulated if they were given the full facts? We surmise that, if the target identifies or agrees with the reasons behind the recommendation, they would not feel they were being manipulated, and if they disagree, they would feel manipulated. With nudges, once one knows what the agent is attempting to get one to do, the target can reject it. With healthy food, the target can see the suitable reason, but decide that on balance they prefer the unhealthy food.Footnote ¹⁴ With the slotting example, the target might feel they are being manipulated, even if they can still override the recommendation. They will know that, even as they try to override the recommendation, they might still be subject to its force. Thus, we can see why, in the example of Krstić and Saville (Reference Krstić and Saville2019) discussed by Klenk (Reference Klenk2022), the Twitter user knows the Twitter algorithm is promoting content that will ‘push her emotional buttons’, but still feels the force of that content. Her knowledge of the intentions of the agents setting up the algorithm does not make it feel any less manipulative, because she cannot override it despite disagreeing with the reasons behind setting up the algorithm.Footnote ¹⁵

Real disclosures of situational influence are more likely where the nudge is meant to help the customer or citizen make better decisions. The four nudge examples above seem to fit this description and therefore represent cases of minimal manipulation. Real disclosures are highly unlikely where the situational influence is designed to help the agent at the expense of the targets. For example, a political party is unlikely to issue the following disclosure, characteristic of gerrymandering:

The hypothetical disclosure test provides a conceptual way to get around the absence of explicit reasons. But all sorts of evidence can turn the hypothetical test into a real one. Leaked corporate memos, regulatory oversight, investigative journalism, public legal proceedings and scientific investigations are some of the many ways in which further information about the reasons behind certain policies, including nudges, eventually become publicly available. Moreover, democratic governments are usually quite open about their nudge policies, precisely because they are paternalistic: that is, they are designed to enhance citizen welfare.

The above helps explain why both critics and defenders of nudge are preoccupied with transparency conditions (Thaler and Sunstein Reference Thaler and Sunstein2008; Bovens Reference Bovens, Grüne-Yanoff and Hansson2009; Schmidt Reference Schmidt2017).Footnote ¹⁷ According to Bovens (Reference Bovens, Grüne-Yanoff and Hansson2009: 217), therefore, ‘every Nudge should be such that it is in principle possible for everyone who is watchful to unmask the manipulation’. However, this is a necessary, but not sufficient, condition for non-manipulation. Mere awareness of the influence is not enough without the reasoning behind it. Thaler and Sunstein also advocate the inclusion of a publicity condition: ‘In its simplest form, the publicity principle bans government from selecting a policy that it would not be able or willing to defend publicly to its own citizens’ (Thaler and Sunstein Reference Thaler and Sunstein2008: 244). Their publicity principle approximates our public relations test. However, Wilkinson (Reference Wilkinson2013) points out that this principle is insufficiently precise, and some governments seem comfortable defending extremely manipulative policies. It also leaves a lot of work to be done by the ability condition (‘able to defend’) that requires further normative theorizing to properly explicate. Our hypothetical disclosure statement better reveals what the analyst considers to be manipulative, recognizing that without disclosure there is always a degree of manipulation, however slight.

4. When is Manipulation Morally Justified?

We have argued that, without supplementary conditions regarding disclosure, situational influence will generally be classified as manipulative. However, the degree of manipulation varies, and some manipulation is only petty. Is such manipulation sometimes justified? An important advantage of our less moralized definition of manipulation is that it allows us to separate when influence is manipulative and then to what degree it is so, allowing us to assess the all-things-considered moral status of such influence. While our approach to whether influence is manipulative is focused on the agent, our assessment of the morality of manipulation focuses on the target. By first assessing whether a nudge is manipulative through the intentions of the influencer, and then determining how morally unacceptable the act is through the target’s perspective, our approach integrates both parties better than earlier accounts.

Much recent discussion of manipulation focuses on the nudge agenda. Here the primary moral concern about manipulative influence is that it reduces the autonomy of the target, where autonomy is generally understood as ‘the control an individual has over his or her own evaluations and choices’ (Hausman and Welch Reference Hausman and Welch2010: 128).Footnote ¹⁸ Non-manipulative influence allows targets to retain control over their own evaluation and choices. However, many argue that even manipulative nudges are morally acceptable under certain conditions where they either (1) do not reduce autonomy (Wilkinson Reference Wilkinson2013; Schmidt Reference Schmidt2017) or (2) the autonomy reduction is compensated for by countervailing welfare gains (Thaler and Sunstein Reference Thaler and Sunstein2008: 244).

Wilkinson (Reference Wilkinson2013) argues that manipulative nudges to which the target consents do not reduce autonomy and are therefore morally acceptable. Imagine, for example, giving permission to a coach or physical trainer to nudge you into exercising more frequently or eating more healthily. Our account will handle such cases by considering the role of the agent i. For example, if i is acting as the agent for the principal j, then i might well give advice in terms of j’s world view and not i’s. For example, a financial adviser might believe that investing in fossil fuels is the most lucrative form of investment, but knowing agent j’s environmental concerns will not suggest such investments. Our approach allows all such agents to be handled in this way, since the communications they give are given with the known interests of their principal in mind. They act as the representative of the principal.Footnote ¹⁹

The more complicated question involves weighing autonomy considerations against welfare ones. Instead of proposing an objective assessment of the relative weights of these two types of considerations, we prefer to use the target’s subjective assessments. Our account of manipulation enables people to identify and assess manipulatory practices. The types of judgements people make are likely to be heterogeneous. Some ‘libertarian’ citizens will weigh autonomy considerations more heavily, requiring a much larger degree of welfare gains to compensate for even small reductions in autonomy. Other ‘welfarist’ citizens will weigh welfare considerations more heavily and so be comfortable with more manipulation provided it results in welfare gains. We can expect the same heterogeneity to apply at the level of various constituencies or electorates. The median voter in the US will likely have a different autonomy-to-welfare tradeoff from the median voter in Sweden. Hence minimally or moderately manipulative nudges will be morally acceptable when implemented in countries where the average citizen weighs welfare more highly, but morally unacceptable in countries where they more strongly prefer autonomy. We consider this to be a helpful and heretofore overlooked consideration in the ethics of nudging.

In addition to our general principle of respecting the weights citizens assign to autonomy and welfare, we offer two further considerations as part of a normative assessment of the justification or appropriateness of situational influence. These are considerations that should guide the moral assessment of citizens and public officials, both with respect to policy tools and to regulating private-sector influence. Such assessment should involve these questions: (1) is this public policy more manipulative than the alternatives? And (2) are the manipulation criteria applied in (unjustifiably) asymmetric ways across cases?

We can transform these questions into conditions to determine when a manipulative situational influence is morally acceptable. Call the first the comparative manipulation condition. For example, for a nudge policy to be problematically manipulative, it would have to be more manipulative than the most likely non-nudge alternative. Some examples of regulation are transparent and meet the relevant conditions. A paradigmatic example of non-manipulative regulations is clearly presented and honestly communicated traffic signs. Not only do citizens have clear signs along every road communicating the relevant traffic laws, but all drivers receive mandatory training in interpreting such signs correctly, meeting the TH condition. Moreover, the reasons behind such rules are often straightforward and endorsed by the relevant decision-makers.

Contrast this with the gerrymandering example we discussed in section 3. New district maps can be drawn for plausibly suitable reasons (for example, to accommodate a growing population) or for plausibly unsuitable reasons (for example, to secure partisan political gains). Adding computer-generated ‘fair maps’ as recommendation nudges to influence the decision of redistricting committees would seem to constitute a move away from manipulation. Similar cases can be found across a range of regulations in domains as diverse as housing, health, education, family law and environmental politics. When scholars criticize nudges as being problematically manipulative, it is important to compare the proposed nudge with alternative regulatory mechanisms. In some cases, nudges might be more manipulative and therefore a matter of moral concern. In other cases, they might be less manipulative. Ceteris paribus, a manipulative nudge should be morally acceptable if it replaces a more manipulative regulatory alternative.

Call the second condition the symmetry condition, requiring the moral assessment of manipulation to be applied symmetrically in private- and public-sector contexts. Some believe that government manipulation is worse than private manipulation. The argument sometimes offered is that government has the monopoly on the legitimate use of violence, hence its acts are backed by these coercive powers. However, any manipulative act by a private firm or an individual, if legitimate, is also backed by the coercive power of the government. That is what entails its legitimacy.

The concern should always be: should this act of manipulation be legitimate? Thus, concern about government-run cafeterias arranging food to encourage healthy eating should also elicit concern if privately run cafeterias arrange the food to make high-value-added but unhealthy choices more salient. Similarly, concern about public-sector announcements cautioning citizens about the risks of smoking, alcohol or diets rich in salt, sugar and fat should equally elicit concern about private-sector advertising encouraging the consumption of these products through similar techniques. Of course, this condition does not imply that these situations are equivalent from an all-things-considered standpoint. Welfare-enhancing nudges have moral advantages over welfare-decreasing ones. Our condition merely articulates that the criteria for identifying manipulation and for judging whether a particular case of manipulation is morally justified should be applied consistently across other domains.

Some people hold an asymmetry thesis. Mark White, for example, argues that private-sector manipulation is morally better than government manipulation, because it is (1) non-coercive; (2) not paternalistic; and (3) expected and guarded against by consumers (White Reference White2013). His claim that manipulation in markets is non-coercive is based on competition: ‘If Jennifer gets sick of her coffee shop’s manipulative promotions, she can try to find another coffee shop with practices she likes better’ (White Reference White2013: 107). It is not paternalistic, because businesses have no intention of benefiting consumers unless it results in increased profits. Finally, and most importantly for White, consumers are already well equipped to deal with manipulation in markets, allowing them to minimize potential damage, whereas citizens’ assumptions about government benevolence make them exceedingly vulnerable: ‘the core difference between profit-motivated manipulation by businesses and paternalistic manipulation by government: we expect businesses to do it, but we expect more from our government’ (White Reference White2013).

All three arguments are problematic. As to the first, one can only avoid nudges if one is aware of them. Even then, one can only exit, as in the Jennifer example, if there is a non-manipulatory alternative – which is often implausible; consider, for example, how cars are uniformly marketed. It is not clear why nudges that are to the target’s benefit are morally wrong, but those that benefit the manipulators, even when known to harm consumers, are not wrong. Calling the former paternalistic does not make it worse than the latter. Furthermore, surveys do not suggest that the government is more trusted than private business.

Most importantly for our purposes, a classification of a given public policy as manipulative depends on the intentions of the putative manipulator. While it is true that the degree of success in manipulating others often depends on the degree of vigilance, experience and disposition of the targets, we do not classify an act as non-manipulative merely because its targets have become better at resisting. This is not to say that developing strategies for resisting manipulation is not worthwhile and should even be of interest to normative theorists. However, this relates to a different type of enquiry from ours.

Overall, situational nudges involve some degree of manipulation, particularly if the nudge is not explicitly disclosed. However, this does not mean the situational nudge is more manipulative than any potential policy alternatives. After all, the status quo might already be manipulative and new nudges could reduce manipulation overall. For example, government could provide eye-level slots in supermarkets for healthy food products, replacing current commercial slotting practices.Footnote ²⁰