Social Sciences and the Humanities

The previous chapters of this part looked inside one person. This chapter steps outside. Sociology, anthropology, law, linguistics, education, and their neighbors ask where a taste comes from, who is in the sample, what happens when the system that measures a preference also shows people what to choose, and what institutions that have collected human judgments for decades have learned. Each question presses on an assumption of preferential Bayesian optimization (PBO): in the formulation of Section 19.1, a person carries a latent utility function that stays fixed during the session, and every answer to "which is better?" is a noisy reading of it.

Three threads bear most directly on PBO. Machine learning has formalized the old observation that measuring something can change it, under the names performative prediction and induced preference shift (Section 42.2). European Union law has begun to prohibit interface designs that, in effect, distort users' autonomous decisions (Section 42.2.3). And information retrieval has accumulated evidence that judges make preference judgments faster and more consistently than graded ones (Section 42.4.1). Several widely repeated claims from these fields turn out to be wrong or overstated; Section 42.6 collects them. Disciplines with weaker links share one section (Section 42.5).

42.1 Where taste comes from #

If the latent utility is formed by class, culture, and life events, then a model of one person's answers is also a model of their social position, and a prior learned from other people carries those people's positions with it.

42.1.1 The sociology of taste #

The classic claim is that taste is socially produced and stratified by class. Bourdieu (1984) explained it with habitus, durable dispositions, acquired through upbringing and social position, that generate likings without deliberate calculation, and cultural capital, familiarity with legitimate culture that works like an asset. Hennion emphasized taste as an active practice that amateurs work at (Hennion, 2001; Hennion, 2007); reading this as saying that every preference query trains the user's taste is an extrapolation, not Hennion's argument. Refining the class picture, Childress et al. (2021) found that higher-status people are more inclusive at the level of genres and more exclusive at the level of specific objects such as particular artists and works. Airoldi's Machine Habitus treats machine learning systems as socialized agents whose training data steer how they classify and whose outputs feed back into culture (as summarized by Rödl, 2022), and Webster (2021) argued that personalization took over the labor of curating music, which weakens one traditional way of gaining distinction, by choosing for oneself.

What it means for PBO. In cultural domains such as music, fonts, or visual style, a surrogate defined on concrete parameters cannot see openness at the genre level, and one defined on genres cannot see exclusiveness at the object level, so the level of the features is a substantive modeling decision (inference). Priors built from population data or from a large language model encode the taste of whoever supplied the training data, so an informative prior (Section 31.2, Section 32.5) should be audited for whose taste it encodes (inference). And where part of a result's value comes from having curated it oneself, interaction in which the designer leads is favored (inference; compare Section 32.2).

42.1.2 Cultural psychology and WEIRD samples #

Citing an analysis of top psychology journals from 2003 to 2007, Henrich et al. (2010) reported that 96% of participants came from Western industrialized countries, which house 12% of the world's population, and called such societies Western, Educated, Industrialized, Rich, and Democratic (WEIRD). Little has changed: in six top psychology journals from 2014 to 2018, a little over 60% of authors and samples were American, and authors and samples from outside the English-speaking world and Western Europe remained at 4% to 5% (Thalmayer et al., 2021).

For aesthetic preferences, a 2025 paper at the annual meeting of the Cognitive Science Society gives unusually broad data. Lee et al. (2025b) collected 401,403 preference judgments from 4,835 participants in 10 countries across shape, curvature, color, musical harmony, and melody, sampling a two-dimensional parameter space for each. Liking for symmetrical forms was consistent across cultures; color preferences were consistent at the level of categories while "ratio-like" preferences varied; and melody showed the largest cross-cultural variation. The authors caution that online samples are exposed to global media, so the results need replication in laboratories and small-scale societies. Smell shows a similar split: when 225 people in 9 non-Western communities ranked ten odors by pleasantness, the identity of the odorant explained 41% of the variance in the rankings, individual variability 54%, and culture only 6% (Arshamian et al., 2022) (the institution's press release gives 235 participants (ScienceDaily, 2022)).

What it means for PBO. The pattern Lee et al. report suggests a hierarchical or multi-task Gaussian process, a shared component plus a group deviation (Section 20.5), with the expected size of the deviation depending on the domain: small for symmetry-like visual features, larger for melody and style; for smells, a model of the individual matters more than a cultural prior (inference). Most PBO user studies recruit Western university populations, so a finding such as "users prefer option A" is bound to that sample (inference). Rating scales are known to be vulnerable to response styles, culturally varying habits such as favoring the extremes or the middle of a scale; whether pairwise comparisons are less affected is open, as we found no study that tests it.

42.1.3 Demography and the life course #

Life-course research (Elder, 1998) and socioemotional selectivity theory, which holds that people shift toward emotionally meaningful goals as they perceive their remaining time as shorter (Carstensen et al., 1999), both predict that preferences change systematically over a life. Schildberg-Hörisch (2018) summarized that risk preferences are persistent but only moderately stable, and suggested treating a preference parameter as a distribution with a fairly stable mean and systematic variance around it. Using Australian panel data, Kettlewell (2019) found that financial changes, the birth of a child, and the death of someone close shift risk preferences, with effects largest near the event and fading afterward. A test-retest correlation measures how well a measurement agrees with itself when repeated on the same people later; for risk preference, the meta-analysis of Bagaïni et al. (2025) estimates a reliability of 0.61 for self-reported propensity to take risks and 0.25 for behavioral measures such as lottery choices.

What it means for PBO. For personalization that runs over months or years, such as tuning a hearing aid or an exoskeleton (Section 33.4, Section 33.1; Chapter 24 works the exoskeleton case end to end), a model with a stable component plus deviations triggered by events that then decay fits this evidence better than a random-walk time kernel that lets preferences drift anywhere (inference; Section 29.10 and Section 46.5 discuss time-varying models). For a single session, the connection is weak.

42.2 When measuring changes what is measured #

This group contains the chapter's strongest links. Its disciplines study the same loop from different sides: a system observes people's choices, acts on what it learned, and in acting changes the choices it will observe next. A PBO session is a small instance of that loop.

42.2.1 Opinion dynamics and recommender feedback #

In the bounded confidence model of opinion dynamics, agents move toward others' opinions only when those opinions are within a confidence bound of their own; with a large bound the population reaches consensus, with a medium bound it polarizes, and with a small bound it fragments (Deffuant et al., 2000). A PBO system serving many users should expect clusters rather than one shared optimum.

Two famous empirical claims need care. The finding of Christakis and Fowler (2007) that obesity spreads through social networks, a person's chance of becoming obese rising by 57% if a friend became obese, faces an identification problem: homophily (similar people become connected) and contagion (connected people become similar) leave the same trace in observational network data, and Shalizi and Thomas (2011) proved that the two are generally confounded there. And a summary in circulation says that on Facebook, individual choice limited exposure to ideologically diverse content by 70% and the ranking algorithm by 15%. Those numbers are not in the source: Bakshy et al. (2015), analyzing 10.1 million American users, found that algorithmic ranking reduced cross-cutting content by 5% for conservatives and 8% for liberals, and individual choice by 17% and 6%: choice outweighed the algorithm for conservatives, not for liberals.

The theory since 2017 writes preference dynamics into the recommender's objective. Dean and Morgenstern (2022) assumed that a user's preferences move toward content they consume and enjoy and away from content they consume and dislike, and proved that under these dynamics standard reward maximization is "an almost trivial goal": a large class of simple algorithms achieves constant regret. (Regret, from Chapter 13, is the utility lost by recommending what the system recommended instead of the best option.) The more meaningful goal, they argue, is to keep the user's preferences approximately stationary. Carroll et al. (2022) pointed out that a recommender optimized over a long horizon has an incentive to push users' preferences toward ones that are easier to satisfy, and proposed constraining the system to a trust region of "safe shifts", for example the shifts that would have happened without the system. Kleinberg et al. (2023) proved that when users' preferences are inconsistent, a platform can raise engagement while lowering user utility, and Carroll et al. (2024) and Williams et al. (2025) showed that a static preference objective implicitly rewards influencing the user and that optimizing on user feedback can learn targeted manipulation.

The mechanism behind the "almost trivial goal" is simple enough to watch. Figure 42.1 puts a simulated person with plastic preferences in front of a deliberately simple optimizer and scores the result twice.

Things to try:

• Press Fixed preferences. The blue and magenta curves coincide, which is the world PBO assumes; the simple local search is still far from the ideal after 40 comparisons, with mean regret 0.43.
• Return η to 0.1. The blue curve falls to about 0.014: by the benchmark, the same weak optimizer now looks excellent, much better than against a fixed person, while judged by the preferences the person arrived with its regret is 0.76. The person moved about 0.36 toward the system, more than the system moved toward the person.
• Set the challenger to Anywhere. The system reaches the person's region before the person has moved far, and the gap between the two scoreboards shrinks (about 0.006 against 0.14 at η = 0.1). How far the person moves depends on which queries the system asks.
• Set the starting design to 0.75. When the system starts where the person already is, there is little to steer.

Does this happen outside a simulation? The evidence splits by design and timescale (Table 42.1). Simulations, observational studies, and short experiments find homogenization, including stories that became more similar to each other when writers used ideas from generative AI (Doshi and Hauser, 2024). Large randomized field experiments over weeks to months mostly find small or null effects on attitudes (Guess et al., 2023), including eight attitude measures that were preregistered, declared before the data were collected, which guards against searching for a positive result (Nyhan et al., 2023).

What it means for PBO. Dean and Morgenstern's result means that low regret can be achieved by moving the user's preferences, so regret or speed of convergence alone is not a sufficient criterion of success. A study should also report how much the preference moved during the session, against a control group with a random or balanced query order and a delayed retest of options the person rejected early, the PBO counterpart of Carroll et al.'s "safe shift" baseline (inference). The timescales differ, though. Platform studies concern weeks to months, while a single PBO session is closer to revaluation within a session, such as the drift in subjective values during a single experimental session that Zylberberg et al. (2024) traced to how value is constructed during deliberation. The link is therefore strong for personalization over months and weaker for a one-off design session (inference). With many users, the bounded confidence results favor mixture or clustered utility models, and argue for not showing users what others chose while eliciting their preferences (inference). Section 36.4 follows the same loop in the recommender systems literature.

42.2.2 Science and technology studies #

Science and technology studies (STS) examines how technical artifacts and social order shape each other. One story often used to show that artifacts carry politics, that Robert Moses built low overpasses on the parkways to Jones Beach so that buses could not reach it (Winner, 1980), is disputed (Joerges, 1999) and should not be retold as fact without that caveat. The general point survives: instruments of measurement take part in producing what they measure, as traders' use of option-pricing theory helped make market prices conform to the theory (MacKenzie and Millo, 2003).

The most important development since 2017 is that machine learning gave this idea a formal theory. Perdomo et al. (2020) called a prediction performative when it influences the outcome it aims to predict, and defined a stability notion for it.

The natural algorithm is repeated risk minimization: deploy, collect the data the deployment induces, retrain, and repeat. Perdomo et al. give conditions that are both necessary and sufficient for it to converge to a stable point whose loss is nearly minimal. Informally, the distribution map is $\varepsilon$-sensitive if moving the parameters by a distance $d$ moves the induced data distribution by at most $\varepsilon d$ in Wasserstein distance; if the loss has curvature at least $\gamma$ ($\gamma$-strongly convex) and a gradient that changes at rate at most $\beta$ ($\beta$-smooth), repeated risk minimization converges at a linear rate to a unique performatively stable point when $\varepsilon < \gamma / \beta$, and it can fail to converge when $\varepsilon \ge \gamma / \beta$. The world's sensitivity to the model must be small relative to how sharply the loss pins the model down. Hardt and Mendler-Dünner (2025) connected the concept to performativity in economics and the social sciences and distinguished two mechanisms, learning (predicting the world better) and steering (moving the world toward the prediction). Asking about behavior also changes it: a meta-analysis of 116 tests of the question-behavior effect found that asking people about their intentions or to predict their own behavior changes their later behavior slightly, with Cohen's $d = 0.24$, a difference between groups in units of the standard deviation (Wood et al., 2016).

What it means for PBO. This is the chapter's strongest formal link. A PBO loop whose queries change the person's preferences is a performative prediction problem: the surrogate's fixed point is the preference that is stable under the queries the surrogate itself induces, which need not be the preference that a procedure without influence would find. In Figure 42.1, the local optimizer ends at such a point. Perdomo et al.'s condition is a candidate criterion for whether the loop still converges when preferences are co-constructed by the system and the person, and the distinction between learning and steering can be used to evaluate PBO directly (inference). The choice of success metric is itself performative: regret against the final inferred utility, and the concentration of the posterior, reward steering the person toward a predictable state, while endorsement at a delayed retest and comparison with a random-order control group are much harder to steer (inference; Section 46.7).

42.2.3 Law #

Legal templates such as Alexy's proportionality test with its weight formula (Alexy, 2024) invite analogies with preference aggregation, but the substantive development is European Union law on manipulation and dark patterns, interface designs that steer users into choices they would not otherwise make, such as pre-ticked boxes or cancellation flows that are harder than sign-up.

In the Planet49 case (C-673/17, 1 October 2019), the Court of Justice of the European Union held that a pre-ticked checkbox does not constitute valid consent, which requires active behavior (Etteldorf, 2019). Article 25(1) of the Digital Services Act (Regulation (EU) 2022/2065) provides that "providers of online platforms shall not design, organise or operate their online interfaces in a way that deceives or manipulates the recipients of their service or in a way that otherwise materially distorts or impairs the ability of the recipients of their service to make free and informed decisions" (European Union, 2022a); paragraph 2 excludes practices already covered by the Unfair Commercial Practices Directive or the General Data Protection Regulation. Recital 67 defines dark patterns as practices that materially distort or impair the user's ability to make autonomous and informed choices "either on purpose or in effect" (European Union, 2022b). Article 5(1)(a) of the AI Act (Regulation (EU) 2024/1689) prohibits AI systems that deploy subliminal techniques, or purposefully manipulative or deceptive techniques, with the objective or the effect of materially distorting behavior by appreciably impairing a person's ability to make an informed decision, in a manner that causes or is reasonably likely to cause significant harm; Article 5(1)(b) covers the exploitation of vulnerabilities due to age, disability, or a specific social or economic situation. The prohibitions apply from 2 February 2025 (European Union, 2024); the Commission's guidelines on them, published on 4 February 2025, are non-binding (European Commission, 2025). A Digital Fairness Act covering dark patterns, addictive design, and unfair personalization has not been adopted; the European Parliament's legislative tracker, updated 20 June 2026, lists the proposal as expected in the fourth quarter of 2026 (European Parliament, 2026).

The effects these rules target can be large. In the experiments of Luguri and Strahilevitz (2021), mild dark patterns more than doubled the rate at which people signed up for a dubious service and aggressive ones nearly quadrupled it. Mathur et al. (2019) found 1,818 instances of dark patterns, in 15 types, on about 11,000 shopping websites. Esposito et al. (2026), by contrast, argue that Article 25 is vague, narrow in scope, and weakened by the exclusion in its paragraph 2.

What it means for PBO. The following are inferences from the text of the law, not legal advice. A PBO system deployed on an online platform decides which options a user sees and in what order, so it is part of the design of an online interface. If its acquisition function systematically steers users, for example by repeatedly reinforcing the incumbent or by favoring options that make the user more predictable, the words "in effect" mean that intent need not be shown under Article 25. The threshold of Article 5 of the AI Act, purposeful manipulation or subliminal technique combined with significant harm, is much higher, and ordinary design tuning is unlikely to reach it; a Digital Fairness Act may reach personalization by optimization directly (inference). By the logic of Planet49, an option the user "accepts" at the end of a long, system-led sequence of queries is only weak evidence of preference if the sequence was steering; logging a balanced query order and ending with an explicit endorsement check serve scientific validity and compliance at once (inference; Section 46.8). We have not reviewed the full text of the Commission's 2025 guidelines, and we found no court or regulatory decision that applies Article 25 to a recommendation or preference elicitation algorithm rather than to static interface elements.

42.2.4 Rhetoric and argumentation #

Rhetoric holds that preferences change through reasons exchanged in dialogue, a mechanism a model of fixed utility ignores. (The argumentative theory of reasoning of Mercier and Sperber (2011) is often compressed to the claim that reasoning evolved to persuade others; the theory concerns both producing arguments and evaluating them.) Classical PBO has no channel for reasons, but BO systems that add one through large language models have appeared (Section 35.2), and conversational systems move views. Costello et al. (2024) had 2,190 people who believed in a conspiracy theory hold three rounds of personalized dialogue with GPT-4 Turbo; belief fell by about 20%, and the effect lasted at least 2 months. Tessler et al. (2024) trained a language-model mediator, the "Habermas Machine", with a reward model to maximize the group's endorsement of statements. Participants ($N = 5{,}734$, in the United Kingdom) preferred its group statements to those written by human mediators, discussants' views converged after mediation, and dissenting views were incorporated into the successful statements.

What it means for PBO. An interface that explains ("this option is brighter because ...") or presents options through a language model is itself a channel of persuasion, so the presence of an explanation should be randomized, or at least recorded, so that its effect can be separated from the comparison data (inference; Section 32.7). The pipeline of Tessler et al. (generate candidates, collect endorsements, refine) is close to PBO for a group, and its convergence can be read as common ground or as homogenization caused by the optimizer, so group PBO needs a criterion such as "are minority positions retained", which Tessler et al. measured (inference).

42.3 Linguistics #

When feedback is given in words, language shapes what is measured. Verbal overshadowing is the finding that describing something can impair later recognition of it: in Schooler and Engstler-Schooler (1990), participants who described a robber's face were about 25% worse at picking him out of a lineup. That figure is correct for the original study but overstates the effect. A multi-laboratory registered replication (Alogna et al., 2014) found that participants who described the robber were 4% less likely to identify him when the description immediately followed the event, and 16% less likely when it came after a 20-minute delay: robust, and strongly dependent on timing. Wilson and Schooler (1991) found that introspecting about one's reasons can lower the quality of preferences and choices. Gradable adjectives, words such as "tall" or "warm" whose meaning is a threshold on a scale, have been modeled with an uncertain threshold (Lassiter and Goodman, 2017) that varies with context and with the adjective (Xiang et al., 2022).

What it means for PBO. When a system accepts graded verbal feedback such as "slightly warmer" or "much better", the likelihood should treat the magnitude a word expresses as uncertain and relative to the current context of comparison, not as a fixed increment of utility (inference). Asking users to write down reasons before they choose may change the judgments that follow, so reasons are better collected after the choice (inference).

42.4 Institutions that already collect judgments #

Search engines, schools, and armies have gathered human judgments at scale for a long time. Their experience speaks to which question to ask, how to read the answer, and when a person's preference is the wrong target.

42.4.1 Information science #

Information retrieval is the discipline closest to the lineage of PBO: it framed interactive optimization as a dueling bandit problem (Yue and Joachims, 2009), in which the learner sees only which of two options wins (Chapter 21). Its evaluation research since 2017 gives direct evidence for preference judgments. Sakai and Zeng (2020) found that the best preference-based evaluation measures agree with users' preferences between search result pages at least as well as an average assessor does. Clarke et al. (2021) found that assessors make preference judgments faster and more consistently than graded judgments, and that preferences can distinguish items that graded judgments treat as equivalent; but fully ordering a pool by preferences takes more than linear effort. Partial preference judgments that identify and order only the top items detected improvements in neural rankers that the standard graded measure (normalized discounted cumulative gain) missed.

Language models as judges have produced results in both directions. Thomas et al. (2024) report that at Bing, relevance labels from a large language model were as accurate as those of human labelers, but that simple rewordings of the prompt changed the accuracy. A 2024 preprint (Upadhyay et al., 2024) found that in the retrieval-augmented generation track of the Text REtrieval Conference (TREC) 2024, the system ranking produced by an automatic judge correlated highly with the ranking from fully manual judgments. Another 2024 preprint (Clarke and Dietz, 2024) built a system designed to exploit automatic evaluation, obtained inflated scores, and pointed out that if every system uses the same language-model judge as a reranker, the resulting rankings are distorted by circularity.

What it means for PBO. The retrieval evidence supports PBO's core design: pairwise queries aimed at locating the best region rather than learning the whole utility well. When the goal is selection, the acquisition function should target identifying the top few options rather than global accuracy (inference; Section 19.4). When a language model serves as a proxy oracle of preference, the final evaluation must use independent human judgments, or it inherits the circularity Clarke and Dietz describe; and sensitivity to the prompt means a language model's "preferences" are partly products of the prompt (inference; Section 30.6, Section 31.7). Section 36.6 follows learning to rank, the machine learning side of the same field.

42.4.2 Education #

Education research supplies the clearest evidence that what people prefer and what works for them can come apart. In a randomized comparison in a large introductory physics course, Deslauriers et al. (2019) found that students in active-learning classes learned more but felt they had learned less, partly because of the greater cognitive effort; the authors warn that evaluating teaching by students' perceptions may favor inferior passive teaching. On learning styles, the claim that instruction matched to a student's preferred modality works better, the meta-analysis of Clinton-Lisell and Litzinger (2024) (21 studies, 1,712 participants) found an overall benefit of matching of $g = 0.31$ (95% confidence interval 0.05 to 0.57), where Hedges' $g$ is an effect size like Cohen's $d$; but only 26% of the outcome measures showed the crossover interaction the hypothesis requires, in which each group does better with its own matched instruction, and the authors concluded that the benefits are too small and too infrequent to justify widespread adoption.

What it means for PBO. Tuning instructional parameters with the learner's preference as the objective risks optimizing the feeling of learning. The objective should be the learning outcome, with preference as a constraint or a secondary objective. Bayesian optimization with preference exploration (BOPE) has this structure: a model maps parameters to measured outcomes, and the person compares outcome vectors rather than parameters (Lin et al., 2022) (inference). Where better options take more effort to appreciate, early pairwise judgments may penalize them, and a delayed retest is a safeguard (inference).

42.4.3 Military decision-making #

A strong claim from research on expert decisions is that experts under time pressure never compare options. Klein's original study of fireground commanders, a 1985 report reprinted in 2010 (Klein et al., 2010), covered 156 decision points: evidence of concurrent comparison of options appeared at fewer than 12% of them, and at more than 80% the commanders recognized the situation as typical and identified the typical action directly. "Never compare" is an overgeneralization. According to a secondary summary with page references (Ambur, 2004), Klein's 1998 book reports that the share of recognitional decisions across studies ranged from 46% to 96%, that people compare options more when they must justify a choice, seek the best option, or meet an unfamiliar situation, and that they use a "face-off" procedure in which one option is compared with a second and the winner meets the next.

What it means for PBO. Searching for the best option in an unfamiliar continuous design space is exactly when Klein expects comparison, and the incumbent-versus-challenger query of many PBO systems matches the face-off procedure (inference). For experts in familiar domains, feedback on whether a single option is acceptable fits better (inference; Section 32.4), and expert status does not justify lowering the noise parameter in a new design space (inference; Section 32.6).

42.5 Briefer links #

Six more disciplines offer one lesson each for PBO.

Museums. Context shifts liking more than it shifts rankings. The same exhibition was liked more in a museum than in a laboratory simulation (Brieber et al., 2015), while a rehang of the Belvedere in Vienna increased viewing time but did not change 259 visitors' interest in specific artworks or their preferences among art forms (Reitstätter et al., 2020). Pairwise comparisons made within one context should transfer better than absolute ratings made across contexts (inference).

Geography. People choose from the options they actually consider. Treating that consideration set as latent, Tsoleridis et al. (2023) found that ignoring constraints on it misestimated preferences. Comparisons involving options outside the range a user considers may be close to random, which a likelihood with a lapse component handles (inference; Section 27.2).

Archaeology. Many different processes of cultural transmission produce the same population-level frequency patterns, so matching a pattern is weak inference (Kandler and Crema, 2019). Pooled comparisons from many users likewise cannot pin down both the individual utilities and the rule that aggregated them, in line with the analysis of hidden context by Siththaranjan et al. (2024) (inference).

Semiotics. Consumer culture theory studies how people use goods to make identities and social meanings (Arnould and Thompson, 2005; McCracken, 1986). If the target of optimization lives in a space of such meanings, a Gaussian process on semantic embeddings should predict held-out comparisons better than one on physical design parameters, a testable comparison (inference).

Narratology. Under radical uncertainty, when outcomes cannot be enumerated and probabilities cannot be assigned, people choose by constructing narratives (Johnson et al., 2023). A typical PBO problem is not radically uncertain, but a short elicitation of purpose and audience before the pairwise queries helps when the user is not yet clear what the design is for (inference).

Organization theory. Adaptive processes refine exploitation faster than exploration, which is effective in the short run and self-destructive in the long run (March, 1991) (Section 11.3). An exploitative acquisition function combined with revaluation after choice will lock in the incumbent prematurely, as the local optimizer does in Figure 42.1, and an exploration budget that does not depend on the posterior counters it (inference).

42.6 Common claims, checked #

The claims in Table 42.2 are found in writing that applies these disciplines to preference learning and recommendation. Each was checked against its source; the sections above give the details.

None of these corrections removes a general point: algorithms still embed choices of value without the Moses story, and verbal reasoning still affects perceptual judgment at 4% to 16%. What changes is how strongly each point can be asserted, and that is the currency of a model's prior (inference).

42.7 Settled, contested, missing #

Further reading #

• Perdomo et al. (2020) and the review by Hardt and Mendler-Dünner (2025) are the formal core of this chapter: what it means for a model to be stable when its deployment changes the data.
• Dean and Morgenstern (2022) and Carroll et al. (2022) show, in recommender settings, why regret stops measuring success when preferences move, and how to bound the shift a system induces.
• Guess et al. (2023) and Nyhan et al. (2023) are the large randomized field experiments to read before claiming that systems reshape what people want.
• Clarke et al. (2021) is the clearest evidence from information retrieval for preference judgments, including their cost.
• Lee et al. (2025b) (CogSci 2025) is the broadest cross-cultural data on aesthetic preferences over parameterized stimuli, close in form to PBO.