Algorithm Aversion: Why Humans Reject Helpful Models

People who watch an algorithm make a mistake become less willing to use it — even when they can see, in the same data, that the algorithm still outperforms the human judges they would otherwise rely on. This article traces that finding from its original 2015 forecasting experiments through a decade of boundary-condition work, distinguishes it from the opposite pattern of algorithm appreciation, examines whether generative AI inherits or escapes the same dynamics, and argues that the standard deployment prescription — "give users control to reduce aversion" — is a narrower lever than it is usually presented to be.

Coverage note: verified through May 2026. Generative-AI-specific findings from 2025 onward are labeled as emerging evidence and weighted lower than the predictive-AI literature.

1. The canonical finding

The term algorithm aversion enters the behavioral literature with Berkeley Dietvorst, Joseph Simmons, and Cade Massey's 2015 paper in the Journal of Experimental Psychology: General (Dietvorst, Simmons, and Massey 2015). Across five studies the authors put participants in incentivized forecasting tasks — predicting MBA student performance from admissions data, predicting state-level airline traffic — and varied what participants observed before being asked to choose between their own forecast and an algorithm's forecast for a payout-relevant decision.

The key manipulation was visibility of error. In one condition participants saw the algorithm forecast a series of cases, with its errors plainly displayed. In another they saw only their own errors. In a third they saw both. The algorithm outperformed human judges on average in every condition. Despite this, participants who had watched the algorithm err were dramatically less likely to choose it for the payout-relevant forecast than participants who had only seen the human forecaster err — even though they had observed, on the same screen, that the algorithm's errors were smaller.

This is the precise behavioral pattern the term denotes. It is not "people dislike algorithms." It is asymmetric forgiveness: a single visible miss costs an algorithm more share-of-reliance than a comparable miss costs a human. Dietvorst and colleagues found the effect in participants who had economic incentives to choose the more accurate forecaster, and the effect held in expert as well as lay samples.

Three features of the setup are important when porting the finding elsewhere:

The model was, in fact, more accurate. Participants were not exercising calibrated skepticism toward an inferior tool. The task is engineered so that rejection is empirically costly.
Accuracy was scoreable. Each forecast had a ground-truth outcome and a clear loss function.
The choice was atomic. Participants chose which forecaster's number to use, not how to combine, edit, or contest it.

Most real-world AI deployments violate at least one of these conditions. The article returns to that point in §6, but the canonical finding should be stated narrowly: under conditions of measurable accuracy, observed error, and discrete choice, people under-rely on a superior algorithm after seeing it fail. Anything broader is extrapolation.

2. Distinguishing algorithm aversion from adjacent phenomena

A persistent failure in popular AI commentary is treating algorithm aversion as a synonym for AI distrust, technophobia, or resistance to automation. None of those equivalences holds. The phenomenon is narrower than "distrust" and broader than "post-error rejection in forecasting tasks," and it sits next to several patterns that produce the opposite behavior or the same behavior for different reasons.

Pattern	Definition	When it appears	Standard citation
Algorithm aversion	Reduced reliance on an algorithmic advisor after observing its errors, even when it outperforms the human alternative on average.	Forecasting and judgment tasks with visible error, scoreable accuracy, and discrete choice between algorithmic and human advice.	Dietvorst, Simmons, and Massey 2015
Algorithm appreciation	Stronger weight given to algorithmic advice than to advice attributed to humans, especially in numeric estimation and external-advisor settings.	Lay participants estimating numeric quantities; advice-taking paradigms; non-expert raters; pre-error baselines.	Logg, Minson, and Moore 2019
Automation bias	Excess reliance on automated suggestions, including failures to detect or correct automation errors.	Cockpit, ICU, and other high-workload, time-pressured monitoring tasks with default-accept interfaces.	Parasuraman and Manzey 2010
Selective adherence	Differential acceptance of algorithmic recommendations as a function of whether they confirm prior beliefs or in-group expectations.	Public-sector decision support, hiring, parole, where decision-makers have priors that algorithms confirm or contradict.	Alon-Barkat and Busuioc 2023
Calibrated reliance	Reliance on a model that tracks its actual accuracy, including appropriate override when the model is outside its competence.	The normative target of trust-calibration research; not a spontaneous behavior.	Lee and See 2004

The taxonomy matters because the four behaviors call for opposite interventions. A user exhibiting algorithm aversion needs evidence that the model is in fact more accurate, ideally including evidence about how its errors compare in magnitude and frequency to the human alternative. A user exhibiting automation bias needs friction, accountability for outcomes, and structured disagreement prompts. A user exhibiting selective adherence needs depersonalization of the recommendation channel and outcome tracking. None of these is "more trust" or "less trust" in a unidimensional sense; they are reliance disorders with different signatures.

The taxonomy also constrains what counts as a successful intervention. A design that increases adoption could be:

Correcting aversion (the model is good, users now use it appropriately);
Inducing automation bias (the model is good but users now defer to it past its competence);
Manufacturing selective adherence (the model is now used when it agrees with existing priors and overridden when it doesn't);
Producing genuine calibrated reliance (use and override track accuracy).

Without behavioral measurement after visible error, these are indistinguishable from a metrics dashboard.

3. Boundary conditions: when aversion appears, when it doesn't

The decade since Dietvorst et al. has clarified that algorithm aversion is conditional. The phenomenon does not present uniformly across tasks, users, source labels, or interface designs. The empirical literature has identified seven boundary conditions that meaningfully change whether and how strongly the asymmetric-forgiveness pattern emerges.

3.1 Error visibility

The Dietvorst effect requires that participants observe the algorithm's errors. In conditions where neither forecaster's track record is visible, baseline preferences dominate, and these baselines are often not anti-algorithmic — in some studies, untested algorithms are preferred to untested humans for numerical tasks. The aversion is post-evidence, not anti-machine prior. This is why presenting accuracy statistics without making errors locally visible is a partial intervention: it tells the user the average but not how the average decomposes.

3.2 Outcome control and bounded adjustment

The cleanest follow-up to the canonical finding is Dietvorst, Simmons, and Massey's 2018 Management Science paper "Overcoming Algorithm Aversion: People Will Use Imperfect Algorithms If They Can (Even Slightly) Modify Them" (Dietvorst, Simmons, and Massey 2018). Across forecasting tasks, participants who were allowed to adjust an algorithm's forecast by even a small bounded amount — sometimes capped at 2 or 10 percentage points — were substantially more likely to use the algorithm than participants offered the algorithm's output as-is. The effect held when the available adjustment was so tightly capped that it could not change the forecast meaningfully.

Two readings of this result coexist in the literature and should be held in tension. The optimistic reading is that bounded adjustment restores agency without sacrificing model value, producing higher adoption of a better forecaster at low accuracy cost. The pessimistic reading is that bounded adjustment is, in effect, an agency placebo: it raises uptake without proving that the resulting human-algorithm team is more accurate, fairer, or safer. Dietvorst et al.'s own data are friendlier to the optimistic reading in forecasting tasks where adjustment is small and the model is genuinely better. They do not generalize to settings where users have rich domain priors that may override the model in systematic, harmful ways, nor to settings where "user adjustment" functions as legal cover for institutional automation.

The design takeaway is therefore narrower than it is usually stated: constrained, logged, outcome-measured adjustment can convert aversion into use without large accuracy costs in forecasting-shaped tasks. Unconstrained override, or adjustment without outcome tracking, is an adoption lever, not a decision-quality lever.

3.3 Task objectivity and perceived human uniqueness

Castelo, Bos, and Lehmann's 2019 Journal of Marketing Research paper "Task-Dependent Algorithm Aversion" (Castelo, Bos, and Lehmann 2019) is the most useful boundary-condition study for deployment teams. Across nine studies the authors show that aversion is markedly stronger on tasks perceived as subjective — dating recommendations, joke ratings, ethical judgments — than on tasks perceived as objective — predicting outcomes, calculating values. They further show that this gradient is mediated by lay theories about whether the task requires "uniquely human" capabilities such as intuition, emotion, or empathy.

Two features of the Castelo result deserve emphasis. First, the subjective/objective gradient is about perceived objectivity, not actual objectivity. Tasks that have measurable ground truth but are folk-categorized as subjective (e.g., predicting which date will be enjoyable) elicit aversion. Tasks that are folk-categorized as objective elicit appreciation even when the underlying problem is poorly defined. This means deployment teams can sometimes reduce aversion by recasting how a task is framed, without changing what the model does — a manipulation that should be reported transparently rather than treated as a neutral UX improvement.

Second, the uniqueness effect is heterogeneous across domains. Longoni, Bonezzi, and Morewedge's 2019 Journal of Consumer Research paper "Resistance to Medical Artificial Intelligence" (Longoni, Bonezzi, and Morewedge 2019) finds that patients resist AI-based medical care because they perceive their own medical case as distinctive in ways an algorithm cannot capture — a phenomenon the authors call uniqueness neglect. Reducing uniqueness neglect (by framing AI as personalized, or by providing physician oversight) reduces resistance. But it does not follow that the resistance was epistemically wrong; medical AI tools that perform well on benchmarks can underperform on individual patients precisely because population averages obscure subgroup variation. The Castelo and Longoni findings are best read together: subjectivity and uniqueness produce aversion, but in some domains the aversion is tracking real features of the deployment problem rather than just a folk theory.

3.4 Transparency, explanation, and understandability

Transparency is the most-recommended and least-empirically-supported intervention against algorithm aversion. The naive prescription — "show users why the model made its decision and they will trust it more" — does not hold up uniformly in the data.

Yeomans, Shah, Mullainathan, and Kleinberg's 2019 Journal of Behavioral Decision Making paper "Making Sense of Recommendations" (Yeomans et al. 2019) shows that perceived understandability is a meaningful predictor of acceptance for joke recommenders, and that providing transparency about how the algorithm works increased preference for the algorithm. But this finding lives in a low-stakes recommender setting. Lehmann, Klaas, and Klesel's 2022 Production and Operations Management paper on transparency in algorithm-assisted decision making (Lehmann, Klaas, and Klesel 2022) finds non-monotonic effects: revealing a too-simple model can reduce advice use, because users update against the model when they see how parsimonious it is. More recent pre-registered work on explainable AI interventions has been mixed, with several studies finding that explanations increase confidence without improving accuracy of reliance — a signature of overreliance, not calibration.

The cleanest current statement of the evidence is: transparency is a tool for auditability and calibration, not a generic adoption lever. Performance information (comparative accuracy, error profiles, uncertainty) tends to move reliance more reliably than process-transparency (algorithm internals, feature importance, counterfactual explanations) — though both can produce overreliance if the displays signal authority more than they signal limits.

3.5 Stakes and moral salience

Algorithm aversion is consistently stronger on decisions with high stakes, irreversibility, or moral salience. Bigman and Gray's 2018 Cognition paper "People Are Averse to Machines Making Moral Decisions" (Bigman and Gray 2018) shows that participants oppose machine involvement in moral decisions even when assured the machine matches human accuracy. The aversion is not eliminated by performance information; it tracks the perceived appropriateness of the decision-maker rather than its competence.

This is the cleanest case where collapsing aversion into "bias" is empirically and normatively wrong. In moral and high-stakes settings, users may be exercising a categorically different judgment than accuracy comparison — a judgment about who is permitted to be a decision-maker, who can be held accountable, and what reasons can be demanded of a decision. §7 returns to this point.

3.6 Source label and identity

Aversion responds to labels independently of behavior. In one of the cleanest recent demonstrations, blind evaluations of identical text often rate AI-authored advice as good or better than human-authored advice; the same text rated lower when the source is disclosed as AI. Osborne and Bailey's 2025 Scientific Reports paper on personal advice from large language models (Osborne and Bailey 2025) is one of several recent studies finding source-label penalties that survive blind quality controls. This is "aversion to the label," not aversion to the content.

The label-effect literature complicates two common interventions. First, "disclose that the system is AI-generated" — often presented as ethical baseline — can substantially reduce uptake of equally good or better outputs. The right ethical response is disclosure; the practical effect on adoption should be reported, not assumed away. Second, "make the AI more human-like" — frequently proposed as a UX fix — can reduce aversion in some settings (Castelo on perceived warmth) and increase it in others (anthropomorphism that triggers uncanny-valley or accountability-laundering reactions).

3.7 Expertise and self-comparison

Logg, Minson, and Moore find that algorithm appreciation weakens or reverses when the comparison is between the algorithm and the user's own judgment rather than between the algorithm and a generic human advisor, and when the user has domain expertise. Domain experts in particular tend to under-rely on algorithmic advice in their domains — including, in some studies, when the algorithm outperforms them. This is consistent with the "egocentric advice discounting" literature: people weight their own judgment more than equivalent advice from any source.

Combined with the Dietvorst result, the implication is that expert users of decision support tools are doubly exposed to aversion: they discount external advice generally, and they over-penalize algorithmic errors specifically. Deployments that target expert users (clinicians, loan officers, hiring managers, judges) cannot rely on either appreciation effects or naive performance disclosure to produce calibrated reliance.

4. The counter-pattern: algorithm appreciation

Logg, Minson, and Moore's 2019 Organizational Behavior and Human Decision Processes paper "Algorithm Appreciation: People Prefer Algorithmic to Human Judgment" (Logg, Minson, and Moore 2019) is the most important counterweight to a one-directional reading of the aversion literature. Across six experiments using a judge-advisor paradigm, lay participants weighted advice attributed to an algorithm more heavily than identical advice attributed to a person. The effect held for numeric estimates (weight from photos, song popularity rankings, romantic attractiveness ratings) and for a forecasting task.

Two findings inside this paper are particularly important:

Appreciation is conditional on the comparison frame. When the comparison is algorithm vs. another person, lay participants prefer the algorithm. When the comparison is algorithm vs. self, the preference shrinks or reverses.
Appreciation weakens with expertise. Professional forecasters (national security experts in one experiment) did not show appreciation; they showed something closer to indifference or mild aversion.

Algorithm aversion and algorithm appreciation are therefore not opposites on a single dial. They are reliance patterns triggered by different combinations of comparison target, task type, error visibility, user expertise, and source label. The honest summary of the field is: untested algorithms tend to receive a modest reliance premium from lay users on objective tasks; observed algorithmic errors produce asymmetric forgiveness penalties that often exceed the premium; experts show muted appreciation and amplified post-error aversion; moral and uniqueness-laden tasks elicit aversion even before any error is observed.

This means the central empirical question for any deployment is not "will users trust this?" but "given this task, this user, and this error profile, what reliance pattern should we expect, and is it calibrated?"

5. The LLM frontier (emerging evidence)

The predictive-AI literature was built on tasks where the model produced a number or a category, the user produced a competing number or category, and accuracy could be scored against ground truth. Generative AI breaks every part of that pipeline. Outputs are open-ended; the user's "competing output" is often a request rather than a forecast; accuracy is replaced by harder-to-score concepts like usefulness, faithfulness, or appropriateness; and the system has conversational presence that recruits social rather than statistical processing.

Porting the algorithm-aversion frame onto generative AI is therefore a genuine extrapolation, and the 2025+ evidence should be read as emerging rather than settled.

The strongest current studies fall into three groups.

Source-label effects. Osborne and Bailey 2025 (Osborne and Bailey 2025) find that personal advice generated by large language models is rated as high or higher than human-generated advice in blind comparisons, but lower when the AI source is disclosed. Several adjacent 2025 papers replicate label penalties in domains ranging from medical communication to creative writing evaluation. The structural pattern is consistent with predictive-AI label effects, but the magnitude and persistence vary widely across domains and stakes.

Domain-dependent reversals. Jin, Yalcin, and Mookerjee's 2025 Journal of Public Policy & Marketing paper on the "human superiority effect" in advice taking with generative AI (Jin et al. 2025) finds that users prefer human to AI advice on identity-laden and morally salient questions even when AI advice is rated as equally informative. By contrast, Merkle's 2025 SSRN working paper on financial-advice AI (Merkle 2025) finds appreciation patterns — participants who experience ChatGPT-based financial advice update toward favoring AI advisors over the human professionals they preferred ex ante. Both findings are credible at the level of the studies they report; together they are best read as evidence that domain and outcome observability dominate generic source effects.

Generative-vs.-predictive comparisons. Recent conference work, including 2025 ICIS papers and several arXiv preprints, has begun to compare aversion patterns when the same underlying recommendation is delivered as a number-from-a-model or as a sentence-from-a-chatbot. Early results suggest that generative framing can attenuate or reverse some predictive-AI aversion effects — for example, conversational explanation may produce higher uptake than equivalent numeric disclosure — but these studies vary in task realism, sample size, and whether the model is actually superior to the human alternative. They should not yet be cited as settling whether generative AI escapes or inherits the canonical effect.

Three features of the LLM setting are genuinely new and matter for any deployment that wants to reason about reliance:

Hallucination is a fundamentally different error class than forecast error. A forecast is wrong by a magnitude; a hallucination is wrong by category and is often delivered in the same fluent register as a correct answer. The "asymmetric forgiveness" mechanism in Dietvorst et al. may be amplified for hallucinations because users cannot easily price the error: a single fabricated citation may cost an LLM more share-of-reliance than a much larger numerical error costs a forecasting model.
Conversational presence recruits social processing. Generative systems produce reasons, hedges, apologies, and apparent perspective-taking. This blurs the boundary between advice and authorship, and it activates trust dynamics from interpersonal contexts (warmth, competence, perceived intentionality) that the predictive-AI literature largely sidestepped.
Open-endedness defeats local accuracy disclosure. Performance disclosures are well-defined for forecasting models (MAPE, AUC, calibration plots). They are ill-defined for general-purpose chat. Domain-specific benchmarks (e.g., medical Q&A accuracy, code-task pass rates) can be reported, but they often do not predict performance on the user's local task.

The honest summary is: there is suggestive evidence that LLM contexts shift the aversion-appreciation balance, that source labels still carry effects, and that the mechanisms differ enough from predictive-AI tasks that the older interventions (transparency, control, performance disclosure) do not transfer one-to-one. Anyone citing 2025 LLM work as a settled extension of Dietvorst et al. is overclaiming.

6. Open empirical questions

Four questions are genuinely unsettled by the current evidence base, and any wiki treatment that pretends otherwise should be discounted.

Does aversion attenuate as model capability becomes overwhelming evidence? The classic finding is that aversion persists after participants see comparable algorithm and human errors — but in those experiments, the algorithm's advantage is modest. It is plausible (and assumed by many deployment teams) that sufficiently large, sufficiently visible capability gaps would dissolve aversion. The supporting evidence is thin. Performance feedback does move reliance, but no study I am aware of has demonstrated that arbitrarily large capability gaps eliminate post-error asymmetric forgiveness in tasks where stakes are high and observation periods are long. The most cited natural-experiment evidence — uptake of GPS navigation, of statistical models in baseball, of LLMs in coding — confounds capability with infrastructure, social proof, and institutional adoption pressure. Treat "capability will solve aversion" as a hypothesis, not a finding.

Does the pattern hold across cultures? Most foundational studies use US online or undergraduate samples; nearly all are WEIRD (Henrich, Heine, and Norenzayan 2010). Some cross-cultural work exists — for example, studies of Chinese-sample reliance on AI advisors — but there is not enough methodologically comparable evidence to claim that asymmetric forgiveness, uniqueness neglect, or appreciation effects generalize across cultures. Hypotheses with face validity: cultures with higher institutional trust in technology may show muted aversion; cultures with stronger relational accountability norms may show stronger aversion on social-judgment tasks. Neither has been adequately tested.

Does generative AI elicit different aversion patterns than predictive AI? §5 lays out the suggestive evidence. The cleanest experiment would hold task and recommendation constant and vary only the framing (numeric output vs. conversational output) and source label (human/AI/anonymous), with both blind quality ratings and behavioral reliance measures, and with observed-error vs. unobserved-error arms. This experiment has been run in fragments; it has not been run end-to-end at adequate scale.

Are LLM-era aversion findings tracking aversion or rational caution? Hallucinations are categorically different errors, and there is principled reason to expect users to distrust open-ended outputs more than calibrated forecasts. Distinguishing "biased over-penalization of LLM errors" from "appropriate caution given the actual error profile" requires studies that establish the ground-truth quality of LLM outputs in the deployment domain, then measure user reliance against that ground truth. Studies that compare blind ratings to disclosed ratings get partway there but do not resolve the question. This is the highest-value open question for deployment teams, because the policy responses are opposite: if aversion, the design fix is calibration; if rational caution, the design fix is reducing the actual error profile (and possibly slowing deployment).

7. The limits of "reduce aversion" as a deployment frame

The dominant practical takeaway in popular treatments of algorithm aversion is some version of: give users control, transparency, and personalization; aversion will fall; adoption will rise; deployment will succeed. This is true in narrow forecasting settings. It is unsafe as a general deployment frame, for three reasons.

7.1 Reducing aversion can produce overreliance

The same interface levers that convert algorithm aversion into use can convert it into automation bias. Bounded adjustment increases adoption (Dietvorst et al. 2018). It does not increase calibration unless override behavior tracks the model's actual competence boundary. The well-documented overreliance literature in clinical decision support, aviation, and intelligence analysis (Parasuraman and Manzey 2010) shows that high-trust interfaces produce predictable patterns of late or absent override even when the system is visibly out of distribution.

A deployment that measures only adoption cannot tell whether it has fixed aversion or manufactured overreliance. The minimum measurement set for distinguishing them is:

Adoption rate (uses model recommendation as-is).
Override rate, both in-distribution and out-of-distribution.
Override quality (edit direction, edit magnitude, edit effect on outcome).
Post-error reliance: does adoption drop after visible model failure? By how much? Does it recover?
Outcome accuracy of the human-model team, not the model alone and not the human alone.

A deployment that shows rising adoption and stable outcome accuracy after errors is plausibly calibrating. A deployment that shows rising adoption and unchanged outcome accuracy after errors is plausibly bias-laundering.

7.2 User control has accountability side effects

"Human in the loop" is the most common reassurance offered to regulators, ethicists, and end-users. It is also the most easily counterfeited. The slot for human review can be filled by a reviewer with no time, no incentive, no expertise, no ability to escalate, or no statutory authority to override; the system is "human-supervised" by an interface designer's diagram and not by anyone who could change a decision. The same Dietvorst-style logic that supports bounded adjustment as an adoption lever supports it as an accountability lever: even cosmetic adjustment rights make a user feel responsible for the outcome, which in regulated settings can substitute the user's accountability for the deploying organization's.

The deployment question is not "is there a human in the loop?" but "what authority does the human in the loop actually have, what information do they receive, what time do they have, and who is accountable for decisions they routinely approve?" Designs that answer these questions in writing — with audit logs, override statistics, escalation paths, and clear allocation of responsibility for systemic vs. case-level error — produce a different kind of human-in-the-loop than designs that answer them with an interface element.

7.3 Some aversion is not bias

In moral, high-stakes, and rights-affecting domains, refusal to delegate to an algorithm can be a legitimacy claim rather than an accuracy mistake. Bigman and Gray's finding that aversion to machine moral decision-making survives accuracy parity is consistent with this reading. A growing philosophical literature on algorithmic decision-making — Brand's 2025 treatment (Brand 2025) is one recent example — argues that some user objections track reciprocal-obligation, contestability, and reason-giving concerns rather than competence concerns. A model can produce statistically superior recommendations and still be the wrong authority if affected people cannot demand reasons, contest decisions, or hold a responsible party accountable.

This does not vindicate every refusal as principled. Many objections to algorithmic decision-making are post-hoc rationalizations of professional discretion that was never evidence-based — the Meehl–Dawes–Grove tradition documents decades of clinical judgment outperformed by simple actuarial models that practitioners refused to adopt. The honest deployment frame distinguishes:

Objections that would be satisfied by accuracy disclosure and bounded control. These are the targets of standard aversion-reduction design.
Objections that would not be satisfied by any accuracy disclosure, because they concern who is permitted to decide. These are governance questions, not UX questions.
Objections that are professional-discretion rationalizations. These are a fourth thing — legitimate to challenge with evidence, but not to flatter with override sliders.

A deployment frame that treats all three as "aversion to overcome" will mishandle the second category and entrench the third.

8. Design implications: calibrated reliance, not maximum trust

The least-marketing-friendly summary of the literature is that the deployment target should be calibrated reliance, not trust. Trust is an attitude; reliance is a behavior; calibration is the alignment between reliance and competence. Maximum trust is a failure mode, not a success metric, whenever the model has a competence boundary — and every model has one.

A non-exhaustive list of design moves that support calibrated reliance and are reasonably well-supported by the evidence reviewed here:

Show comparative performance, not absolute performance. Headline accuracy numbers ("our model is 94% accurate") do not give users the information needed to allocate reliance. Comparative performance — how the model's error rate, error magnitudes, and error types compare to the realistic human alternative on the user's task — is what Dietvorst-style results actually update. This includes showing where humans outperform the model.

Make errors locally inspectable. Aggregate accuracy claims paired with example errors (the model's misses, including the kinds of cases it tends to get wrong) produce better calibration than aggregate accuracy alone. This is asymmetric in a useful direction: showing errors raises the credibility of the accuracy claim while reducing the magnitude of post-error abandonment, because users have been primed to expect that errors exist.

Permit bounded, logged adjustment. The Dietvorst 2018 result is robust enough to justify designs that allow users to modify model outputs within an outcome-tested envelope. The design contract is: adjustments are logged, override outcomes are measured, the envelope is widened or narrowed based on whether user overrides improve or degrade accuracy in that envelope. Adjustment without measurement is an adoption lever masquerading as a quality lever.

Distinguish advice from automation. A model that recommends and a model that acts have different reliance dynamics. Recommender-style deployments inherit the algorithm-aversion / algorithm-appreciation literature relatively cleanly. Automation deployments inherit the automation-bias literature, including its over-reliance failure modes. Hybrid designs (model acts unless user objects within N seconds) tend to inherit the worst of both. Make the distinction visible to the user and to the design team.

Measure behavior after visible error. The single most diagnostic measurement for whether a deployment is calibrating reliance or manufacturing trust is what happens to override behavior after the model is observed to fail. Healthy calibration shows local override increases in cases similar to the failure, no broad collapse of usage, and a return to baseline as additional cases accumulate evidence. Unhealthy patterns include: no override change after a clearly visible failure (overreliance), broad usage collapse after a single failure (Dietvorst-style aversion), and override change that does not track the actual locus of model failure (e.g., users abandon the model in cases unlike the failure case).

Preserve audit, appeal, and contestability where stakes warrant. This is the governance complement to the UX prescriptions above. For decisions affecting rights, employment, credit, health, or liberty, the legitimacy-objection literature (§7.3) implies that deployment cannot rely on UX-level aversion reduction. Audit logs, appeal channels with real authority to overturn, and explicit accountability for the deploying organization are not nice-to-haves; they are the prerequisites for treating user refusal as bias rather than as a substantive objection.

Tier evidence for generative AI separately. Until LLM-specific evidence is mature, designers should not assume that interventions validated in forecasting tasks transfer to chat-based tools. In particular: bounded adjustment, transparency, and source-disclosure prescriptions all have weaker and more mixed evidence in generative contexts than in predictive ones. Where the deployment uses generative AI, the prudent stance is to run local A/B tests on reliance behavior rather than to import design conclusions from predictive-AI studies.

9. What a careful reader should take away

Algorithm aversion is a real, replicable, behaviorally specific finding. It is also a term that has expanded beyond its evidentiary base in popular usage, where it now functions as a label for any user resistance to AI. The honest summary of the literature is more conditional than the label suggests:

Under conditions of measurable accuracy, observed error, and discrete choice between algorithmic and human advice, people under-rely on a superior algorithm — the canonical Dietvorst pattern. This is robust in forecasting tasks.
The same setting reverses into appreciation when users have not yet observed errors and when the comparison is algorithm-vs.-other-human rather than algorithm-vs.-self.
Aversion strengthens with task subjectivity, perceived uniqueness, expertise of the user, moral salience, and source labels — all of which describe the deployments that matter most institutionally.
Bounded user adjustment is a well-supported lever for increasing adoption of forecasting models. It is a weaker lever for improving decision quality and a poor substitute for accountability.
Generative-AI evidence is emerging and should be cited as such. The mechanisms differ enough from predictive AI that the older interventions do not transfer one-to-one.
The right deployment target is calibrated reliance, not maximum trust; the right diagnostic measure is behavior after visible error, not stated trust before error.
In high-stakes domains, some user refusal is a legitimacy objection rather than a calibration error, and is not addressable by UX.

The strongest practical instruction this literature supports is also the least flashy: measure reliance behavior, not trust attitudes; measure it after visible errors, not only at first use; measure the accuracy of the human-model team, not the model alone; and design under the assumption that the goal is appropriate use of a system with a competence boundary, not maximum use of a system without one.

Companion entries

Core theory:

Algorithm Appreciation
Automation Bias
Trust Calibration
Calibrated Reliance
Clinical vs. Statistical Prediction (Meehl Tradition)
Asymmetric Forgiveness of Errors
Uniqueness Neglect in AI Acceptance

Practice:

Human-in-the-Loop AI
Decision Support System Design
Explainable AI
Model Cards and Performance Disclosure
Uncertainty Communication in AI Interfaces
Override Logging and Outcome Measurement
AI Advice-Taking and Source Labels
Bounded Adjustment as an Adoption Lever

Counterarguments and governance:

Legitimacy Objections to Algorithmic Decision-Making
Contestability of AI Decisions
Accountability Laundering via Human Oversight
Professional Discretion vs. Statistical Prediction
Cross-Cultural Generalizability of WEIRD Findings

LLM frontier:

Hallucination as an Error Class
Source-Label Penalties in Generative AI
Generative vs. Predictive AI Reliance Patterns
Conversational Agency and Anthropomorphism