Algorithm Aversion

At a Glance

1. Quick Summary

Algorithm aversion is what happens when a forecasting model misses once and we stop using it, even though the human forecaster we switch to misses more often. Berkeley Dietvorst, Joseph Simmons and Cade Massey named the pattern in 2015 after watching participants abandon a statistical model that beat them, simply because they had seen it err. Four years later Jennifer Logg, Julia Minson and Don Moore showed the opposite: when people had not seen an algorithm fail and the task was a matter of fact rather than taste, they followed algorithmic advice more readily than human advice. Visible errors, subjective tasks and personal expertise push toward aversion; objective tasks and a clean record push toward appreciation. Since statistical rules have matched or beaten expert judgment in most head-to-head tests since the 1950s, the aversion side costs us accuracy and the appreciation side costs us scrutiny.

Simply About the Bias

Algorithm aversion means giving up on a computer's advice the first time it gets something wrong, even when the computer is still right more often than a person would be. The flip side is trusting a computer too much when you have never seen it mess up. Both come from the same habit: we judge machines by a stricter rule than we judge people.

Imagine your phone's map app sends you down a road that turns out to be closed. You are late and annoyed, and you decide the app is useless. The next day you ask a friend for directions instead. Your friend gets you lost twice, but that feels fine, because people make mistakes. The app made one mistake and you wrote it off forever.

Here is a trick for catching it. When something lets you down, ask a plain question: "Over the last ten times, who got it right more often?" Count honestly. If the app wins eight out of ten and your friend wins five, the app is still the better choice. And if you have never seen the app fail, that deserves a second look too, because nothing is right every time.

2. The Science Behind It

2.1. Discovery and History

The story starts long before anyone used the word "algorithm" in a psychology paper. In 1954 the clinical psychologist Paul Meehl published a short, uncomfortable book called Clinical versus Statistical Prediction. He gathered the roughly twenty studies then available that pitted trained clinicians against simple statistical formulas, and in nearly every comparison the formula matched or beat the clinician. Practice barely changed; clinicians found reasons to distrust the tables, and that resistance was the first large-scale demonstration of this bias, decades before it had a name.

The evidence kept piling up. In 2000 William Grove and colleagues published a meta-analysis of 136 studies comparing clinical and mechanical prediction. Mechanical prediction was about 10 percent more accurate on average, won outright in roughly half the comparisons, tied in most of the rest, and lost clearly in only a handful.

The laboratory demonstration came in 2015. Dietvorst, Simmons and Massey, then at the University of Pennsylvania, ran incentivized forecasting experiments and coined "algorithm aversion." Participants who could choose between their own forecasts and a statistical model's were far less willing to pick the model once they had watched it make mistakes, even though they had also watched it beat them.

The field then complicated itself, which is what healthy fields do. Logg, Minson and Moore's 2019 "algorithm appreciation" paper showed that laypeople updated more toward advice labeled algorithmic than advice labeled human, though the effect shrank among experts. Systematic reviews by Jussupow, Benbasat and Heinzl (2020), Burton, Stein and Jensen (2020) and Mahmud and colleagues (2022) reached one verdict: aversion is real but conditional. A 2024 working paper by Logg and Schlund argued that most "aversion" in the literature comes from hypothetical preference questions, while people facing actual decisions use algorithmic advice more than human advice. A 2025 meta-analysis presented at the Hawaii International Conference on System Sciences pooled 32 intervention studies and found a small average reduction in aversion, with the largest gains from changing how people interact with the algorithm. At least one preregistered attempt to reproduce the basic effect has failed.

Key milestones:

• 1954: Meehl's Clinical versus Statistical Prediction shows formulas matching or beating clinicians, and clinicians declining to use them
• 2000: Grove and colleagues' meta-analysis of 136 studies finds mechanical prediction about 10 percent more accurate on average
• 2015: Dietvorst, Simmons and Massey coin "algorithm aversion"
• 2018: The same team shows that letting people adjust a model's output even slightly restores its use
• 2019: Logg, Minson and Moore publish "algorithm appreciation"; Castelo, Bos and Lehmann show task subjectivity governs trust; Longoni, Bonezzi and Morewedge identify uniqueness neglect
• 2024-2025: Logg and Schlund's reconciliation paper and a HICSS meta-analysis of 32 intervention studies frame aversion and appreciation as two settings of one trust dial

2.2. Key Researchers

2.3. Landmark Studies

"Algorithm Aversion: People Erroneously Avoid Algorithms After Seeing Them Err" (Dietvorst, Simmons & Massey, 2015)

Participants forecast either how MBA applicants would perform from their admissions data, or how many airline passengers would fly from each US state. A statistical model built on the same data was available, and some participants watched it perform before choosing whose forecasts would set their bonus. The model outperformed the humans in every condition. Yet participants who had watched the model perform, and therefore watched it err, were substantially less likely to bet on it than those who had seen nothing.

"Overcoming Algorithm Aversion" (Dietvorst, Simmons & Massey, 2018)

The same team asked what would bring people back. Participants could use the model's forecast but adjust it, in some conditions freely and in others by only a few points. Being allowed to adjust, even within a narrow band, sharply raised the share who chose the model, and they ended up more accurate than those who forecast alone. What people want is a say, not total control.

"Algorithm Appreciation: People Prefer Algorithmic to Human Judgment" (Logg, Minson & Moore, 2019)

Across six experiments, participants estimated things like a stranger's weight from a photo or a song's chart position, then received advice labeled as coming from an algorithm or from a person. Laypeople moved closer to the algorithm's advice. The effect weakened when participants chose between the algorithm and their own judgment, and reversed among experienced professionals.

2.4. Neurological Basis

The neuroscience of algorithm aversion specifically is thin. The named studies in this literature are behavioral; nobody has yet scanned participants while they watch a forecasting model fail. What exists is a set of established findings about how the brain treats human and non-human partners.

Neuroimaging work on economic games shows that regions linked to mentalizing, especially the medial prefrontal cortex, respond more strongly when people believe they are interacting with another human than with a computer running a fixed rule; McCabe and colleagues reported this in 2001 and later studies replicated the pattern. We build a mental model of a human advisor's intentions, while a machine gets filed as a mechanism. A mechanism that errs looks defective; a person who errs looks like a person. Any claim of a dedicated "algorithm aversion circuit" should be treated as speculation.

3. Evolutionary Origins

Our ancestors never met an algorithm, but they met two things that matter here: tools and advisors. The brain's rules for each were tuned to different problems.

Advice from another person came bundled with a face, a history and a reputation. Evaluating it meant modeling the advisor's knowledge and incentives, and that machinery is forgiving of occasional error, because a reliable ally who sometimes misjudges is still worth keeping. A tool was judged by whether it worked. A snare that failed to spring was repaired or discarded; tools do not learn or apologize, so there was no reason to extend one the benefit of the doubt. When a modern person watches a statistical model miss, the tool rule fires: discard.

A second old pattern is the pull toward the explainable. A human advisor can be asked "why?" and will produce a story. A formula that says "73 percent" offers no narrative. Yeomans, Shah, Mullainathan and Kleinberg showed in 2019 that people preferred joke recommendations from a human to equally good ones from an algorithm, and that supplying an explanation closed much of the gap. The habits are not broken; they are being applied to a new kind of object that is a tool in its mechanics and an advisor in its function.

4. How This Bias Manifests

4.1. In Everyday Life

• Deleting a navigation app after one bad route, then driving longer routes for months on memory and habit
• Calling a weather forecast "always wrong" after a rained-out picnic, while forgetting the dozens of days it was right
• Following a recipe app's cooking time to the second on first use, and never checking the oven

4.2. In the Workplace

• Sales teams overriding a lead-scoring model after it ranks one obviously good prospect low, then returning to gut-feel prioritization that converts worse
• Senior professionals, as in Logg's national security sample, discounting advice that junior colleagues would have accepted

4.3. In Business and Marketing

Castelo's work explains a pattern marketers see constantly: customers accept an algorithm for the "objective" parts of a product and resist it for the "subjective" ones. A bank can advertise that a model sets your credit limit; a dating app that says "our model picked your match" risks sounding cold. Companies respond by hiding the algorithm (a "curated" playlist), by humanizing it with a name, or by adding a token human step.

4.4. In Politics and Media

Public algorithms attract public aversion. When a government uses a model to allocate school places or moderate exam grades, every visible error becomes a news story, while the errors of the human process it replaced were never counted. Britain's 2020 grading episode (Case Study 2) is the textbook case. But the same public that abandons a grading model trusts a recommendation feed it has never audited.

4.5. In Healthcare

Clinicians have resisted statistical prediction since Meehl's day, and the Grove meta-analysis found the resistance survived decades of contrary evidence. Longoni and colleagues showed that patients share it: offered identical accuracy, they prefer a human diagnosis, because they believe their case is unusual. The preference reverses when a doctor "reviews" the AI's output, even nominally. On the other side of the swing sits automation bias, covered separately on this site.

4.6. In Finance and Investing

An investor who follows a rules-based rebalancing strategy through two good years will often drop it after one quarter in which the rule sold too early. Dietvorst and Bharti's 2020 finding, that people reject algorithms most in inherently uncertain domains, lands hardest here.

5. Real-World Case Studies

Case Study 1: Zillow Offers and the Price of Trusting a Model Blindly

• Context: Zillow, the US property listings company, launched Zillow Offers in 2018, using its home-valuation models to make instant cash offers on houses and resell them.
• What happened: Through 2021 the model's offers ran ahead of resale prices as the pandemic housing market shifted. By autumn the company held roughly 7,000 houses. In November 2021 Zillow announced it would shut the business down, reported a loss of around 420 million dollars for the division in the third quarter alone, and said it would cut about a quarter of its workforce.
• The bias at work: This is the appreciation end of the swing at corporate scale. The model had not visibly failed before and the firm's identity was built on trusting its valuations, so human checks were thin. Once the failure became visible, the response was total abandonment rather than recalibration.
• Consequences: Hundreds of millions in losses, a 25 percent workforce reduction, and the end of a flagship business line.
• Lessons learned: A model that has never failed in front of you is not a model that cannot fail. Continuous monitoring with pre-set tripwires turns a visible error into an inspection rather than denial or panic.

Case Study 2: The 2020 English Exam Grading Algorithm

• Context: When the pandemic cancelled A-level examinations in England in 2020, the regulator Ofqual built a standardization model to moderate teacher-predicted grades and prevent grade inflation. Results were published on 13 August 2020.
• What happened: The model lowered nearly 36 percent of A-level grades by one step below the teacher's prediction and about 3 percent by two steps. Because it leaned on each school's historical results, strong students at historically weaker schools were marked down most. Protests followed within hours. On 17 August, four days after results, the government and Ofqual reversed course and reissued grades based on unmoderated teacher predictions.
• The bias at work: Part of the reaction was a fairness judgment, and a reasonable one: the errors fell on identifiable groups. But the episode also shows the aversion mechanism in pure form. The algorithm's errors were visible and personal; the errors of the human alternative (teacher predictions, long known to run optimistic) were invisible and diffuse. The country discarded the visible error source entirely rather than fixing it.
• Consequences: Grades rose sharply against prior years, universities scrambled to place students, and the head of Ofqual resigned later that month.
• Lessons learned: Visible errors are not the same as more errors. A model judged against human judgment deserves a fair accounting of the human errors too, and its failure cases should be tested on the groups most likely to be hurt before it ships.

Historical Example: Oakland's Scouts and the Spreadsheet (2002)

In the early 2000s the Oakland Athletics, a baseball team with one of the smallest payrolls in the league, began using statistical analysis to find players that scouting undervalued. General manager Billy Beane and his assistant Paul DePodesta leaned on measures like on-base percentage, which predicted runs better than the judgments scouts had used for a century.

The scouts' reaction was the Meehl story replayed on a ballfield. They pointed to every player the model liked who flopped as proof that "you can't reduce baseball to a spreadsheet," and kept no comparable tally of their own misses. In 2002 Oakland won 103 games, including a record twenty in a row, on a budget a third the size of the New York Yankees', and Michael Lewis's 2003 book Moneyball made the episode famous. Once the approach had visibly worked, nearly every team adopted it, and some swung to the far end, trusting models where the data were thin.

6. The Cost of This Bias

6.1. Personal Costs

• Worse decisions in domains where a good tool exists: longer commutes, poorer financial allocations, missed health signals
• The mirror cost: unexamined reliance on a recommendation, credit score or diagnostic app that was never checked, until it fails in a way that matters

6.2. Professional Costs

• Forecasting accuracy left on the table: in Dietvorst's experiments, participants who abandoned the model earned smaller bonuses than those who kept it
• Expensive "return to manual" decisions after a single visible error, which quietly reinstate a less accurate process
• The opposite failure in automated workflows: nobody owns the output because "the system handles it," so drift goes unnoticed for quarters

6.3. Societal Costs

• Public services that abandon a fixable model and revert to a less fair or less accurate human process
• Under-use of validated risk tools in medicine, criminal justice and child protection, where the evidence for statistical methods is decades old
• A public conversation in which algorithmic errors are counted and human errors are not

6.4. Statistical Impact

• Grove and colleagues' 2000 meta-analysis of 136 studies found mechanical prediction about 10 percent more accurate than clinical judgment on average, with clinicians clearly better in only a handful of comparisons
• In Dietvorst, Simmons and Massey's 2015 experiments, seeing the model perform (and err) substantially reduced the share of participants who chose it, although it outperformed them in every condition
• Letting participants adjust a model's forecast within a narrow range (Dietvorst et al., 2018) raised model use markedly and improved accuracy over unaided forecasting
• A 2025 meta-analysis of 32 intervention studies (89 effect sizes) found a small average reduction in aversion, with interventions that changed the interaction environment roughly twice as effective as the average

7. The Hidden Benefits

Algorithm aversion is not simply a mistake with a good press agent. Several parts of it are worth keeping.

A model that has erred once may have erred because the world changed, and a model cannot tell you that. The forecaster who reacts strongly to a visible miss is sometimes detecting the kind of regime shift a backtested model is blind to. Zillow's models did not know the housing market had turned; a skeptical human could have.

The demand for a human in the loop, which Longoni's patients expressed, is also a demand for accountability. A doctor can be asked why. A model in most deployments cannot, and a decision that affects a life ought to come with someone who can answer for it.

And the aversion side of the swing is the only brake on the appreciation side. The problem is not that we scrutinize algorithms. It is that we do it at the wrong moment: after one error rather than before deployment and throughout use.

8. Self-Assessment: Do You Have This Bias?

8.1. Warning Signs Checklist

• You have stopped using a tool after one bad recommendation without checking its overall record
• When a model is wrong you say "I knew it," but when a person is wrong you say "it happens"
• You trust a system completely until you see it fail, then distrust it completely
• You have overridden a model's output by a large amount on a hunch, without recording the result
• You believe your situation is too unusual for any statistical tool to apply
• You accept algorithmic advice for "number" questions but reject it for "people" questions
• You have never asked how often a tool you rely on daily is wrong
• You feel insulted when an algorithm's answer is better than yours
• You prefer an explanation you can follow over a result with the better record
• You have said "you can't reduce this to a formula" about a decision where formulas have beaten experts

Scoring:

• 0-2 checked: Low susceptibility
• 3-5 checked: Moderate susceptibility
• 6-8 checked: High susceptibility
• 9-10 checked: Very high susceptibility

8.2. Self-Reflection Questions

1. Name a tool or model you abandoned. Do you know its actual hit rate compared with whatever replaced it?
2. Name a tool you rely on without question. When did you last check whether it had made an error you did not notice?
3. Which decisions in your work are "objective" and which "subjective"? Have you ever tested that classification?
4. When you override a system's output, do you write it down and revisit it later?
5. If a colleague and a model each missed the same forecast by the same amount, would you treat them the same way?

8.3. Quick Diagnostic Scenario

Scenario: Your team's demand-forecasting model has beaten the planners' manual forecasts in eleven of the last twelve months. This month it badly overestimated demand for one product line. Your manager asks what to do.

How would you respond?

• A) Recommend dropping the model and returning to planner forecasts, since it clearly cannot be trusted → High susceptibility
• B) Keep the model but let planners override it freely whenever they disagree → Moderate susceptibility
• C) Keep the model, investigate what changed in that product line, and allow planners bounded adjustments with a record of each one → Low susceptibility

9. Identifying This Bias in Others

9.1. Behavioral Indicators

• Treats a single model error as proof the model is worthless, with no comparison to the human error rate
• Treats a long run of model successes as proof the model needs no checking
• Makes large, undocumented manual overrides of a system's outputs
• Insists their case, customer or patient is "different" in ways no tool could capture
• Is an acknowledged expert in the domain and discounts the tool more than novices do

9.2. Conversational Red Flags

Phrases people say when under this bias:

• "See? The model got it wrong. I told you."
• "You can't put a number on this."
• "The system has never been wrong, why would we check?"
• "It doesn't know my situation."
• "Just let me adjust it, I've been doing this for twenty years."

Types of arguments they make:

• One memorable algorithmic failure against a vague sense that humans "usually get it right"
• Appeals to the specialness of this case rather than evidence about cases like it

Questions they avoid asking:

• "How often is the model right compared with us?"
• "When did we last audit this?"

9.3. Situational Triggers

• A visible, attributable algorithmic error, especially a public one
• Tasks that feel subjective, personal or moral
• A hypothetical "which would you trust?" question rather than a real decision with stakes
• Expertise in the domain, which raises confidence in one's own judgment
• No visible failure yet, combined with pressure to move fast, which triggers the appreciation end

10. Cognitive Debiasing Strategies

10.1. Immediate Techniques

• The two-column tally: When a tool errs, write its last ten results in one column and the human alternative's in another. Compare before deciding anything.
• The same-error test: Ask, "If a colleague had made exactly this mistake, would I fire them?" If not, do not fire the model.
• Bounded adjustment: Instead of accepting or rejecting a model's output wholesale, adjust it within a pre-set range and log the change. Dietvorst's 2018 participants did well with a few percentage points.
• Flip the question: If you have never seen a system fail, ask what its failure would look like and whether you would notice.

10.2. Long-Term Strategies

• Keep an override journal: Record every manual override of a model and revisit it quarterly. Most people who do this find their overrides hurt more than they helped, which recalibrates trust faster than any lecture.
• Learn the base rates: Read the Grove meta-analysis or Kahneman's chapter on intuition versus formulas once. It changes how a single error feels.
• Audit the unfailing: Schedule reviews of the systems that have never visibly erred. They are the ones most likely to be drifting unseen.

10.3. Environmental Design

• Build in bounded modification: Give users a small, logged adjustment range rather than an accept/reject switch. The 2018 study found this raised use and accuracy at the same time.
• Display comparative error rates: Show the model's hit rate next to the human baseline on the same screen, so a single miss is seen against its denominator.
• Design tripwires, not kill switches: Pre-commit that a visible error triggers an investigation with a named owner and a deadline, and that only a documented finding can retire the model.

10.4. When to Seek External Input

• When you are about to retire a model after a visible error and have not compared error rates
• When you are about to deploy a model in a public or high-stakes setting and have not tested its failure cases on the groups most likely to be affected
• When a system has run for a year without anyone being able to say how often it is wrong

Whom to ask: someone who can run the comparison, not someone who shares your frustration.

11. Practical Exercises

Exercise 1: The Fair Trial

• Objective: Replace a one-error verdict with an evidence-based comparison
• Time required: 45 minutes
• Materials needed: Records of a tool's past outputs and the human alternative's past judgments; a spreadsheet
• Difficulty level: Beginner
• Instructions:
  1. Pick a tool you have abandoned or distrust
  2. Collect its last 20 outputs and what actually happened
  3. Collect the human alternative's last 20 judgments on comparable cases
  4. Score both on the same metric
  5. Write one sentence on which source you should use going forward and why
• Reflection questions:
  • Did the comparison match your gut feeling before you started?
  • Which single error had been doing the most work in your opinion of the tool?
  • What would it take for you to trust the winner, whichever it was?
• Frequency: Once per tool you are tempted to drop

Exercise 2: The Bounded Override Log

• Objective: Learn whether your manual adjustments help or hurt
• Time required: 2 minutes per decision, 30 minutes monthly review
• Materials needed: A notebook or a shared sheet
• Difficulty level: Intermediate
• Instructions:
  1. For one month, every time you adjust a model's output, log the model's value, your value, and your reason in one line
  2. Set a maximum adjustment size in advance (for example 10 percent)
  3. At month end, add the actual outcomes and compute the error of the model alone versus the error after your adjustment
  4. Mark each override as "helped," "hurt" or "no difference"
  5. Decide whether to narrow or widen your range next month
• Reflection questions:
  • What share of your overrides helped?
  • Were the helpful ones the ones with a specific reason or the ones based on feel?
  • Did knowing you would be scored change how often you overrode?
• Frequency: Monthly, for at least three months

Exercise 3: Auditing the Unfailing

• Objective: Apply scrutiny to a system you trust by default
• Time required: 60 minutes
• Materials needed: Access to one system you rely on without checking
• Difficulty level: Intermediate
• Instructions:
  1. Write down what you believe the system's error rate is. Be specific.
  2. Write down what an error would look like and whether you would see it
  3. Sample 20 recent outputs and check each against ground truth or your best judgment
  4. Compare the measured error rate with your belief from step 1
  5. Draft a one-paragraph monitoring plan: what to check, how often, who owns it
• Reflection questions:
  • Was your belief about the error rate too high, too low or absent?
  • Which errors would have stayed invisible without this exercise?
  • What is the cheapest tripwire you could install?
• Frequency: Quarterly, rotating across systems

Daily Practice

The Denominator Question

Once a day, when any tool or system lets you down, ask out loud: "Out of how many?" Answer with a number before you react.

• Suggested duration: 30 seconds
• Best time of day: Whenever a tool errs
• How to track progress: Tick a box each time you asked the question before you complained

Weekly Challenge

Swap the Standard

Each week, pick one judgment you would normally make by feel (which email to answer first, how long a task will take, which route to drive) and follow a simple rule or tool instead. Score both at the end of the week.

• Expected outcomes after 4 weeks: A personal record of where rules beat you and where they do not
• Journaling prompts for reflection:
  • Which rule surprised you by winning?
  • Where did your judgment beat the rule, and what did it know that the rule did not?

12. For Specific Audiences

For Leaders and Managers

• Set the retirement rule in advance: Decide before deployment what evidence would justify dropping a model. A single visible error should trigger an inquiry, never a decision.
• Make error rates visible: Publish the model's accuracy alongside the human baseline. People cannot weigh one miss against a denominator they have never seen.
• Give bounded control: Staff who can adjust a model's output within limits use it more and perform better than staff given an accept/reject switch.

For Parents and Educators

• Age-appropriate explanation: "Computers and apps make mistakes, and so do people. The trick is to count how often each one gets it right, instead of remembering just the one time it went wrong."
• Teach the denominator early: When a child calls the app or the calculator "stupid," ask how many times it was right this week.
• Model both sides: Show that you check a map against a road sign sometimes, and that you do not throw the map away after one closed road.

For Healthcare Professionals

• Know the evidence: Statistical prediction has matched or beaten clinical judgment in most comparisons for seventy years. A risk score that misses once is not thereby discredited.
• Address uniqueness neglect directly: Patients resist algorithmic care because they feel unseen. Explain what the tool does with their specific data and stay visibly in the loop.
• Guard the other flank: Automation bias, where staff accept a wrong alert without checking, is the same trust dial turned the other way. Train for both.

For Financial Professionals

• Expect the drawdown reaction: Clients who embraced a rules-based strategy will want to abandon it after the first bad quarter. Set the expectation of periodic underperformance before it happens.
• Show the full record: Present a strategy's history as a distribution, not its latest result, and compare it with what the client would have done unaided.
• Do not confuse unfailing with safe: A model that has not failed in a client's memory has not been tested. Explain its failure modes before the market does.

13. Interactions with Other Biases

Biases That Amplify This One

Biases That Counteract This One

Common Bias Chains

Algorithm Aversion → Outcome Bias → Hindsight Bias → Not Invented Here

A model makes a visible error. The team judges it by that result rather than its process (Outcome Bias), soon recalls having "always known" it was unreliable (Hindsight Bias), and replaces it with an in-house method trusted on origin rather than evidence (Not Invented Here), whose own errors go uncounted.

The reverse chain runs Automation Bias → Illusion of Validity → Algorithm Aversion: a system is trusted without checking, and when silent drift finally produces a visible failure the swing to aversion is total.

Interrupting the chain: The intervention point is the same in both directions: a running, visible comparison of the model's error rate and the human alternative's.

14. Cultural Perspectives

Most of the algorithm aversion literature comes from American and Western European samples, and the few cross-cultural comparisons should be read as early findings rather than settled fact.

The first pattern concerns individualism. Longoni's uniqueness neglect, the feeling that "my case is special," is a more natural thought in cultures that emphasize the individual self. Cross-country work summarized in the topic's reviews reports less aversion in more collectivist samples, which fits that idea but is not yet strongly replicated.

The second concerns the authority attached to technology. Where the state deploys algorithms widely, people show more acceptance of algorithmic decisions in public life, though trust in the algorithm is hard to separate from deference to the institution. Where there has been a recent public scandal (Britain's 2020 grading episode, the Dutch childcare benefits affair), aversion to government models runs higher and is partly a reasonable response to experience.

15. Myths and Misconceptions

16. Expert Insights

"There is no controversy in social science which shows such a large body of qualitatively diverse studies coming out so uniformly in the same direction as this one." — Paul Meehl, 1986, on the clinical-versus-statistical prediction evidence

"To maximize predictive accuracy, final decisions should be left to formulas, especially in low-validity environments." — Daniel Kahneman, Thinking, Fast and Slow, 2011

People lose confidence in an algorithm faster than in a human after watching both make the same mistake, and so they choose the human even when the algorithm is more accurate. — Berkeley Dietvorst, Joseph Simmons and Cade Massey, 2015 (paraphrasing the paper's central finding)

17. Key Takeaways

1. It is a swing, not a setting: The same person who abandons a model after one error will over-trust a different model that has never visibly failed.

2. Visible error is the trigger: Dietvorst, Simmons and Massey showed that watching a model err does more damage to trust than watching a human make the same error.

3. The base rates favor formulas: From Meehl's twenty studies in 1954 to Grove's 136 in 2000, statistical prediction has matched or beaten expert judgment in most tested domains. One miss does not reverse that record.

4. Task framing and expertise matter: Objective tasks and clean records draw appreciation; subjective tasks, personal stakes and professional expertise draw aversion.

5. A little control goes a long way: Letting people adjust a model's output within a small, logged range restored use and improved accuracy in the 2018 experiments, and interaction changes beat explanations in the 2025 meta-analysis.

6. The effect is conditional and partly an artifact of how we ask: Reviews list many moderators, Logg and Schlund argue hypothetical questions inflate aversion, and a preregistered replication attempt has failed.

7. Scrutinize before the error, not after: Audit the systems that have never visibly failed, set retirement rules in advance, and keep an override log. Calibrated trust is the goal, in both directions.

18. Further Resources

Academic Papers

• Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114-126.
• Dietvorst, B. J., Simmons, J. P., & Massey, C. (2018). Overcoming algorithm aversion: People will use imperfect algorithms if they can (even slightly) modify them. Management Science, 64(3), 1155-1170.
• Logg, J. M., Minson, J. A., & Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes, 151, 90-103.
• Castelo, N., Bos, M. W., & Lehmann, D. R. (2019). Task-dependent algorithm aversion. Journal of Marketing Research, 56(5), 809-825.
• Longoni, C., Bonezzi, A., & Morewedge, C. K. (2019). Resistance to medical artificial intelligence. Journal of Consumer Research, 46(4), 629-650.
• Burton, J. W., Stein, M.-K., & Jensen, T. B. (2020). A systematic review of algorithm aversion in augmented decision making. Journal of Behavioral Decision Making, 33(2), 220-239.
• Jussupow, E., Benbasat, I., & Heinzl, A. (2020). Why are we averse towards algorithms? A comprehensive literature review on algorithm aversion. Proceedings of the 28th European Conference on Information Systems (ECIS 2020).
• Mahmud, H., Islam, A. K. M. N., Ahmed, S. I., & Smolander, K. (2022). What influences algorithmic decision-making? A systematic literature review on algorithm aversion. Technological Forecasting and Social Change, 175, 121390.
• Grove, W. M., Zald, D. H., Lebow, B. S., Snitz, B. E., & Nelson, C. (2000). Clinical versus mechanical prediction: A meta-analysis. Psychological Assessment, 12(1), 19-30.
• Dietvorst, B. J., & Bharti, S. (2020). People reject algorithms in uncertain decision domains because they have diminishing sensitivity to forecasting error. Psychological Science, 31(10), 1302-1314.

Books

• Meehl, P. E. (1954). Clinical versus Statistical Prediction: A Theoretical Analysis and a Review of the Evidence. University of Minnesota Press.
• Kahneman, D., Sibony, O., & Sunstein, C. R. (2021). Noise: A Flaw in Human Judgment. Little, Brown Spark.
• Lewis, M. (2003). Moneyball: The Art of Winning an Unfair Game. W. W. Norton.

Book Chapters

• Kahneman, D. (2011). Intuitions vs. formulas. In Thinking, Fast and Slow (Chapter 21). Farrar, Straus and Giroux.
• Mosier, K. L., & Skitka, L. J. (1996). Human decision makers and automated decision aids: Made for each other? In R. Parasuraman & M. Mouloua (Eds.), Automation and Human Performance: Theory and Applications (pp. 201-220). Lawrence Erlbaum Associates.

19. Summary Card

20. Glossary of Terms Used

21. Discussion Questions

For book clubs, classrooms, or self-reflection:

1. Think of a tool or app you stopped using after it let you down. If you had to defend that decision with its actual hit rate versus the alternative's, could you?

2. Why might a person who errs feel like "someone having a bad day" while a model that errs feels "broken"? Is that distinction ever justified?

3. The British public rejected an exam-grading algorithm whose errors were visible, in favor of teacher predictions whose errors were invisible. Was that aversion, a fairness judgment, or both?

4. Logg and Schlund argue that asking people which advisor they would prefer produces aversion, while watching what they actually do produces appreciation. What does that gap tell us about how to design surveys, and products?

5. Where in your own life or work is a system running that has never visibly failed? What would its failure look like, and would you notice?

Cập nhật lần cuối 15 tháng 10 2026