012 min
Where hindsight bias shows up
An illustrative case, with a role rather than a person. A staff engineer runs an incident review after a payment service was unavailable for forty minutes. The cause was a configuration change shipped the previous afternoon. Everyone in the room can now see that the change touched a retry setting, and that the same setting had caused a smaller problem a year earlier. Someone says the change should never have been approved. The reviewer who approved it agrees, and cannot explain what they were thinking at the time.
The moment the effect starts is visible in the room. It starts when the cause is named out loud. Before that sentence, the group is listing many things that could have gone wrong. After it, the group is explaining the one thing that did.
Now change a single input and run it again. The identical change ships, the retry setting behaves, and nothing breaks. There is no review. Nobody calls the approval careless. The decision was the same in both runs. Only the result differed.
That contrast is the pattern. Hindsight bias is not a judgement about a decision. It is a judgement about a decision made after its result is known, by people who can no longer recover how uncertain they were. Neal Roese and Kathleen Vohs (2012) describe it as the inability to recapture the feeling of uncertainty that came before the event.
022 min
Why hindsight bias happens
Learning an outcome does not leave what you already knew unchanged. It changes it. Roese and Vohs (2012) group the causes into three kinds of input.
Knowledge updating
Human memory is reconstructive, which means a memory is rebuilt at the moment of recall rather than played back unchanged. Knowledge updating is the first input. A new fact is connected to older knowledge. The parts of that knowledge which fit it are strengthened. The parts that do not fit stay inactive. After an outage, the earlier retry incident is easy to bring to mind. The five other changes shipped that week are not.
Sensemaking
The second input is sensemaking, which is the work of building a causal account of why the outcome happened. Roese and Vohs report that the better the story, the larger the bias. One clean cause produces more hindsight bias than a set of causes that interact messily.
Ease and motivation
The third input is metacognitive, which means it is a judgement about your own thinking rather than about the world. When an explanation arrives easily, that ease is misread as evidence the outcome was likely. The fourth input is motivational. People prefer a world that is orderly and predictable, and prefer not to be blamed.
The adaptive reason is plain. Updating what you know when new facts arrive is what learning requires. The cost is that the update leaves no record of the earlier state, so the person cannot measure how far they moved. Baruch Fischhoff (1975) named this creeping determinism: in retrospect, events appear to have unfolded with an inner necessity that was never visible in advance.
One hedge: this four-input account is a synthesis across many experiments, not a single model tested end to end.
031 min
The three levels of hindsight bias
Hindsight bias is not one belief. Blank, Nestler, von Collani and Fischer (2008) proposed three components. Nestler, Blank and Egloff (2010) then showed experimentally that the components can be separated. Roese and Vohs (2012) arranged them as stacked levels, where each level contains the one below it.
| Level | The claim the person is making | How it is measured |
|---|---|---|
| Memory distortion | "I said it would happen" | Compare a recalled estimate against the estimate the person actually recorded earlier |
| Inevitability | "It had to happen" | Rate whether any different outcome was possible in the circumstances |
| Foreseeability | "I knew it would happen" | Self-report on whether the outcome was predictable at the time |
The levels are worth keeping apart because they break at different points. Memory distortion needs an earlier judgement to distort, so recording the judgement removes the material it works on. Inevitability survives that fix. A person can accept their own written estimate and still insist that no other outcome was possible. Foreseeability is the level that carries blame, because it is a claim about a person's ability rather than about the world.
Roese and Vohs map inputs onto levels: recollection and knowledge updating feed memory distortion, sensemaking feeds inevitability, and the metacognitive and motivational inputs feed foreseeability.
042 min
Where hindsight bias comes from
Baruch Fischhoff ran the founding experiments at the Hebrew University of Jerusalem and published them in 1975 in the Journal of Experimental Psychology: Human Perception and Performance.
Subjects read a 150-word passage about the British campaign against the Gurkhas of Nepal, which began in November 1814. Fischhoff took it from E. L. Woodward's book The Age of Reform (1938). Four outcomes were offered: a British victory, a Gurkha victory, a stalemate with a peace settlement, and a stalemate without one. One group read the passage alone. Four other groups read the same passage with one extra sentence naming one outcome as what had happened. Three of those four groups were given an outcome that had not happened. Everyone then estimated the probability of all four.
Six sub-experiments gave 24 chances to test the prediction. In all 24, being told an outcome raised the probability assigned to it. The rise ranged from 3.6 percentage points to 23.4, and averaged 10.8. Around 70% of the told-an-outcome subjects placed it above the average given by subjects who were told nothing.
A second experiment asked subjects to answer as they would have if they had not known the outcome. They could not. The average gap was 9.2 points, against 10.8 when nobody was asked to compensate.
"In retrospect, we seem to perceive the logic of the events which unfold themselves in a regular or linear fashion according to a recognizable pattern with an alleged inner necessity."
Fischhoff and Ruth Beyth published the companion study the same year, in Organizational Behavior and Human Performance. They asked people who had estimated the likely outcomes of President Nixon's 1972 trips to Peking and Moscow to recall their own earlier estimates after the visits ended. Recalled probabilities were higher than the originals for events the person believed had happened.
051 min
Hindsight bias in a negligence trial
Kim Kamin and Jeffrey Rachlinski moved the effect out of the vignette and into a legal decision. Their study appeared in Law and Human Behavior in 1995, pages 89 to 104.
Participants assessed whether a town should pay for precautions to protect a riverside property owner against flood damage. One group saw the question in foresight, framed as an administrative hearing held before anything happened. They were asked whether the precaution was worth taking. A second group saw parallel material in hindsight, framed as a trial held after a flood had occurred. They were asked whether the failure to take the precaution had been negligent.
Three quarters of the foresight group concluded that a flood was too unlikely to justify further precautions. A majority of the hindsight group judged that same conclusion negligent. The hindsight group also gave higher probability estimates for the flood itself.
The study then tested the remedy that courts already use. Judges instruct jurors to set the outcome aside and assess the decision as it stood at the time. Those instructions did not reduce the bias.
What this case adds to the laboratory work is the stake. The foresight group's answer was the defendant's answer. Two groups reached opposite verdicts about one decision, and the only difference between them was knowing how it ended.
061 min
What hindsight bias costs one person
For one person, the cost is losing the ability to grade their own forecasts.
Take an estimate. An engineer says a migration will take three weeks. It takes seven. In the review afterwards, every cause of the overrun is visible, and each one looks like something a careful person would have counted. The engineer concludes they were careless. The direction of the error is consistent. A past estimate gets judged against information that did not exist when it was made, so the estimator looks worse than they actually were.
The effect works in the opposite direction on successes. A founder who set a price that worked will remember being more confident about that number than they were. The next pricing decision then gets less analysis, because the previous one is remembered as obvious.
Roese and Vohs (2012) name overconfidence as one of the two main consequences: exaggerating your own ability to analyse a situation, which leads to riskier positions later. Fischhoff (1975) closed his paper with the same point, stated more directly. If you cannot recover how uncertain you were, you cannot measure what the outcome actually taught you.
072 min
How hindsight bias compounds across a team
One person's distorted memory stays with that person. Inside an organisation it turns into a process.
The mechanism of compounding is blame. Roese and Vohs (2012) describe myopia as the second main consequence. The search for causes stops at the first plausible candidate. People explain events through the intentional acts of other people, so that candidate is usually a person. Their example is the Concorde crash near Paris in July 2000, which killed 109 people. An observer can settle on pilot error and stop, and so miss the strip of metal on the runway that the investigation actually identified.
Once a review names a person as the cause, the effect sustains itself. Roese and Vohs report that hindsight bias raises punitiveness, drawing on studies of jurors. People who expect to be judged that way report less, and report later. That removes the raw material the next review depends on.
Software operations built a counter-practice around exactly this. John Allspaw, then at Etsy, published "Blameless PostMortems and a Just Culture" in May 2012 and named the problem directly:
"We accept that the Hindsight Bias will continue to cloud our assessment of past events, and work hard to eliminate it."
Google's Site Reliability Engineering book carries the same rule in Chapter 15, written by John Lunney and Sue Lueder. A postmortem is blameless when it identifies the contributing causes of an incident "without indicting any individual or team for bad or inappropriate behavior".
082 min
Hindsight bias in product, design, and AI
Three places this reaches people who build software.
Experiment results
An A/B test finishes and one variant wins. Within a day the result is obvious to everyone, and the losing variant is described as something nobody really expected to work. The team then runs fewer tests, because outcomes now feel predictable. This is the foreseeability level at work. The belief is about the team's own judgement rather than about the data. The guard is one record: write down the predicted winner and the confidence before the test starts, then read that record when it ends.
Model evaluation
An analyst reviewing a model's failures has the labels in front of them. Every misclassification has a visible reason, and each reason suggests a feature the model should have used. Adding those features tunes the model to a test set that has already been read. The level at work here is inevitability. The failure looks like it had to happen for a reason the analyst can now name, so a fix aimed at that one reason looks sufficient.
Dashboards and data visualisation
Roese and Vohs close their 2012 review with a warning about visualisation technology. Rendering a dataset as a picture raises confidence in whatever reading the picture supports. They give two rules. View several visualisations at once, so one arbitrary angle cannot dominate. Rebuild the visualisation with parts of the data removed, and check whether the reading survives.
One version is produced deliberately rather than suffered. A company retrospective written for recruiting or investor audiences presents a run of decisions as though it had been a plan. The tell is that no abandoned option is named.
092 min
How to guard against hindsight bias
Awareness is not a technique. Telling people the bias exists does not remove it, and Kamin and Rachlinski (1995) watched the courtroom version of that warning fail on jurors. What works changes the process, not the attitude.
Write the prediction down first
Memory distortion needs an earlier judgement to distort. A recorded estimate, carrying a date and a number, takes that material away. Fischhoff and Beyth's (1975) design is the template: collect probabilities before, then compare them with what people recall afterwards. In daily work that is a decision log entry, a written estimate on a ticket, or a stated confidence level on an experiment.
Consider the opposite
Roese and Vohs (2012) single this out as the one debiasing strategy that has worked consistently. The decision maker explains how an outcome that did not occur could well have occurred. Arkes and colleagues (1988) tested it on physicians. Those asked to give a reason why each candidate diagnosis might be true showed less hindsight bias. It was reduced rather than removed.
Prompt: name two ways this could have gone differently, and give one reason each of them was plausible at the time.
Two alternatives, not ten. Roese and Vohs report the strategy backfires past about three: when generating more alternatives starts to feel difficult, people misread that difficulty as evidence the alternatives were implausible.
Keep the decision and the result apart
Ask two separate questions in any review. What information existed at the time? What would a reasonable person have chosen with only that information? Allspaw's just-culture writing turns this into a standing meeting rule, which matters, because no individual can decide to switch the effect off.
101 min
Common misunderstandings about hindsight bias
People repeat the saying "hindsight is 20/20", which means the past is seen clearly. Fischhoff's (1975) results say something close to the opposite. His subjects did not see the past clearly. They saw a version of it that had been edited to fit the ending, and they could not tell it had been edited. Three specific errors follow from that saying.
- "I knew it all along, so my judgement is good." The claim describes a prediction that was never written down. Roese and Vohs (2012) classify foreseeability as the level most driven by the wish to appear knowledgeable.
- "It is only a memory problem, so writing things down fixes it." Records remove one of the three levels. Someone can accept their own recorded estimate and still believe the world allowed no other outcome.
- "It needs a dramatic outcome." Fischhoff used a 150-word account of an obscure 1814 campaign and a set of clinical case notes. Small, dull outcomes produce the effect too.
There is also a case where guarding against it would be a mistake. The target is never the update itself, only the missing record of what you believed before it.
111 min
Hindsight bias vs. nearby concepts
Four other ideas get called hindsight bias. Each is separated from it by one axis.
| Nearby concept | What it says | The axis that separates it |
|---|---|---|
| Not in the library yetOutcome bias | ||
| Not in the library yetOverconfidence | ||
| CommunicationCurse of knowledge | ||
| ResearchSurvivorship bias |
Hindsight bias has no named sub-effects the way availability does. It has three separable beliefs instead.
121 min
Where hindsight bias evidence is contested
Whether the effect exists is not in dispute. Three things around it are.
Effect size under replication. Jieying Chen and Gilad Feldman ran a preregistered replication and published it in the Journal of Experimental Social Psychology in 2021. Study 1, with 890 participants, reproduced Fischhoff's second experiment at a mean effect size of d = 0.60. Study 2, with 608 participants, reproduced Slovic and Fischhoff (1977) at d = 0.40. The direction held in both. The sizes were smaller than the original reports implied. Their extensions covering surprise, confidence and task difficulty produced weak effects or none.
Expertise. Christensen-Szalanski and Willham (1991) meta-analysed the literature and found that expertise went with less hindsight bias. Guilbault, Bryant, Brockway and Posavac (2004) meta-analysed 95 studies, applied a stricter definition of expertise, and found no relation. Their overall mean effect size was d = .39, with a 95% confidence interval of .36 to .42. This disagreement is unresolved.
How far debiasing goes. Considering the opposite reduces the bias without removing it, and Roese and Vohs (2012) report it reversing when too many alternatives are generated.
What to do while this is unsettled: do not treat seniority as a control. Apply the recorded-estimate procedure to everyone, including the people most confident that they do not need it.
131 min
How hindsight bias research changed since 1975
Fischhoff's 1975 papers established the effect and gave it a name, creeping determinism. That name did not last. "Hindsight bias" and the "knew-it-all-along effect" replaced it, and the second phrase traces back to the title of the Fischhoff and Beyth paper, "I knew it would happen".
What changed after that:
- 1990 — Hawkins and Hastie publish a review that frames the effect as a problem of reconstructing an earlier judgement.
- 1991 — Christensen-Szalanski and Willham publish the first meta-analysis and report that expertise reduces the bias.
- 2004 — Guilbault and colleagues publish a larger meta-analysis and overturn that expertise finding.
- 2008 to 2010 — Blank, Nestler and colleagues split the effect into three components and show they come apart.
- 2012 — Roese and Vohs organise those components into levels, and Allspaw makes the bias an operating rule inside software incident reviews.
- 2021 — Chen and Feldman publish a preregistered replication with smaller effect sizes than the originals.
- 2025 — Fischhoff publishes a fifty-year reflection in the same journal. It names clinical diagnosis, patent evaluation, legal adjudication and safety engineering as fields that took the bias up.
Current practice still matches the original meaning on the central claim. It has moved past it in one respect. Fischhoff measured a single number. The field now measures three beliefs and does not assume they move together.
?7 questions
Questions people ask
What causes hindsight bias?
What is an example of hindsight bias in engineering?
How do you avoid hindsight bias?
What is the difference between hindsight bias and outcome bias?
Is hindsight bias real, and does it still replicate?
Does knowing about hindsight bias protect you from it?
Are experts less affected by hindsight bias?
§10 sources
Sources on hindsight bias
Fischhoff, B. (1975). Hindsight is not equal to foresight. Journal of Experimental Psychology: Human Perception and Performance, 1(3), 288–299.
Fischhoff, B., & Beyth, R. (1975). "I knew it would happen". Organizational Behavior and Human Performance, 13, 1–16.
Roese, N. J., & Vohs, K. D. (2012). Hindsight bias. Perspectives on Psychological Science, 7(5), 411–426.
Kamin, K. A., & Rachlinski, J. J. (1995). Ex post does not equal ex ante. Law and Human Behavior, 19, 89–104.
Show all 10 sourcesShow fewer sources
Guilbault, R. L., Bryant, F. B., Brockway, J. H., & Posavac, E. J. (2004). A meta-analysis of research on hindsight bias. Basic and Applied Social Psychology, 26, 103–117.
Chen, J., & Feldman, G. (2021). Retrospective and prospective hindsight bias. Journal of Experimental Social Psychology.
Fischhoff, B. (2025). Fifty years of hindsight bias research. Journal of Experimental Psychology: Human Perception and Performance, 51(2), 143–150.
Allspaw, J. (2012). Blameless PostMortems and a Just Culture. Etsy.
Lunney, J., & Lueder, S. Postmortem Culture. In Site Reliability Engineering (Google).
The Decision Lab. Hindsight bias.
