Hindsight bias

“I called it”

Before the finale of a baking competition, a viewer tells friends the contest is wide open and leans slightly toward one of the other two finalists. After the winner is announced, they say: “Honestly, it was obvious from week three that she’d win.”

At the time, the viewer’s own forecast was uncertain and favored someone else. The outcome has rewritten both the memory of what they thought and their sense of how clear it was. This is the core case: a prediction remembered as better than it was.

“You should have known”

A friend plans a hike for a day with a 20% chance of afternoon storms. A storm arrives and the group gets soaked. Afterward, someone says: “The weather was obviously going to turn. You should never have picked that day.”

A 20% forecast means storms were possible but unlikely. Knowing that one came makes it feel as though the warning signs were clear. Judging the friend’s choice by what’s known now holds them responsible for information nobody had. This is the version aimed at other people, and it’s the one with the most practical cost: in reviews of medical and legal decisions, it can make a reasonable decision look negligent. (Judging a decision’s quality by its result is also called outcome bias, a close relative.)

The warning signs that were “all there”

After a club’s summer fundraiser draws a small crowd, the organizers meet to review it. “Looking back, the signs were all there: the date clashed with a local fair, and ticket sales were slow the first week.” Nobody mentions that ticket sales had also been slow the first week of last year’s event, which sold out.

This is the subtler form. The details that fit the outcome now look like evidence; the ones that pointed the other way are forgotten. Fischhoff’s original study found something similar: being told the outcome changed which details of a story people judged to be relevant. The review feels like careful analysis, but it’s sorting evidence by an answer that nobody had at the time.

Neal Roese and Kathleen Vohs describe three levels, which can occur separately or together:

• Memory distortion: misremembering your earlier prediction as closer to the outcome. “I always said it would rain.”
• Inevitability: believing the outcome was bound to happen. “Given the clash with the fair, it was never going to work.”
• Foreseeability: believing you, or someone, could have predicted it. “Anyone could have seen that coming.”

• When the outcome really was foreseeable with what was known at the time. If there was a clear forecast, a written warning or a documented risk that someone ignored, saying “that was predictable” is a fair judgment, not a distortion.
• When there’s a record made beforehand. A prediction written down in advance settles what you actually expected, so pointing to it isn’t hindsight.
• When you update for next time. Outcomes are evidence. Changing your expectations for future cases because of what happened is learning. Baruch Fischhoff draws the line this way: if an event that seemed 50/50 goes one way, raising your estimate for the next similar case can be reasonable; believing you’d have said so before this one is the bias.

The test: could this have been predicted with only the information available then, and is there any record that someone did?

A warning that was posted in advance

The city posts signs on a street two weeks ahead saying cars must be moved on Tuesday for repaving. A resident ignores them and their car is towed. A neighbor says: “You really should have seen that coming.”

The neighbor’s judgment doesn’t rely on knowing the outcome. The information was available, in writing, before the event, and it pointed directly at what happened. That’s the foreseeable at the time condition above.

Learning from a string of results

A gardener has planted tomatoes in early April for four years. Three of those years a late frost killed the seedlings. “Next year I’ll wait until May.”

This uses outcomes, but it doesn’t claim any of the earlier frosts was predictable. It treats them as a track record and adjusts the next decision. That’s updating for next time, not rewriting the past.

Fischhoff’s original explanation was that outcome knowledge is absorbed automatically. Once you learn what happened, you make sense of the situation in light of it, and your earlier view isn’t kept anywhere you can go back to. He called the result creeping determinism.

Roese and Vohs’s review groups the causes into three kinds:

• Cognitive: people selectively recall information that fits what they now know to be true, and build an explanation around it.
• Metacognitive: an outcome that’s easy to understand feels as if it was always likely. The ease of explaining it is mistaken for how predictable it was.
• Motivational: people want the world to seem orderly and predictable, and want to avoid blame.

Daniel Kahneman places hindsight bias within what he calls the illusion of understanding: good stories about the past make it seem as if the outcome could have been known. He draws on Phil Rosenzweig’s The Halo Effect to show one consequence. When a company does well, its chief executive gets described as flexible and decisive; if it later does badly, the same person is called rigid and confused. Knowing the outcome colors the judgment of the person’s traits, a Halo effect driven by the result.

• Warnings don’t work. Explaining hindsight bias to people and urging them to avoid it failed in early studies, and Fischhoff’s 2025 review of the field still finds no discernible effect of warnings.
• Consider how it could have gone otherwise, briefly. In a study of 194 neuropsychologists, Hal Arkes and colleagues asked some to note one reason why each possible diagnosis in a case might have been correct. Fewer of them showed the bias (41%, against 58% of those who didn’t list reasons). Roese and Vohs’s review also treats considering alternative explanations as a way to reduce it.
• But don’t overdo it. Lawrence Sanna and colleagues found that listing two ways an event could have turned out differently left no significant bias, while listing ten increased it. Struggling to think of alternatives seemed to suggest there weren’t many.
• Judge a decision by what was knowable then. When reviewing someone’s choice, ask what they knew at the time and what the odds looked like, before looking at what happened. Writing predictions down in advance gives you something to check against; that’s ordinary record-keeping rather than a tested remedy, but it removes the need to rely on memory.

Status: replicates robustly. Hindsight bias has been found across well over a hundred studies, and a preregistered replication of the original experiments found it again, though smaller than first reported.

• Fischhoff (1975) gave students short passages describing a historical or clinical event (one was a war between the British and the Gurkhas in 1814), each with four possible outcomes. Those told which outcome happened, and asked to answer as if they didn’t know, gave it a higher probability than those given no outcome. They also judged details that fit the outcome as more relevant, and were largely unaware of the influence. Fischhoff and Beyth (1975) showed the memory version: people who gave probabilities for possible events before a presidential trip abroad later remembered having given higher probabilities to the events they believed had since happened.
• Two meta-analyses, Christensen-Szalanski and Willham (1991, 122 studies) and Guilbault and colleagues (2004, 95 studies with 252 effect sizes), both found a consistent but fairly small average effect, around d = 0.35 to 0.4. (Here d measures the size of a difference between groups; 0.2 is conventionally “small” and 0.5 “medium”.) The status rests mainly on the preregistered replication below, not on these pooled estimates.
• Chen and colleagues (2021) ran preregistered close replications of Fischhoff (1975) and of a related 1977 experiment by Slovic and Fischhoff, with American online participants. Both found hindsight bias (average d = 0.60 for judging past events and 0.40 for predicting whether a result would repeat). The original 1975 effect, recalculated, was much larger (d = 1.13). Results for two of the four original passages, clinical cases that no longer read as they did in 1975, were set aside in consultation with Fischhoff and the journal’s editor.
• A failed replication: a preregistered replication of Fischhoff’s second experiment in Brazil (Seda and colleagues, n = 431, posted in 2025 as a preprint and not yet peer reviewed) found no significant difference on its main test. An exploratory analysis pooling all comparisons found a small effect (d = 0.12).

What remains uncertain is how large the bias is in any given setting: it varies with the task, the kind of judgment and how surprising the outcome was, and it’s smaller in modern replications than in the original studies. Fischhoff has also written that his memory study might better have been called the “I wasn’t all that surprised” effect, since people rarely remembered having been certain. That the bias exists is not in serious doubt.