I Knew It All Along
Opening
Think of a major event from the past few years that surprised most people when it happened. An election result. A company collapse. A scientific finding that upended the consensus. Pick one and hold it in your mind.
Now ask yourself honestly: were you surprised? Or does it feel, looking back, like you always had a sense that it might happen? Like the signs were there, if you had only paid closer attention?
If the event now feels somewhat inevitable in retrospect, you are in excellent company. You are also experiencing one of the most underappreciated distortions in human memory, a bias so seamlessly woven into how we reconstruct the past that most people never notice it is happening at all.
The Researcher Who Couldn't Make People Forget
Baruch Fischhoff was a graduate student at Hebrew University in Jerusalem in the early 1970s, working under Daniel Kahneman during the most productive period of the heuristics-and-biases research programme. The question that interested him was one that no one had thought to study formally: once people know how something turned out, can they accurately remember what they thought before they knew?
The answer, which he began investigating in 1975, turned out to be a firm and consequential no.
In his initial studies, Fischhoff gave participants descriptions of historical events and their actual outcomes, then asked them to estimate the probability they would have assigned to those outcomes beforehand. A control group was given the same descriptions without the outcomes and made genuine probability estimates. The finding was striking: people who had been told the outcome consistently believed the outcome had been more probable all along, and they were largely unable to reconstruct the uncertainty that had genuinely existed beforehand.
Fischhoff called this "creeping determinism," the tendency to see past events as more inevitable than they actually were. He had a catchier name for it too, borrowed from everyday speech: the "I knew it all along" effect. It has since come to be known as hindsight bias.
The Doctors Who Would Have Caught It
The most unsettling application of hindsight bias is in professional judgment, where the stakes are high and the illusion of foresight can do real damage.
In a classic follow-up study, participants were given detailed clinical descriptions of patients and asked to diagnose the condition. The cases were chosen for their genuine ambiguity: multiple diagnoses were plausible from the available information. One group received only the clinical details. Another group received the same details along with the actual diagnosis. Both groups were asked what a competent physician should have concluded from the information.
The group told the actual outcome rated it as significantly more diagnosable than the group who had no outcome information. More troubling, when asked explicitly to set aside their knowledge of the outcome and judge the case as presented, they found they largely could not. The outcome had contaminated their reading of the evidence in a way that felt to them like good clinical reasoning.
This has since been replicated in legal contexts, where jurors told a defendant was convicted rate the prosecution's evidence as stronger than jurors not told the outcome. It has been replicated in financial contexts, where analysts told a company went bankrupt rate the warning signs as more obvious than analysts told the company survived. The pattern is stable: knowledge of outcomes reshapes the memory of what the evidence said, and people are rarely aware it is happening.
Why This Makes Learning So Hard
Hindsight bias is not merely a curiosity about memory. It has a specific and damaging effect on the thing organizations and individuals most need to do after something goes wrong: learn from it.
Post-mortems and after-action reviews exist precisely because we want to understand what happened and why, so that we can do better next time. But hindsight bias systematically corrupts this process. If the people reviewing the failure now believe that the failure was obvious in advance, they will draw the wrong lessons. They will look for someone who "should have seen it coming" rather than examining the actual information environment at the time. They will underestimate the genuine uncertainty that existed before the event. They will design interventions aimed at a clearer signal that often was not actually there.
The result is that the post-mortem produces false confidence rather than real understanding. The team walks away believing they have fixed the problem when they have actually only explained away the outcome in terms that were never available in advance.
Fischhoff noted something particularly demoralizing about this: simply warning people about hindsight bias does not reliably correct for it. Telling someone "try to remember what you thought before you knew the outcome" does not work well in practice, because the original state of uncertainty is genuinely difficult to reconstruct once the outcome is known. The knowledge is sticky in a way that instructions cannot easily override.
Hindsight Bias: The tendency, after learning an outcome, to believe that one would have predicted it in advance. People consistently overestimate the predictability of past events and struggle to reconstruct the genuine uncertainty that existed beforehand. The bias corrupts learning from experience and makes errors of judgment seem more avoidable in retrospect than they actually were.
What this means for a regular Tuesday
Write down your predictions before the outcome is known.
The only reliable protection against hindsight bias in your own reasoning is a written record made before the event. A dated note, a time-stamped email, a prediction journal: anything that preserves what you actually thought before you knew. Without this, your memory of your prior beliefs will silently update toward the outcome, and you will learn nothing from the comparison.
In post-mortems, reconstruct the information environment, not the outcome.
When reviewing something that went wrong, the question to ask is not "why did this happen?" but "what did the information look like at the time the decision was made?" Insist on reconstructing the actual evidence, the actual uncertainty, the actual options available, before the outcome was known. This is hard and uncomfortable, but it is the only version of a post-mortem that produces lessons that generalize rather than merely explains away the specific failure.
Be suspicious of confident retrospective analysis.
In business, in politics, in sports: the people who confidently explain after the fact why something was always going to happen are often experiencing hindsight bias, not displaying genuine insight. Their confidence is a product of knowing the answer, not of having had unusual foresight. Apply the same skepticism to yourself when you find yourself thinking "I always had a feeling this would happen."
How AI can help here
Use the pushback-oriented setup from The Man Who Robbed Banks With Lemon Juice in the main book for prompts that challenge your first instinct.
Use AI to timestamp your predictions and hold them against outcomes later. Before a decision or an uncertain event, describe what you think will happen and why. Ask the AI to store or repeat this back to you after the outcome is known, as a way of creating an outcome-independent record.
I want to record a prediction before I know the outcome, so I can compare it honestly afterward. The situation: [describe the decision, event, or uncertainty]. My prediction: [what you think will happen, and why, in as much detail as you can]. The key uncertainties I see: [what you genuinely don't know]. Please record this prediction. When I come back after the outcome is known, remind me of exactly what I said here, and help me assess whether my prediction was actually as good or as obvious as it feels in retrospect.When conducting a post-mortem on something that went wrong, use AI to help reconstruct the genuine information environment that existed before the failure, before hindsight contaminates the analysis.
I am reviewing a decision or outcome that did not go as hoped: [describe]. The actual outcome was: [describe]. I want to avoid hindsight bias in this review. Help me reconstruct the situation as it actually appeared before the outcome was known: What information was genuinely available? What was legitimately uncertain? What were the reasonable options given that uncertainty? What would a competent person with the available information have been expected to know? Do not treat the outcome as a signal that the decision was obviously wrong. Evaluate the decision based on the information that existed at the time it was made.Before reading analysis of a past event, record your current level of surprise, then use AI to help you separate genuine foresight from hindsight inflation.
Before I read the explanation of why [past event] happened, I want to record my honest reaction. On a scale of 1 to 10, my sense of how surprising this outcome was: [number]. My honest memory of what I expected beforehand: [describe]. After I have read the explanation, I will come back. At that point, ask me the same questions again and help me notice whether my sense of "I would have expected this" has increased after reading the explanation.
References
Baruch Fischhoff. "Hindsight Is Not Equal to Foresight: The Effect of Outcome Knowledge on Judgment Under Uncertainty." Journal of Experimental Psychology: Human Perception and Performance, Vol. 1, No. 3, 1975, pp. 288-299.
The original paper that named and documented the effect.
Baruch Fischhoff and Ruth Beyth. "'I Knew It Would Happen': Remembered Probabilities of Once-Future Things." Organizational Behavior and Human Performance, Vol. 13, No. 1, 1975, pp. 1-16.
A companion paper examining memory for one's own past predictions.
Scott A. Hawkins and Reid Hastie. "Hindsight: Biased Judgments of Past Events After the Outcomes Are Known." Psychological Bulletin, Vol. 107, No. 3, 1990, pp. 311-327.
A comprehensive meta-analysis of the hindsight bias literature.
Daniel Kahneman. Thinking, Fast and Slow. Farrar, Straus and Giroux, 2011. Chapter 19 covers hindsight bias alongside the closely related illusion of understanding.
The asterisk
Hindsight bias is one of the most robustly replicated effects in cognitive psychology, documented across cultures, ages, and domains ranging from medical diagnosis to financial forecasting to sports commentary. But a few nuances are worth knowing.
Ulrich Hoffrage and colleagues have shown that the magnitude of hindsight bias varies with domain expertise. Experts sometimes show less hindsight bias than novices in their own area, possibly because they have more developed mental models that allow them to represent the genuine uncertainty that existed at the time. This is not grounds for complacency among experts, but it does suggest that structured reasoning habits can offer partial protection.
Neal Roese and Kathleen Vohs reviewed the literature in 2012 and distinguished three components that often travel together under the "hindsight bias" label: memory distortion (misremembering what you thought before), inevitability inflation (believing the outcome was bound to happen), and foreseeability inflation (believing a reasonable person should have seen it coming). These three can come apart in interesting ways: you can believe an outcome was inevitable without believing you personally would have predicted it. Separating them matters for the design of learning interventions.
The practical literature on debiasing is modest in its ambitions. The most reliable approach is structural rather than cognitive: written pre-event predictions, explicit documentation of the information environment at decision time, and post-mortems that insist on reconstructing that environment before touching the outcome. Simply being told about hindsight bias reduces it slightly but does not come close to eliminating it.
The past is the only thing that feels more certain the less you remember of how uncertain it actually was.