Evidence status: conditional support Debiasing education produces a small average improvement on experimental measures, supported by a 2025 systematic review and meta-analysis. Results differ considerably across biases and instructional formats, and evidence for transfer to new settings and lasting behaviour remains limited. Knowing a term may help without amounting to recognition and correction in a real decision. This is research draft version 0.1.
How Everyday Psychology Takes Shape · Season One: Why Do We Think This Way? · Article 12
A person reads about anchoring and learns that the first number encountered can influence a later estimate. During a property negotiation, he still adjusts from the seller’s asking price. Another person can explain confirmation bias while opening only the search results that support her conclusion. Naming a bias afterwards is easy; noticing it while it is helping form a decision is a different event.
Psychological knowledge therefore faces an uncomfortable question. If we can name the error, why does it continue? One pessimistic answer treats bias as unchangeable. An optimistic one treats reading a list as an upgrade to rationality. Research falls between them. Training can improve performance, but the average effect is modest, variation is substantial, and transfer to a new task is particularly difficult.
The issue is not only whether a person knows more. It is whether knowledge can enter the structure in which judgment happens.
Knowing a definition is not recognising an event
Anchoring has a clean textbook form: present an irrelevant number, then ask for an estimate. The opening number in a real negotiation may simultaneously contain market information, a strategic signal and power. The decision-maker must notice that the starting point requires checking, distinguish relevant information from unjustified pull, and then use a better estimation method.
This requires at least a trigger, diagnosis and replacement. A trigger indicates that checking is needed now. Diagnosis identifies which process may be distorting the judgment. Replacement supplies an executable alternative. Remembering the term completes none of these automatically.
Hindsight bias illustrates the problem especially well. Once an outcome is known, prior uncertainty has already been reconstructed in memory. Someone may say accurately, “This is hindsight bias”, but cannot restore an unrecorded earlier probability by will. A timestamped forecast or independent record is needed to constrain the rewritten past.
Many biases are not simple error buttons either. Heuristics conserve time and can perform well in some environments. Reducing bias requires knowing when a shortcut is mismatched to the task, not demanding that intuition never be used.
How large are the effects of debiasing education?
A 2025 systematic review and meta-analysis in Nature Human Behaviour examined educational interventions among students. The authors screened randomised controlled research and included 54 studies with a pooled sample of 10,941 participants; 41 studies entered the meta-analysis. Compared with control conditions, debiasing education had a small positive effect on measured bias, Hedges’ g = 0.26, with a 95% confidence interval from 0.14 to 0.39.
The average is not a universal guarantee. Heterogeneity was high, and bias type significantly moderated results. Mixed-bias measures and overconfidence showed clearer improvement, while individual analyses for anchoring, representativeness and several other outcomes were not significant. Educational games had a larger average effect, but relatively few studies tested them, and the format alone cannot guarantee success.
The review also reported that every included study had unclear or high risk of bias in at least some respects, while funnel-plot asymmetry suggested possible publication bias. Much of the research asks whether students improve on similar tests, not whether they make better medical, career or financial decisions months later. Transfer from a training task to a changed context remains a central limitation.
“Training never works” is inconsistent with the synthesis. So is “once taught, bias has been corrected”.
Why other people’s biases are easier to see
Research on the bias blind spot finds that people acknowledge bias in human judgment generally while regarding themselves as comparatively objective. We observe another person’s conclusion and possible motive without seeing their entire thought process. For ourselves, reasons are immediately available, and that accessibility can be mistaken for evidence of impartiality.
Introspection does not reveal how many units an initial number shifted an estimate, how much familiarity contributed, or how a category changed attention. Many processes never appear in consciousness as “a bias is now being applied”. A sincere desire to be fair is valuable, but it does not measure one’s actual error.
Bias language can also become a weapon. “That is just your confirmation bias” often ends a discussion instead of examining evidence. Selective evidence does not make the other person’s conclusion necessarily false, and the accuser may be selectively noticing the other side’s selection. The name of a psychological mechanism cannot replace a factual argument.
What works better than a reminder alone
Research and practice suggest a general principle: changing a task before a predictable bias appears is often more reliable than asking for self-correction afterwards.
When making a numerical estimate, form an independent baseline before seeing another person’s number. In group discussion, have members record their judgments and reasons separately before comparing them. This reduces the ability of the first speaker to determine everyone’s starting point. For consequential forecasts, preserve probabilities and conditions, then reread them after the outcome.
Considering alternatives must be concrete. “Think of the opposite” can become another instruction completed superficially. Ask which evidence would be most likely to make the present conclusion wrong, assign someone to search for counterexamples, or apply criteria agreed before either side’s evidence was seen. For repeated processes, checklists, structured fields and mandatory pauses can place the prompt at the moment it is needed.
Feedback is indispensable. Without learning how a judgment differed from an outcome, calibration is impossible. Feedback must also be timely, interpretable and linked to controllable action. Some domains are noisy or reveal outcomes years later, so intuition cannot readily improve through experience alone. Statistical records and external audit become more important there.
Experiments by Carey Morewedge and colleagues in 2015 found that instructional videos and serious games could reduce several biases on experimental tasks, with some effects persisting for weeks. The interventions did more than supply definitions: participants practised, received feedback and identified bias across examples. Later synthesis warns against assuming broad transfer, but the work illustrates why training design has more promise than a list of labels.
Why individual correction also needs institutions
Bias in hiring, medicine, credit and public policy is not located only in one mind. Data selection, process order, performance targets and appeal mechanisms determine which judgments are amplified. If an organisation rewards speed, records no reasons and never examines outcomes across groups, a one-off “unconscious bias” course is unlikely to alter the structure.
Institutional methods are not automatically fair. A structured score can encode old bias in its criteria, and an algorithm can learn inequity from historical data. Their potential advantage is that processes can be recorded, compared and audited—provided someone tests whether the measurement is valid and affected people can correct data and challenge an outcome.
Individual responsibility remains, but its form changes. Rationality is not merely an effort to make the mind neutral. It includes choosing procedures that expose error: inviting independent review, preserving information available before the decision, giving dissent a place, and revising the method after results appear.
Will AI help debiasing or reproduce bias?
AI can prompt neglected options, produce a counterargument and organise unstructured material into common fields. It can also accommodate the framing of a question, make the user’s premise more fluent, or create a falsely balanced list. Asking a model to “check my biases” without independent factual sources may only produce another persuasive passage.
A better use gives the system verifiable tasks. List the assumptions on which this conclusion depends and identify which require external data. Compare alternatives using the same criteria. Separate known facts, inference and uncertainty. Its output still needs verification. AI may serve as one generator of objections within a procedure; it cannot be an unbiased final observer.
Systems themselves also require calibration, audit and recourse. Replacing a human judgment with an opaque score does not remove bias. It moves the places in which bias can arise.
The judgment that closes Season One
Why is knowing a bias usually insufficient? Because a term resides in memory while bias emerges through the attention, sequence, motivation, data and social relations of a particular moment. Without a trigger, an alternative procedure and feedback, knowledge may not enter that moment. Even when it does, the same intervention will not fit every bias or setting.
This does not make psychological knowledge useless. Learning about an effect identifies a risk, and education can produce modest but real average improvement. The mature goal is not to become “a person without biases”. It is to develop a way of judging in which error can become visible, be challenged and be corrected.
Across twelve articles, this season has moved from tip-of-the-tongue experience, hindsight and remembered episodes to AI, emotion and debiasing. Together they show that psychological phenomena are not hidden objects wholly inside the individual, waiting to receive a label. Retrieval and records shape memory. Judgment borrows cues from its environment. Emotion forms through body and meaning. People and tools share cognitive tasks. Psychological knowledge becomes useful not when we can recite a famous effect, but when it helps us reorganise how the next judgment will occur.
Primary research and further reading
- Swaryandini, G. et al. (2025), “Systematic review and meta-analysis of educational approaches to reduce cognitive biases among students”.
- Morewedge, C. K. et al. (2015), “Debiasing Decisions: Improved Decision Making With a Single Training Intervention”.
- Pronin, E., Lin, D. Y., & Ross, L. (2002), “The Bias Blind Spot: Perceptions of Bias in Self Versus Others”.
- Larrick, R. P. (2004), “Debiasing”.
- Korteling, J. E. H., Gerritsma, J. Y. J., & Toet, A. (2021), “Retention and Transfer of Cognitive Bias Mitigation Interventions”.
- Milkman, K. L., Chugh, D., & Bazerman, M. H. (2009), “How Can Decision Making Be Improved?”.
Series navigation: Season One overview: How Everyday Psychology Takes Shape
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.