Learn it → Apply it → Save as Skill
A single piece of evidence feels more convincing than it is when you forget how rare the thing already was. The base rate, how common something is in the wider population, should anchor your estimate. New evidence should shift that anchor, not replace it.
The Dr. Wallace’s Favorite Mental Models is designed to turn useful ways of thinking into reusable tools. Each model gives you a fast visual overview, deeper interactive learning, practical application exercises and an AI-ready Markdown skill you can save for future use.
Instead of merely reading about mental models, you can gradually build your own AI skills library: a collection of portable reasoning skills that can help ChatGPT, Claude, Grok and other AI assistants apply the mental models you value.
Before you weigh a test result, an interview red flag, or a dramatic story, ask how often the underlying thing actually happens in the group you're really drawing from. That number, the base rate, should anchor your estimate before any new evidence adjusts it.
Typical guess after one positive test vs the actual probability, once the base rate is included
Important: the exact gap depends on how rare the condition is and how accurate the test is. The pattern, evidence looking stronger than it is, is what to watch for.
The model is most useful whenever a test, alert, report, or story updates your belief about something whose overall frequency you can estimate.
Base rate neglect happens when the second step, weighing new evidence, crowds out the first step entirely.
Before any new evidence arrives, establish how common the thing actually is in the relevant population. This is the base rate.
Note how reliable the new evidence is: how often it's right when the thing is true, and how often it's wrong when it isn't.
Update the base rate with the evidence's reliability. The evidence should shift your estimate, not replace it outright.
Ask whether the resulting number still feels overconfident, and check whether the reference class you used actually fits the case.
Open a use case to see what the model changes.
Even an accurate test produces mostly false alarms when the condition itself is rare, because the large healthy population generates more false positives than the small affected population generates true ones. Ask for prevalence, sensitivity, and the false-positive rate before reacting to one result.
A single hesitant answer or unusual resume gap is weak evidence next to how common strong and weak candidates already are in your applicant pool. Weigh the signal's reliability against that starting distribution instead of letting one moment decide.
If genuine fraud is rare and the detector has even a modest false-positive rate, most alerts will be innocent. Prioritize investigation using the base rate of real incidents, not the raw count of alerts.
Spam filters, content moderation systems, and AI hallucination or safety detectors face the identical math: a rare target event plus an imperfect detector produces many false alarms. Report precision alongside any flag rate.
A DNA partial match or an eyewitness identification means something very different depending on the size and makeup of the population it's drawn from. The reference class used in court materially changes the real probability.
Compare the specific story's implied odds against the base rate for that category of venture or prediction before letting a compelling narrative override it.
This is a widely used illustrative example, built from the kind of numbers real screening programs report, designed to teach the method.
Expand for the studies behind this model and why the framing of the question matters so much.
Probability theorists including Thomas Bayes and Pierre-Simon Laplace had long since formalized how a prior probability should combine with new evidence. Psychologists Daniel Kahneman and Amos Tversky, in a series of studies in the early 1970s, documented how reliably people ignore that combination in practice and named the resulting error base-rate neglect, sometimes called the base-rate fallacy. Maya Bar-Hillel's later work extended and refined the account of when and why it occurs.
In Kahneman and Tversky's classic version, a city has two cab companies: 85% of cabs are Green and 15% are Blue. A witness identifies the cab in a hit-and-run as Blue, and testing shows the witness correctly names the color 80% of the time. Most people estimate the cab was probably Blue with well over 50% confidence, some above 80%. The correct Bayesian answer is about 41%, because the much larger pool of Green cabs still produces more false "Blue" identifications than the smaller pool of true Blue cabs produces correct ones.
Psychologist Gerd Gigerenzer's research found that reframing the same problem as natural frequencies, "10 out of every 1,000 people," rather than as a percentage, "a 1% prevalence," sharply improves people's accuracy, including among trained professionals. In a widely cited 1982 study by David Eddy, most physicians given prevalence, sensitivity, and false-positive rate for a screening test estimated the chance of true disease after a positive result at around 75%; worked through the natural frequencies, the correct figure was roughly 7 to 8%.
Treating a positive result as near-certain proof while ignoring how rare the underlying thing is.
Applying a base rate from a population that doesn't match your actual case, such as a national average for a sharply different local group.
Disease prevalence, fraud rates, and market conditions all shift over time. An outdated number quietly poisons the whole estimate.
A dramatic anecdote can feel more persuasive than a well-established base rate, even though one story isn't evidence about frequency.
Folding one piece of evidence into an estimate twice, under two different names, inflates confidence without adding information.
A low base rate is not the same as an acceptable risk when the cost of being wrong is severe or irreversible.
the individual-level evidence is overwhelming and specific, such as a confession or a confirmed audit trail. At that point it has effectively superseded the base rate.
the consequences are catastrophic or irreversible. A low probability is not the same as an acceptable risk when the downside is severe.
the reference population may not match your case, or the number was measured under conditions that no longer hold.
This browser tool structures your thinking. It does not send your information anywhere or perform AI analysis.
Save as Skill converts the model into portable Markdown instructions that teach an AI when and how to apply base rate reasoning, what questions to ask, what evidence to seek, what mistakes to avoid, and how to structure its analysis.
Save the skills you find useful and organize them into your own reusable library. The Markdown is intentionally model-neutral, so it can be used wherever an AI system accepts uploaded files, project knowledge, custom instructions, or reusable text context. Platform capabilities differ, so this is a portable skill file rather than a claim of one-click installation.
No, though both come from the same research tradition. Base rate neglect is under-weighting how common something already is. The conjunction fallacy is judging a specific, detailed scenario as more probable than a broader category that necessarily includes it. They often appear together but are distinct errors.
No. Probability and consequence are different questions. A rare event with a severe or irreversible cost can still deserve real attention; the base rate tells you how likely it is, not how much it would matter if it happened.
Use the best available reference class and say plainly how confident you are in it. Natural frequencies, "how many out of a round number," are easier to reason with than percentages. An uncertain estimate stated honestly is still more useful than defaulting to fifty-fifty or ignoring the question entirely.