What we look for
We follow research on cognitive biases, judgment and decision making. We are especially interested in findings that confirm, narrow or challenge familiar claims, and in the growing work on decisions made with AI.
How a review works
We start with the strongest source we can find, preferably the original paper or a serious review. Then we ask what was actually tested, what the result supports, where the limits are, and whether it changes anything we currently say.
Not every new paper becomes a new page. Sometimes the right result is simply to add a source, soften a claim, or leave the library unchanged.
What we publish
Reviewed entries show what we found in the literature, when we last checked it, and where the evidence is strong, limited or mixed. Comparisons help separate ideas that are easy to confuse. Decision pages connect the research to situations where a person may actually need it.
Current focus
Our current work includes decisions made with AI, forecasting, information evaluation, product decisions, project estimation and well-known findings whose popular versions may be stronger than the evidence.
Research notes
These are short syntheses of sources we have reviewed. We mark preprints and unsettled findings instead of treating recency as certainty.
Recent studies make a stronger case that LLM outputs can show repeatable decision patterns that resemble named cognitive biases. They also show why we should be careful with the label: the effects depend on the task, model, prompt and conversational context, and a debiasing prompt that helps one kind of task can hurt another.
Research supports a recurring pattern of optimistic completion-time estimates. The useful response is to compare the plan with similar completed work and keep uncertainty visible, not to multiply every estimate by one fixed number.
Corrections usually improve factual accuracy, but old misinformation can still influence later reasoning. That continued influence is not the same as a backfire effect, and the distinction changes how corrections should be evaluated.
Anchoring is strongly supported in human numerical judgment, and newer studies show two AI-related risks: people can anchor on AI recommendations, while model outputs can also shift when prompts contain numerical anchors.
Availability is a useful shortcut when accessible examples track real frequency. It becomes misleading when accessibility comes from non-diagnostic causes, and some popular claims about dramatic media risks and ease of retrieval are less universal than textbook examples suggest.
Confirmation bias is a broad family of belief-consistent information-processing tendencies. Classic hypothesis-testing work also shows why a narrower problem matters: a test can fit the favored hypothesis while offering little information against plausible alternatives.
Once you know an answer, your own knowledge can distort estimates of what another person knows. The most useful correction is not 'explain more simply' in the abstract, but feedback from the less-informed person's actual perspective.
Several different processes can make past periods look better from the present, but they should not be collapsed into one universal bias. Event recollection, changing emotional intensity, exposure to current negative information, and real historical trends need separate evidence.
Adding an inferior option can change the choice between stronger alternatives, but not every three-tier pricing table creates a reliable decoy effect. The position of the options, prior preferences and the decision context all matter.
Defaults often change what people choose, but staying with a pre-selected or existing option does not by itself show strong preference. The useful question is what made the option sticky and whether the result still serves the person's goals.
The Dunning–Kruger Effect is better understood as a question about calibration between performance and self-assessment than as the internet rule that incompetent people are always extremely confident.
Risky-choice framing is a robust finding, but the word 'framing' covers several different research paradigms. Treating every positive-versus-negative message as the same effect hides important differences in evidence and mechanism.
The famous parole-board study found a strong sequence pattern around food breaks, but it did not measure hunger or randomly manipulate meals. Later critiques and simulations provide plausible alternative explanations for part of the striking result.
Loss aversion is one of the best-known ideas in behavioural economics, but recent meta-analyses disagree substantially about its average size and about which experimental designs provide clean evidence for it.
Past, unrecoverable costs can influence whether people continue, but escalation of commitment is a broader process. Responsibility for the original choice, completion pressure, project structure and advice can also shape the next investment.
Surrogation is more specific than the slogan 'what gets measured gets managed.' Research in strategic performance systems shows how a measure can start to stand in for the construct it was designed to represent, and how decision processes can reduce that substitution.
Systematic bias is not one mental shortcut. It is a measurement and statistical problem: a process can keep pushing results away from a target in a consistent direction, and collecting more data from the same biased process does not automatically fix it.