Understanding cognitive biases

A bias is not a fault: it is a brain shortcut that works almost every time, and fails in predictable situations.

≈ 7 min readReading

In brief

A bias is a systematic, predictable error, not an oversight or bad faith.
Projects and teams combine the conditions where biases strike: uncertainty, pressure, delayed feedback.
Willpower is not enough: you protect yourself with methods (outside view, independent estimates, pre-mortem, written criteria).
Each entry gives a frequency, domains, signals and reflexes.

On this page

What a cognitive bias is
Why the brain does this
Why in IT and in management
How to read the entries
Three ideas to protect yourself
Limits and precautions
Sources and verification

What a cognitive bias is#

A cognitive bias is a systematic gap between what we judge or decide and what a rigorous analysis would give. The word matters: systematic. A random error cancels out over time; a bias always pushes the same way, which is what makes it dangerous, but also predictable.

It is not the same thing as:

ignorance: you don't know, whereas with a bias you could know;
bad faith: someone who falls victim to a bias is sincere;
emotion: some biases are emotional, many are purely mechanical.

The “heuristics and biases” research program was launched by Amos Tversky and Daniel Kahneman in the 1970s. Their 1974 paper describes three founding shortcuts: representativeness (we judge by resemblance to a model), availability (we judge by how easily cases come to mind) and anchoring and adjustment (we start from a value and adjust too little).

Why the brain does this#

Thinking costs energy and time. The brain therefore uses heuristics: simple, fast rules, often accurate. Kahneman groups them under the image of a “System 1” (fast, automatic) and a “System 2” (slow, deliberate). It is a convenient model, not an anatomical description.

A contrary view worth knowing

Gerd Gigerenzer argues that these heuristics are not defects: in a real, uncertain environment, a simple rule often beats a sophisticated calculation. The two views agree on the essential point: the shortcut is good in an environment it knows, and bad in an environment that traps it.

The shortcut is useful in an environment it knows. In a project, the environment is rarely regular and feedback comes late: we go wrong without learning from it.Hover over or tap an element to read its details.

Why in IT and in management#

Kahneman and Gary Klein (2009) spelled out when an expert's intuition is reliable: you need a sufficiently regular environment and fast, clear feedback that allows learning. An IT project rarely meets these conditions:

each project is unique: experience transfers poorly;
feedback is late: you learn whether an estimate was good months later;
we decide on estimates, which are uncertain by nature;
we decide in groups and under a hierarchy;
pressure (deadlines, budget, expectations) degrades deliberation.

Management adds a difficulty: judging people on few observations, with personal stakes.

How to read the entries#

Each entry follows the same outline: definition, mechanism, where it occurs, detection signals, avoidance reflexes, then See also and sources. At the top, two markers.

Frequency#

A qualitative assessment, in the context of an IT project or a team:

Level	Gauge	Meaning
Ubiquitous		You meet it on almost every project or every week.
Frequent		Several times per project.
Occasional		A few times a year, in typical situations.
Rare		Only in particular contexts.

This is not a measurement

There is no reliable statistic for how often each bias occurs in IT projects. These levels are a judgment based on the literature and everyday practice of the profession, to be discussed, not quoted as a figure.

AI help#

When AI can help spot or mitigate the bias, the entry shows a second gauge, in blue, and clicking it leads straight to the “With AI” section, where you find what it can do (detect, remedy, mitigate), ready-to-use prompts and precautions.

Level	Gauge	Meaning
Strong		AI does most of the detection or correction work.
Good		A real gain, provided you keep safeguards.
Partial		It helps, but it too falls into the trap: the order and wording of requests matter.
Limited		Little contribution, or a risk of making the bias worse.

AI has its own biases

Language models are not neutral: they tend to approve what the user already thinks (sycophancy), are subject to anchoring and reproduce biases present in their data. The prompts on this site therefore explicitly ask for the opposing view, and put AI on the side of method rather than judgment. A study from Aalto University also suggests that using AI can lead people to overestimate their own performance.

Domains#

The “Where does it occur” badges indicate the precise moments when the bias shows up: estimating, committee, hiring, code review, negotiation… The same bias appears in several domains.

Authors#

When a bias's authors are known, a frame to the right of the definition says who they are: an avatar drawn by the site (its features are inspired by the person when a photo exists, otherwise they are drawn at random; it does not resemble the person), nationality (flag), dates of birth and death. Hover over the avatar to enlarge it.

Links between biases#

Biases combine. The “See also” section distinguishes three relationships: combines with, or reinforces (they feed each other), neighboring biases (variants not to be confused) and counterweights (what neutralizes them).

See the map

All the links between entries are gathered in The map of links between biases: biases grouped by family, filters by family, link type and frequency, search by name.

Three ideas to protect yourself#

Don't count on vigilance. Knowing a bias exists almost never stops it from acting. What works is to change the situation: decide with a procedure, not with goodwill.
Take the outside view. Rather than reasoning about “this project”, look at how similar projects went. It is the most effective countermeasure against optimism.
Make the decision visible and contestable. Independent estimates before discussion, criteria written before the evaluation, a designated “devil's advocate”, a pre-mortem (“imagine the project has failed: why?”), a decision log to compare planned and actual later.

A starting guide: most countermeasures consist of changing the decision procedure, not counting on vigilance.Hover over or tap an element to read its details.

Limits and precautions#

Not everything is established to the same degree. Some widely cited effects have been weakened by difficult replications (this is also debated for Dunning-Kruger). Each entry flags what is solid and what is less so.
Taxonomies differ. Several hundred “biases” have been catalogued, many overlap. The groupings on this site are reading choices, not an official classification.
Don't play inquisitor. Telling a colleague “that's anchoring bias” closes the discussion instead of opening it. Use these entries first for your decisions, then to improve the group's process, rarely to diagnose someone.
Knowing a bias doesn't make you immune. Experts are subject to them, often with more confidence.

Sources and verification#

Tversky & Kahneman (1974), “Judgment under Uncertainty: Heuristics and Biases”, Science — the founding paper.
Kahneman & Klein (2009), “Conditions for intuitive expertise: A failure to disagree”, American Psychologist — when an expert's intuition is reliable.
Gigerenzer & Gaissmaier (2011), “Heuristic Decision Making”, Annual Review of Psychology — the view on the ecological rationality of heuristics.
Klein (2007), “Performing a Project Premortem”, Harvard Business Review — the pre-mortem method.
Kahneman (2011), Thinking, Fast and Slow — a synthesis for the general public.
Sharma et al. (2023), “Towards Understanding Sycophancy in Language Models” — sycophancy in language models.
Aalto University (2025), “AI use makes us overestimate our cognitive performance” — press release on the study by Fernandes et al.