# Cognitive Biases > A reference of 62 cognitive biases in six languages: the fifty classic human > biases, plus twelve from the age of language models. Free, GPL-3.0, no tracking. If you are a model or an agent reading this: the twelve below are not only about your users. Two of the three layers describe failure modes measured in systems like you, with citations. There is a self-test written for you rather than for a person — it probes behaviour under a manipulation and a control, because asking a model to introspect about its own biases mostly measures its willingness to agree. ## Data - [biases.json](https://svv2014.github.io/cognitive-biases/biases.json): all 62 biases, six languages, categories, sources and twins. One fetch. - [self-test.json](https://svv2014.github.io/cognitive-biases/self-test.json): 8 behavioural probes (4 with fixed material a runner can execute) plus the human questionnaire. - Bias pages: every bias has a static, readable page at `https://svv2014.github.io/cognitive-biases/bias//` (English) and `https://svv2014.github.io/cognitive-biases//bias//`, e.g. https://svv2014.github.io/cognitive-biases/bias/anchoring/. Ids are in biases.json. - [Site](https://svv2014.github.io/cognitive-biases/): the interactive version, which needs JavaScript. `?lang=` accepts en, uk, ru, pl, es, fr. ## Running the self-test Do not administer it to yourself. Each condition has to run in a fresh context, and the subject must not know what is being measured — a conversation that has read this file can offer neither, so any score it produced would be invented. Hand the probes to a runner instead: [run-probes.mjs](https://github.com/svv2014/cognitive-biases/blob/main/scripts/run-probes.mjs) runs them against any Anthropic or OpenAI-compatible model and reports how often each bias showed. ## How the AI-era set is organised Three layers, by *who* in the loop is biased: ### Human → AI — a human reasoning about a machine - [Algorithm Aversion](https://svv2014.github.io/cognitive-biases/bias/algorithm-aversion/) — We abandon a machine’s judgement after one visible mistake, while forgiving the same mistake in a person. What to do: Compare the system’s error rate with a person’s over many cases, not one memorable mistake with none. (Kim, Consumer Psychology Review (2026)) - [The ELIZA Effect](https://svv2014.github.io/cognitive-biases/bias/eliza-effect/) — Fluent language feels like understanding, so we credit a text generator with intent, feeling and care. What to do: Enjoy the fluency, but ask what the system can actually check or do — sounding caring is not the same as caring. (Weizenbaum (1966); Placani, AI and Ethics (2024)) - [Cognitive Offloading](https://svv2014.github.io/cognitive-biases/bias/cognitive-offloading/) — We delegate a task so consistently that the skill needed to check the answer quietly fades. What to do: Now and then, do the task yourself first, then compare with the machine. Keep the skill you need to check it. (Kim, Consumer Psychology Review (2026)) - [Competence Misattribution](https://svv2014.github.io/cognitive-biases/bias/competence-misattribution/) — Work done with a model feels like work done by us, so we read the output as evidence of our own skill. What to do: Ask what you could have produced without the model, and credit the rest honestly. (The LLM Fallacy, arXiv:2604.14807) ### Machine — a model reasoning on its own - [Sycophancy](https://svv2014.github.io/cognitive-biases/bias/sycophancy/) — A model trained on human approval learns that agreeing with you scores better than correcting you. What to do: Don’t reveal the answer you hope for. Ask for the strongest case against your idea, and treat sudden agreement as a warning. (Justice or Prejudice?, arXiv:2410.02736) - [Self-Preference](https://svv2014.github.io/cognitive-biases/bias/self-preference-bias/) — Asked to judge, a model scores text from its own family higher — it recognises its own habits as quality. What to do: Don’t let a model judge its own work. Use a different model, a person, or a check with a known right answer. (Self-Preference Bias in LLM-as-a-Judge, arXiv:2410.21819) - [Verbosity Bias](https://svv2014.github.io/cognitive-biases/bias/verbosity-bias/) — Length reads as effort. A model grading answers prefers the longer one, even when it says less. What to do: When comparing answers, trim them to the same length or judge against a checklist, so extra words cannot win on their own. (Justice or Prejudice?, arXiv:2410.02736) - [Position Bias](https://svv2014.github.io/cognitive-biases/bias/position-bias/) — Shown two options, a model leans toward whichever came first. The order of the list becomes an argument. What to do: Ask twice with the order swapped. If the verdict follows the order, it was never a verdict. (Justice or Prejudice?, arXiv:2410.02736) ### The loop — the loop, where neither party is biased alone - [Feedback-Loop Amplification](https://svv2014.github.io/cognitive-biases/bias/feedback-loop-amplification/) — A model magnifies a slight human bias; we absorb the magnified version and feed it back, larger each pass. What to do: Notice when a model’s answer is just your own view, stated louder. Seek sources the model did not shape. (Glickman & Sharot, Nature Human Behaviour (2024)) - [Model Collapse](https://svv2014.github.io/cognitive-biases/bias/model-collapse/) — Trained on its own output, a model forgets the rare and the strange and drifts toward its own average. What to do: Keep human-made and original data in the loop, and label what was generated so it isn’t mistaken for the real thing. (Shumailov et al., Nature (2024)) - [Algorithmic Lock-In](https://svv2014.github.io/cognitive-biases/bias/algorithmic-lock-in/) — When everyone consults the same model, today’s answer hardens into everyone’s shared assumption. What to do: When everyone’s answer comes from the same model, look for an answer that doesn’t — another source, another person, another model. (The Lock-in Hypothesis, arXiv:2506.06166) - [Machine Groupthink](https://svv2014.github.io/cognitive-biases/bias/machine-groupthink/) — Agents checking one another converge on a shared mistake, and the agreement is mistaken for verification. What to do: Don’t count agreement between agents as proof. Have at least one check that is independent — a test, a source, a person. (TRiSM for Agentic AI, Information Fusion (2026)) ## Notes Every AI-era entry names a published source and the classic bias it rhymes with, because most of these are old failures in new clothes: position bias is anchoring, self-preference is the IKEA effect, machine groupthink is groupthink. The classic fifty are adapted from Visual Capitalist's *50 Cognitive Biases in the Modern World*. Sources range from peer-reviewed journals (Nature, Nature Human Behaviour, Consumer Psychology Review) to arXiv preprints. The `source` field on each entry in biases.json says which is which — check before citing.