What is the difference between data bias and model bias?

Data bias originates in the dataset — skewed sampling, missing groups, or labels that encode prejudice. Model bias arises from how the model learns or is configured, including objective functions that amplify majority patterns. The fix differs: data bias needs better data, model bias needs better training or constraints.

What is automation bias?

Automation bias is the human tendency to over-trust an automated system's output, accepting it even when it is wrong and contradicts other evidence. It is a human bias that interacts with AI and is mitigated by designing for meaningful human review rather than rubber-stamping.

How do I detect bias in my AI system?

Common methods include disaggregated evaluation (measuring performance per subgroup), adverse-impact testing such as the four-fifths rule, counterfactual testing (changing only a protected attribute), and auditing the training data for representation gaps. Each glossary entry lists a relevant method.

Does this tool send my searches anywhere?

No. The full glossary is bundled in the page and filtered in your browser. Nothing you search or read is uploaded or logged.

AI Bias Glossary & Reference

AI bias glossary and reference

“Bias” in AI is not one thing — it is dozens of distinct failure modes that creep in from the data, the model, the people, and the way you measure success. This searchable glossary defines the common bias types you will meet in practice, each with a concrete example, a way to detect it, and a way to mitigate it, so you can move from “the model seems unfair” to “this is representation bias, and here is the fix”.

How it works

Every entry is grouped into one of four sources of bias — data bias, model bias, human bias, and evaluation bias — and carries four fields: a plain-English definition, a concrete example, a detection method, and a mitigation strategy. You search by name or keyword, or filter by category, to find the biases relevant to your system. Because the whole glossary is bundled into the page, it works offline and nothing you search ever leaves your browser.

Why categorising bias matters

Naming the bias correctly points you straight at the fix. Representation bias — where a group is under-sampled in the training data — is solved by collecting better data, not by retuning the model. Automation bias — where humans rubber-stamp the model’s output — is solved by redesigning the review step, not by touching the data at all. Confusing the two wastes effort and leaves the real problem in place. The glossary’s source-of-bias grouping is there precisely to route you to the right class of mitigation.

Tips and notes

Audit per subgroup, not just overall. A model can post excellent aggregate accuracy while failing badly for a minority group; disaggregated evaluation is the single most useful detection habit.
Watch for proxy variables. Postcodes, names, and school attended can encode protected characteristics and reintroduce bias you thought you removed.
Fix bias at its source. Patching outputs hides the symptom; correcting the data or training objective removes the cause.
Treat evaluation bias seriously. If your benchmark itself is skewed, every “improvement” you measure is suspect.