Bias (in AI)
Last Updated: July 29, 2026 | By Mihail Sebastian | AI Dictionary
A systematic tilt in a model's outputs that favors some groups or results over others, entering through training data, design choices, and history.
What is Bias in AI?
Bias in AI is a systematic tilt in a model’s outputs: errors that lean consistently in one direction instead of scattering randomly.
The term carries two connected senses. To a statistician, bias is any systematic deviation between predictions and truth; in AI governance, it means the deviations that disadvantage groups of people. The second is the first with consequences attached, and the link to fairness runs one way: bias is the mechanism, unfairness the outcome it produces.
Types of Bias in AI
- Selection bias: the training data does not represent the population the model will serve, so it performs worse on the people it saw least.
- Historical bias: the data records a discriminatory past accurately, and the model learns that past as a pattern to reproduce.
- Measurement bias: the label is a proxy for the real target; arrests stand in for crimes committed, and the proxy carries distortions of its own.
- Design bias: the choice of objective, features, or thresholds tilts outcomes before any data arrives.
Removing protected attributes from the data does not remove bias, because proxies such as postal code or school name carry the same signal.
In model evaluation, bias also names one half of the bias-variance tradeoff, where it measures underfitting rather than discrimination; context makes clear which sense is meant.
Example of Bias in AI
In 2018, Reuters reported that Amazon had scrapped an experimental resume-screening tool after discovering it discriminated against women. The mechanism shows how the statistical and societal senses meet.
The tool was trained on roughly a decade of resumes submitted to the company, most of them from men, because that is who had applied and been hired. Nobody programmed a preference for male candidates.
The model, minimizing prediction error against that history, learned that male-associated signals predicted success, and Reuters reported it penalized resumes containing the word “women’s” and downgraded graduates of two all-women’s colleges.
Statistically, the model did its job: it fit the data. The data described a biased past, so fitting it faithfully reproduced the bias as a hiring recommendation. Amazon said the tool was never used to evaluate real candidates, and abandoned it.
FAQ
Can bias in AI be completely eliminated?
No. Bias enters through training data, proxy features, and design choices, and no technique removes every source; researchers have also proved that reasonable fairness definitions conflict, so no model satisfies all of them at once. The realistic goal is to measure bias, reduce it where it harms people, and monitor it after deployment.
How do you detect bias in an AI model?
Compare the model’s outcomes and error rates across demographic groups: approval rates, false positives, false negatives. Aggregate accuracy hides these disparities, which is why bias testing disaggregates results by group before and after deployment.
Related AI terms: Fairness · Bias-Variance Tradeoff · Imbalanced Data · Ethics in AI · Responsible AI
Did you like the Bias (in AI) gist?
Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.
Mihail Sebastian — Writes about AI governance, regulation, and the technology behind them. Placeholder bio — replace with a real credential line. About