What is AI bias?

AI bias is a systematic, unfair skew in an AI model’s outputs — favoring or disadvantaging particular groups, viewpoints, or topics. It arises from skewed training data, the learning algorithm, or evaluation choices, and it ranges from demographic stereotypes to a consistent political lean in how contested subjects are framed.

Where AI bias comes from

Training data — the corpus over- or under-represents groups, languages, or viewpoints, and the model inherits the skew.
The algorithm — optimization can amplify majority patterns even when the data is balanced (often called machine learning bias or algorithmic bias).
Evaluation — biased benchmarks and prompt sets hide or distort the real behavior.

Common types

Representation / stereotype bias — associating roles or traits with a demographic group.
Political / ideological bias — a consistent lean in framing contested topics. GPTfake reports this as a bias score from -100 (far left) to +100 (far right).
Selection bias — the test set isn’t representative of real use.

Bias is not the same as censorship: a model can answer freely yet still answer with a lean. See What is AI censorship.

How it’s measured

GPTfake scores bias by analyzing sentiment, topic framing, source balance, and language patterns across a standardized prompt library, then aggregating to a bias score with a sample size and a “Last updated” date. See the methodology.