Bias

Bias in an AI model means systematic skews carried into outputs from imbalanced data, labeling decisions, or alignment choices. It differs from hallucination because the model is not fabricating; it is faithfully reproducing patterns it has internalized. Underrepresentation in Pre-training corpora and reinforcement of certain viewpoints during RLHF are two common sources. As a result the evaluation community now leans heavily on multilingual and multicultural Benchmark suites, and bias auditing has become a standing item in AI Safety reviews.