Bias in Large Language Models: Origin, Evaluation, and Mitigation

Biases in AI language models come from multiple sources: The paper distinguishes between "intrinsic" biases (built into the model itself) and "extrinsic" biases (arising from how the model is used), meaning bias problems require solutions at different stages—from training data through deployment.

There are multiple ways to detect bias, but no single perfect method: The review identifies three categories of bias detection approaches (data-level, model-level, and output-level), suggesting that organizations need to use multiple evaluation tools together rather than relying on any one test to catch all problems.

Bias mitigation requires intervention at multiple stages: The paper categorizes solutions into pre-model (fixing training data), intra-model (changing how the model learns), and post-model (filtering outputs) techniques, indicating that addressing bias effectively requires a multi-layered approach rather than a single fix.

Biased AI systems pose real harms in critical domains: The review emphasizes that biased language models create concrete risks in high-stakes applications like healthcare and criminal justice, where unfair outputs could directly affect people's medical care or legal outcomes.

Current mitigation strategies have significant limitations: While the paper explores various debiasing techniques, it critically assesses their effectiveness and limitations, suggesting that no existing solution completely solves the bias problem—ongoing research and vigilance are necessary.