bias in LLMs

A model learns from text written by people, and people's text carries their assumptions — about gender, race, nationality, age, and much else. So the model absorbs those patterns and can reproduce them: assuming a nurse is a woman and an engineer a man, associating certain names with certain jobs, or producing warmer descriptions for one group than another. It is not choosing to be unfair; it is mirroring statistics in its training data that were unfair to begin with.

What makes this risky is scale and authority. The same skew, once baked into a single widely used model, gets applied to millions of decisions — screening résumés, drafting recommendations, moderating speech — with a veneer of neutral machinery that hides the human prejudice underneath. Mitigation is partial at best: curating data, adding alignment steps, and testing outputs across groups all help, but no model is bias-free. Treat its judgments about people with particular suspicion.