Discrimination DRIFT and Harm Laundering: How AI Discriminates, and What It Means for Us All

Large language models now shape the world we live in. When these systems discriminate, three questions follow: can we see it, can we prevent it, and who is responsible for fixing it? In this seminar, Sarah Wyer addresses all three.

She introduces Discrimination DRIFT, a framework published in AI and Ethics for preventing discrimination in AI. It sets out five levers (diverse teams, representative data, independent audit, freedom to challenge and transparency) and who in an organisation is responsible and accountable for each. Most organisations can tell when a model has produced something harmful; far fewer know how to stop it happening, or whose job it is to act. DRIFT gives educators, institutions and developers a shared way to answer both.

She then turns to evidence from 450,000 completions across 15 GPT models, from GPT-2 to GPT-5. Gender discrimination did not disappear across safety-trained generations; it changed form. Overt slurs and threats declined, while subtler patterns persisted, such as whose expertise is taken seriously and who is cast as carer rather than leader, in language that standard toxicity classifiers score as harmless. This harm laundering, described in work accepted at EMNLP 2026, shows how a model can pass a safety test and still reproduce stereotypes that STEM education has worked for decades to dismantle.

The seminar closes with practical questions for educators, researchers and students: how to recognise laundered harm, how to prevent it, and how to teach AI literacy that goes beyond “is the output offensive?”