LLMs now get tricked by irrelevant details in high-stakes decisions — a method cuts the bias by 84%

When AI models evaluate teachers or make other consequential decisions, they get thrown off by irrelevant details about who the person is — race, gender, experience level — even when those details shouldn't matter. A new training method called Debiasing-DPO teaches models to ignore these spurious signals while maintaining accuracy, reducing bias by 84% on average across several large language models.

AI is moving into decisions that affect people's careers and livelihoods. Right now, these models are sensitive to details that have nothing to do with the actual judgment — a teacher's gender can shift how an AI scores their classroom performance by up to 1.5 points on a 7-point scale. The problem is that the models are already quite capable and accurate, so their biases hide inside what looks like competence. This paper shows those biases are fixable without sacrificing the accuracy that made people trust the model in the first place. If this scales to real deployment, it means the same AI systems can be both more fair and more useful.