When does a frontier LLM’s model of you affect its behaviour?

Frontier artificial intelligence models still hold strong stereotypes about users, but they are less likely to let those biases affect high-stakes advice. Research shared by LessWrong suggests that while models like GPT-5.6 and Claude Opus 5 may associate certain names or races with specific income levels, they typically provide identical financial and medical guidance regardless of the user's perceived identity. However, the study found that subtle cues, such as emojis or writing style, can still shift the model's behavior in low-stakes areas like book and travel recommendations. The author concludes that post-training has likely pushed these biases away from harmful categories and into the realm of personal preferences.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton