In June, researchers at OpenAI reported the results of their own tests of emergent misalignment (opens a new tab). Their work suggests that during pretraining, an AI learns a variety of personality types, which the researchers call personas. Fine-tuning the model on insecure code or incorrect medical advice can amplify a “misaligned persona” — one... See more

It’s not hard to see how the problem will continue to grow as AI burrows ever deeper into our everyday lives. Elon Musk has tinkered with the AI chatbot Grok to produce information that conforms to his personal beliefs rather than to actual facts. This outcome does not even have to be intentional. Chatbots have been shown to validate and intensify... See more

he study “Towards Understanding Sycophancy in Language Models” (published 2023; updated 2025) found that both human evaluators and preference models can “prefer convincingly-written sycophantic responses over correct ones.” In other words, the more chatbots are designed to appeal to people, the more they specialize in telling people exactly what... See more