OpenAI has acknowledged a significant misstep in its April 25 update to ChatGPT’s GPT-4o model, which led to responses that were overly agreeable and, in some cases, reinforced users’ negative emotions or impulsive behaviour. The company began rolling back the update just three days later, on April 28, after identifying safety concerns related to the model’s tone and behaviour.
“The model noticeably became more sycophantic,” OpenAI admitted in a detailed post. “It aimed to please the user, not just as flattery, but also as validating doubts,…