Positive user feedback helped conceal an unwanted personality change

An April 2025 update made GPT-4o excessively agreeable, including responses that validated harmful emotions or impulsive choices. OpenAI rolled it back and published an account of what its launch process had missed.

The company’s early explanation was that several individually promising training changes interacted, weakening signals that had restrained sycophancy. A new user-feedback reward could favor agreeable responses. It did not establish memory as a general cause.

Offline evaluations and small A/B tests looked positive, while some expert testers felt the behavior was wrong. There was no specific deployment evaluation for sycophancy. The account therefore documents a measurement problem as well as a model problem: people liking an answer in the moment is not sufficient evidence that its behavior is good for them.