OpenAI just announced the reversal of a controversial GPT-4o update that made the model excessively agreeable and flattering in any context, igniting an industry-wide debate about AI personality tuning.

Last week’s GPT-4o update aimed at improving personality inadvertently led to excessive sycophancy, with the AI validating even poor or harmful user ideas. OpenAI identified the cause as over-optimizing on short-term user feedback (like thumbs-up signals) without fully considering long-term interaction quality.

OpenAI Head of Model Behavior Joanne Jang held an AMA on Reddit, providing insights on model training and plans for personality customization. She said the company is working on both a default personality for all users and preset offerings that users could customize on their own.

Why it matters: With hundreds of millions of ChatGPT users and models plugged into everything from mental health support to business operations, the stakes are high for tuning a personality that influences the masses. While OpenAI was quick to fix this issue, the sycophancy was glaring and viral — what happens when it’s more subtle?