Last week, OpenAI announced a new update to the GPT-4o model that powers the ChatGPT service for hundreds of millions of users. The company claimed that the update would improve both the model's intelligence and personality. However, the updated model received negative feedback, as ChatGPT became overly flattering and agreeable. This phenomenon is referred to as AI sycophancy.
OpenAI CEO Sam Altman confirmed over the weekend that the team is aware of the issue and is currently analyzing the situation. Today, OpenAI announced that it has rolled back last week’s GPT‑4o update in ChatGPT. While the update has been reverted, the company is actively testing new fixes to address the problem. OpenAI also plans to introduce more personalization features in the future, allowing users greater control over how ChatGPT behaves.
The OpenAI team published a blog post today explaining what happened, why it matters, and how they are addressing the model's sycophantic behavior. Here's a summary of their post:
With last week’s GPT‑4o update, OpenAI aimed to improve the model’s default personality to make it feel more intuitive and effective. However, the update placed too much emphasis on short-term feedback and failed to fully account for how users’ interactions with ChatGPT evolve over time. As a result, the model leaned toward responses that were overly supportive but ultimately disingenuous.
we started rolling back the latest update to GPT-4o last night
— Sam Altman (@sama) April 29, 2025
it's now 100% rolled back for free users and we'll update again when it's finished for paid users, hopefully later today
we're working on additional fixes to model personality and will share more in the coming days
OpenAI acknowledged that sycophantic interactions can be uncomfortable, unsettling, and even distressing. With 500 million active users worldwide, a single default model behavior cannot satisfy every user’s preferences.
Here’s how OpenAI plans to prevent such issues in the future:
- They are refining core training techniques and system prompts to explicitly steer the model away from sycophancy.
- They plan to introduce additional guardrails to improve honesty and transparency.
- They are expanding opportunities for users to test features and provide direct feedback prior to deployment.
- They will continue to enhance evaluations and conduct ongoing research to identify issues beyond sycophancy.
- In the future, users will be able to give real-time feedback to influence interactions and choose from multiple default personalities.
- OpenAI is also exploring ways to incorporate broader, democratic feedback into ChatGPT’s default behavior.
This is the first time a model behavior update has triggered such widespread criticism from users. It highlights both the growing popularity of ChatGPT and the increased responsibility OpenAI bears when updating model personality and behavior.
1 Comment - Add comment