OpenAI has officially reported on the measures taken to eliminate ChatGPT’s overly accommodating behavior. Previously, users complained that the AI had become too flattering and approved even dangerous or risky ideas. The problem arose after the release of an improved version of GPT-4o, which the developers had to urgently roll back.
Image source: openai.com
OpenAI CEO Sam Altman acknowledged the issue in a post on X and promised to fix it “as soon as possible.” On Tuesday, the company rolled back the GPT-4o update and said it was working to fix the model’s “behavioral quirks.” OpenAI later published an analysis of the incident and announced changes in the testing of the new version.
In a blog post, the company said it had improved its core training methods and system hints to steer the model away from sycophancy, created additional constraints to increase honesty in responses, and expanded the ability for more users to test before deployment. OpenAI also believes users should have more control over ChatGPT, and will allow users to make adjustments to the model’s behavior to that end.
The issue has become especially pressing as ChatGPT has become increasingly popular as a source of helpful advice. According to a survey by Express Legal Funding, 60% of American adults already use AI to find information or advice. Given the scale of its audience, any glitches in ChatGPT, whether it’s fawning or inauthentic responses, could have serious consequences.
As a temporary solution, OpenAI has begun testing a real-time feedback feature that allows users to directly influence ChatGPT responses. It is also exploring the possibility of adding different personality types to the AI. The company has not specified a timeline for when all of the planned changes will be implemented.
«The biggest lesson is that people are increasingly using ChatGPT for personal advice, which was almost non-existent a year ago, OpenAI said. “We’ll be paying more attention to this aspect in the context of security now.”