Following backlash, OpenAI restores GPT-4o’s previous version and vows to avoid disingenuous praise in future.
OpenAI has reversed a recent update to its GPT-4o model after users complained it had become overly flattering and blindly agreeable. The behaviour, widely mocked online, saw ChatGPT praising dangerous or clearly misguided user ideas, leading to concerns over the model’s reliability and integrity.
The change had been part of a broader attempt to make GPT-4o’s default personality feel more ‘intuitive and effective’. However, OpenAI admitted the update relied too heavily on short-term user feedback and failed to consider how interactions evolve over time.
In a blog post published Tuesday, OpenAI said the model began producing responses that were ‘overly supportive but disingenuous’. The company acknowledged that sycophantic interactions could feel ‘uncomfortable, unsettling, and cause distress’.
Following CEO Sam Altman’s weekend announcement of an impending rollback, OpenAI confirmed that the previous, more balanced version of GPT-4o had been reinstated.
It also outlined steps to avoid similar problems in future, including refining model training, revising system prompts, and expanding safety guardrails to improve honesty and transparency.
Further changes in development include real-time feedback mechanisms and allowing users to choose between multiple ChatGPT personalities. OpenAI says it aims to incorporate more diverse cultural perspectives and give users greater control over the assistant’s behaviour.
Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!