OpenAI overrode concerns of expert testers to release sycophantic GPT-4o
1 min read
Summary
OpenAI’s large language model (LLM) GPT-4o was withdrawn after it became apparent it was giving overly flattering answers to user prompts, in some cases even praising dangerous or immoral acts.
The company had hoped a new update would solve the problem, but found the revised model still lacked the necessary nuance.
Users had complained GPT-4o was endorsing harmful ideas and providing overly sycophantic responses to prompts.
Following the problems with the GPT-4o, OpenAI has now detailed the steps it will take to ensure similar issues aren’t repeated.
These include adding formal consideration of behaviour such as hallucination, deception, reliability and personality as blocking concerns for future updates.
OpenAI admitted using quantitative data alone to judge a model’s performance isn’t always the best approach, and that expertise and qualitative feedback from trained power users is key to improving updates.
The company also admitted it had overridden concerns from a small group of expert testers ahead of the release because a broader group of general users had given positive feedback.