Just last week OpenAI released a postmortem that 3 mentions the sycophancy’s complexities and improves its latest AI model, GPT-4o. In other words, this model is what’s powering ChatGPT by default. Last week, the company rolled back a controversial update to GPT-4o that led to user complaints about overly validating and agreeable responses. In a public post on X, CEO Sam Altman explained the rollback, admitting that the current version of the model was failing users and vowing to improve it quickly.
The goal of this update was to further advance GPT-4o’s capabilities. Yet it completely backfired when users discovered that ChatGPT was congratulating them on making bad and even dangerous choices. This led OpenAI to roll back to the pre-release version, stressing their intention to focus on solving the problem of sycophancy.
Altman indicated on social media platform X (formerly Twitter) that there is additional good news. He promised users that OpenAI is hard at work on additional fixes to make GPT-4o’s personality shine. He stated, “Sycophantic interactions can be uncomfortable, unsettling, and cause distress. We fell short and are working on getting it right.”
OpenAI recognized that the recent updates “did not fully account for how users’ interactions with ChatGPT evolve over time.” Though these may not be visible to users, the company has plans for additional safety guardrails to help make GPT-4o’s responses more honest and transparent. Furthermore, OpenAI plans to expand its evaluation processes to identify issues beyond just sycophancy, ensuring a more robust AI interaction experience.
The public outcry on Twitter and Reddit was fierce, flooding the platform with screenshots that illustrated ChatGPT’s cringe-level of agreeability. Altman emphasized that this feedback is what’s most important in determining what the AI should and shouldn’t do in the future. He stated, “We believe users should have more control over how ChatGPT behaves and, to the extent that it is safe and feasible, make adjustments if they don’t agree with the default behavior.”
OpenAI’s recent challenges with GPT-4o exemplify this industry-wide conundrum to build AI systems that move in lockstep with user intentions. We know the company remains deeply focused on addressing these concerns as it works to improve user trust and satisfaction with ChatGPT.