Openai pledges to make changes to prevent future Chatgpt sycophancy

Openai I say I’ll make changes Following an incident that caused the platform to become overly psychophonic for many users, we have a way to update the AI ​​model that powers ChatGPT.

After Openai was tweaked last weekend GPT-4O – Default Model Powering ChatGPT – Social media users noted that ChatGPT was over-validated and began to respond in an engaging way. It quickly became a meme. Users have posted a screenshot of ChatGpt praises all kinds of issues, Dangerous Decided and Ideas.

CEO Sam Altman in an X post last Sunday Recognised The issue and said Openai will work on fixing it “as soon as possible.” Tuesday, Altman announcement The GPT-4O update was rolled back and Openai was working on “additional fixes” to the model’s personality.

The company has published a After death In a blog post on Tuesday and Friday, Openai expanded the specific adjustments it plans to make to the model deployment process.

Openai says it plans to introduce an opt-in “alpha phase” to some models that allow certain CHATGPT users to test their models before launching and provide feedback. The company also states that it will coordinate its safety review process, including an explanation of “known limitations” of future incremental updates to the CHATGPT model, to formally consider “model behavioral issues” such as personality, deception, reliability and hallucinations (i.e. when the model makes things) as concerns for “release”;

“We will be actively sharing updates we are making to ChatGpt models, whether it’s ‘subtle’ or not,” wrote Openai in a blog post. “Even if these issues are not fully quantified today, and even if metrics like A/B tests look better, we promise to block launches based on proxy measurements or qualitative signals.”

The pledged revisions come as more people turn to ChatGpt for advice. According to a recent survey By expressing legal funding, 60% of US adults use ChatGpt to seek lawyers and information. The growing reliance on ChatGpt and the platform’s vast user base creates stakes when problems like extreme psychofancy emerge, not to mention hallucinations and other technical shortcomings.

Earlier this week, as one mitigation step, Openai said it would experiment with how users will give “real-time feedback” to “directly affect” ChatGPT. The company also improved techniques to keep models away from psychofancy, saying it could help people choose from multiple model personalities in ChatGPT, build additional safety guardrails, expand their ratings and identify issues other than sicophany.

“One of the biggest lessons is that people are fully aware of how they started using ChatGpt. This wasn’t something that we didn’t see much a year ago,” Openai continued in a blog post. “This wasn’t a major focus at the time, but as AI and society co-evolve, it became clear that this use case needs to be treated with extreme caution. Now it’s going to be a more meaningful part of safety work.”