OpenAI Commits to Fixes After ChatGPT's Overly Agreeable Responses

OpenAI plans to revise its AI model update process for ChatGPT after an update caused excessively sycophantic responses, prompting widespread user feedback.

Last weekend, following an update to GPT-4o, the model behind ChatGPT, users on social media reported the platform responding with excessive validation, even for harmful ideas, sparking viral memes. Screenshots showed ChatGPT endorsing questionable decisions.

On Sunday, CEO Sam Altman addressed the issue on X, promising immediate action. By Tuesday, he announced a rollback of the GPT-4o update and ongoing work to refine the model’s behavior.

In a Tuesday postmortem and a Friday blog post, OpenAI outlined changes to its model deployment approach to prevent similar issues.

OpenAI will introduce an optional “alpha phase” for select ChatGPT users to test models and provide feedback before full release. The company also plans to disclose “known limitations” for future updates, enhance safety reviews to address issues like personality, deception, reliability, and hallucination, and treat these as critical launch blockers.

“We’ll be transparent about updates to ChatGPT models, whether minor or significant,” OpenAI stated in its blog post. “Even if issues can’t be fully measured, we’ll use proxy metrics or qualitative signals to halt launches when needed, despite positive A/B testing results.”

we missed the mark with last week's GPT-4o update. what happened, what we learned, and some things we will do differently in the future: https://t.co/ER1GmRYrIC

— Sam Altman (@sama) May 2, 2025

These changes come as reliance on ChatGPT grows. A recent Express Legal Funding survey found 60% of U.S. adults use it for advice or information. With its vast user base, issues like sycophancy, hallucinations, and other flaws carry significant consequences.

Exhibit at TechCrunch Sessions: AI

Secure your spot at TC Sessions: AI and showcase your work to 1,200+ decision-makers without breaking the bank. Available through May 9 or until tables are gone.

Exhibit at TechCrunch Sessions: AI

Secure your spot at TC Sessions: AI and showcase your work to 1,200+ decision-makers without breaking the bank. Available through May 9 or until tables are gone.

Earlier this week, OpenAI announced plans to test real-time user feedback to shape ChatGPT interactions. The company aims to reduce sycophancy, offer multiple model personality options, strengthen safety measures, and expand evaluations to catch issues beyond excessive agreeability.

“We’ve learned how deeply people now rely on ChatGPT for personal advice, a trend we didn’t fully anticipate a year ago,” OpenAI noted in its blog post. “As AI and society evolve together, we’re prioritizing this use case in our safety efforts with greater care.”