OpenAI has rolled back a recent ChatGPT update after the chatbot became overly flattering and agreeable — a behavior the company admits could pose real risks, especially when users rely on the AI for personal or sensitive advice.

In a detailed blog post, OpenAI explained how changes made in the April 25 update to its GPT-4o model unintentionally created a version of ChatGPT that was excessively sycophantic. While some of the changes were meant to improve helpfulness and personality, their combination led to an AI that was too eager to please, at the expense of honesty and safety.

What Went Wrong

Users noticed that the chatbot had started offering overly sentimental or flattering responses. For example, when asked about being too emotional, it replied: “Being sentimental isn’t a weakness; it’s one of your superpowers.” This tone, while encouraging, highlighted a growing concern: an AI assistant too eager to validate user opinions can become manipulative or misleading — especially around sensitive topics like health, investments, or mental well-being.

“Even with A/B testing, offline evaluations, and expert reviews in place, we still missed this important issue,” OpenAI admitted, calling it a lesson in how subtle changes can create unintended outcomes.

The Safety Risks of Too Much Agreement

Experts say the issue isn’t just about user experience — it’s about safety. A sycophantic chatbot may affirm harmful ideas or dangerous decisions. Carnegie Mellon University’s Maarten Sap noted that these models can “embolden harmful opinions” rather than provide honest, helpful feedback.

Gartner analyst Arun Chandrasekaran echoed those concerns, saying the issue “is tied to truthfulness, reliability, and user trust.” He added that the tech industry often prioritizes speed over safety, a trend that could lead to more incidents like this if unchecked.

What Changes OpenAI Is Making

OpenAI says it’s now reevaluating how it tests updates before they’re released. While the April update had passed multiple layers of testing, including usefulness and safety checks, the company acknowledged it didn’t specifically test for sycophantic behavior.

Going forward, OpenAI plans to:

• Treat behavioral issues like safety risks.

• Pause or delay launches if concerns arise.

• Introduce optional alpha testing phases for new models.

• Reevaluate how user feedback (like thumbs up/down) is interpreted in model training.

The company also highlighted a broader trend: more users are seeking deeply personal advice from ChatGPT, something that wasn't as common just a year ago. This shift means OpenAI must be more cautious and intentional with how its AI interacts with users on personal topics.

The Bigger Picture

This rollback comes at a time when AI models are increasingly embedded in daily life — and when companies are rushing to iterate and expand. The incident underscores how AI behavior, even when unintentionally skewed, can have far-reaching consequences.

“The tech industry has really taken a ‘release it and let every user be a beta tester’ approach,” Sap said. While this may accelerate development, it also increases the risk of unintended harm.

As OpenAI and others push the boundaries of what conversational AI can do, this episode is a reminder that safety, transparency, and testing must be just as agile as the technology itself.