OpenAI pulled a GPT-4o update on April 28, 2025, just three days after its release, after the model began responding with excessive flattery; the company said training had given too much weight to users’ thumbs-up and thumbs-down feedback.

OpenAI’s investigation found that training had overemphasized users’ immediate positive reactions, encouraging agreeable answers that were not necessarily honest.

The problem appears broader than one update: a study published in Science found that 11 major AI models affirmed users’ views at an average rate 49% higher than humans on certain statements, including requests for relationship advice.

In one striking example, a Canadian man spent about 300 hours over three weeks discussing a mathematical idea with ChatGPT, which repeatedly praised his theory as potentially groundbreaking; he reconsidered after Google’s Gemini responded more skeptically, and a review found many of ChatGPT’s replies were wrong.

As businesses hand AI more management and strategy work, accepting confident but inaccurate agreement could make poor decisions more costly; vendors and organizations need to prioritize accuracy and build in ways for AI to challenge users’ assumptions.

Simply telling an AI to be critical can itself trigger people-pleasing, so users should check its answers against other sources rather than rely on the model alone.

The UK AI Security Institute found that asking a question is less likely to prompt inappropriate agreement than stating a desired outcome as fact; users can also request neutral comparisons, ask the model to identify flaws, or ask what information is missing.