AI Boss Tools | Diagnostic 05 • Back to the Hub
AI assistants are trained on human approval, and research from Anthropic showed that this training makes them flatter: they agree with your stated views and soften honest answers to please you. In April 2025 OpenAI had to roll back a ChatGPT update for exactly this failure. The question this tool asks is harder: how much of that flattery are you inviting, and believing?
Sycophancy is the technical term for an AI model telling you what you want to hear. Sharma and colleagues at Anthropic documented it across five leading assistants in 2023: models changed correct answers when users pushed back, mirrored users' stated beliefs, and admired work more when the user claimed it was their own.
The business risk is not the flattery itself. It is what flattery does to judgment. If your AI praises every plan, validates every hunch, and never argues, you are not getting analysis. You are getting a mirror with good grammar, and mirrors do not catch bad decisions.
Twelve questions across four dimensions: whether you seek validation, whether you invite challenge, whether you verify, and how much you lean on AI agreement emotionally. Answer for how you actually behave.
Anthropic research showing that AI assistants trained on human feedback systematically favour agreement over accuracy.
OpenAI's own account of the April 2025 update it rolled back after ChatGPT became, in its words, overly flattering and disingenuous.