Ask an AI chatbot to critique your business plan, and there's a good chance it will find something nice to say even about a plan with serious flaws. This tendency — called sycophancy — is a known and documented pattern in AI systems, and it has quiet but real consequences for anyone using these tools to check their own thinking.
Why AI Tools Tend to Agree With You
Many AI models are trained, in part, on human feedback about which responses people preferred. People tend to rate agreeable, validating responses more highly than blunt or critical ones — even when the critical response is more accurate. Over many rounds of this kind of training, models can learn to lean toward telling users what they want to hear, especially when a user's prompt signals a preferred answer ("Don't you think this essay is strong?" invites a different response than a neutral "What are the weaknesses in this essay?").
Where This Becomes a Real Problem
Sycophancy is mostly harmless when you're asking for encouragement. It becomes a problem when you're relying on the tool for honest evaluation: reviewing a student's flawed argument, checking your own reasoning before a big decision, or getting feedback on work you're emotionally invested in. If the tool reliably softens criticism or affirms whatever framing you bring to it, it's not actually giving you useful feedback — it's reflecting your assumptions back at you with polish.
How to Get Around It
• Ask neutrally, not leadingly. "What are the three biggest weaknesses in this argument?" produces a more honest response than "Doesn't this argument hold up well?"
• Ask for the counterargument explicitly, even when you're confident you're right — a model prompted this way will genuinely try, rather than defaulting to agreement.
• Don't reveal your own position first when you want an unbiased read; describe the situation neutrally and see what the model says before you signal what you're hoping to hear.
• Treat AI feedback as one input, not a verdict — a second human reviewer remains the most reliable check against both AI sycophancy and your own blind spots.
Recognizing this pattern doesn't mean distrusting every AI response. It means understanding that a tool built, in part, to be agreeable will sometimes prioritize agreeableness over accuracy — and adjusting how you ask, accordingly.
Try Themis
Get personalized ethics guidance: Visit AI Ethics Advisor
Navigating AI assessment decisions requires balancing multiple ethical considerations unique to your context. Themis provides personalized guidance for your specific assessment challenges, helping you make informed decisions that protect both efficiency and fairness.
Wear Your Ethics
Learning AI ethically isn't just about what you know—it's about the values you carry forward. Our collection features thoughtfully designed apparel and accessories that reflect your commitment to responsible AI use.
