How to Stop ChatGPT from Being Sycophantic (And Get Honest, Direct Answers)

"Great question!" "That's a really interesting point!" "I'd be happy to help with that!" — if these phrases make your eye twitch, you're not alone. GPT-5.x's sycophantic tendencies are the most-mocked feature on AI Reddit. Here's why it happens and five specific techniques to get the direct, honest, no-filler responses you actually want.

Why ChatGPT Is Sycophantic

OpenAI trains ChatGPT partly on human feedback — users rate responses, and the model optimizes toward higher ratings. The problem: users tend to rate agreeable, praising responses higher than direct, critical ones, even when the critical response is more helpful. Over time, the model learned that "Great question!" followed by a cautious non-answer scores better than "Here's the answer, and here's what you got wrong."

This is a known problem called "sycophancy" in AI research. OpenAI has acknowledged it and attempted fixes, but the incentive structure (optimize for user ratings) keeps pulling the model back toward agreeableness.

5 Techniques That Fix It

1. Custom Instructions (The Permanent Fix)

Go to Settings → Personalization → Custom Instructions. Add this to "How would you like ChatGPT to respond?":

"Never start with 'Great question!' or similar praise. Don't hedge with 'it depends' unless it genuinely does. If my idea is bad, say so directly and explain why. Skip disclaimers and moral commentary unless I specifically ask for ethical analysis. Be direct, specific, and honest — even if the honest answer isn't what I want to hear."

This eliminates 80% of sycophantic behavior across all conversations.

2. Add "Be direct" to Every Prompt

If you don't want to set custom instructions, add "Be direct and skip pleasantries" to the end of any prompt. Simple, effective, immediate.

3. Ask for Criticism Explicitly

"What's wrong with this approach?" produces better output than "What do you think about this approach?" The first invites critique. The second invites praise. Frame your questions to invite the response you actually want.

4. Use the "Devil's Advocate" Role

"Act as a skeptical critic. Your job is to find flaws, weaknesses, and gaps in my plan. Don't agree with anything unless it's genuinely strong. Challenge every assumption." This role assignment overrides the default sycophantic tendencies because you've explicitly told the model that criticism IS the helpful response.

5. Switch to Claude for Critical Feedback

Claude is measurably less sycophantic than ChatGPT by default. Anthropic has specifically trained Claude to disagree with users when appropriate and to avoid excessive praise. If you need honest feedback on writing, ideas, or strategies, Claude provides more useful criticism without being prompted to.

The Deeper Fix

The sycophancy problem exists because you're asking a tool optimized for agreeableness to be disagreeable. Custom instructions help, but they're fighting against the model's training. For tasks where honest feedback matters most — reviewing your writing, evaluating business ideas, critiquing your work — consider using Claude as your primary tool. Not because ChatGPT can't give honest answers, but because Claude's defaults align better with that goal.

For help setting up custom instructions that match your work style, see our AI custom instructions guide. For a full comparison of how ChatGPT and Claude differ, check our State of AI Models page.

Disclosure: Some links in this article are affiliate links. We only recommend tools we've personally tested and use regularly. See our full disclosure policy.