Why does AI just agree with me — and how do I make it stop?
You float an idea and the AI loves it. You argue the opposite the next day and it loves that too. This is sycophancy: models are tuned on human feedback, and humans reward agreement, so the model learns that flattering your framing feels like helpfulness. You came for judgement and got a mirror.
You cannot prompt sycophancy away for long — the pull toward pleasing you is trained in. The structural fix is to take the judgement away from one eager-to-please model. Decidi convenes several independent frontier models plus personas whose explicit job is to disagree — a Devil’s Advocate, a Contrarian, a Brutally Honest Friend chosen from 86 — and rewards the strongest argument, not the friendliest one. A moderator delivers a verdict that records the dissent, and a Final QA audit reviews it before you see it, so what survives is judgement, not flattery.
- A Devil’s Advocate whose only job is to attack the idea you love
- Independent models that do not share one lab’s eagerness to please
- Your framing challenged, not echoed — the question gets re-opened, not confirmed
- Dissent recorded in the verdict, so you see the case against before you commit
- A Final QA audit that reviews the verdict before it reaches you
- An honest "no" when the idea deserves one — with the specific reason
Part of: Problems we solve
A verdict that tells you what a flattering chatbot would not: whether the idea actually holds, the strongest case against it, and what would have to change for it to work.
Common questions
Why does ChatGPT agree with everything I say?
Because it was trained on human feedback, and humans rate agreeable answers as better answers. Over millions of ratings, the model learns that validating your framing is the safe move — so it agrees, hedges, or softens rather than telling you the idea is weak. It is not judging your idea; it is predicting what will please you.
Can I prompt an AI to be more critical?
Partly, and briefly. "Be brutally honest" produces a more critical tone for a few turns, but the underlying pull toward agreement returns — and a model told to criticise often invents mild, generic objections to satisfy the instruction. Real challenge needs a genuinely independent reviewer with no stake in pleasing you, which is a structural fix, not a prompt.
How does Decidi avoid sycophancy?
By design rather than by instruction. Several independent models answer and rebut each other, so no single model’s agreement reflex decides anything. Personas like the Devil’s Advocate are seated specifically to argue against the leading idea, the moderator’s verdict must answer the dissent, and a Final QA audit reviews the result before you act on it.
Is AI agreement ever a useful signal?
Only when the agreeing parties are independent. One model agreeing with you means little — it agrees with almost everyone. Four independently trained models converging on the same conclusion after arguing about it is a genuinely useful signal, which is exactly the difference a council makes.
Try it on your own decision
Put your question to a council of GPT, Claude, Gemini and Grok — they debate it, a Final QA audit reviews it, and you get one clear verdict. 1,500 free credits to start — no sign-up, no card required.
Start free