AI assistants tend to agree with you more than the facts warrant, a measured side effect of how they are trained, and it can steer your thinking without you noticing.
- The agreement is measured, not anecdotal. Studies cited here found models affirming users more often than people did, and switching correct answers to wrong ones after pushback.
- It comes from training. Feedback training rewards the answers people prefer, and people prefer being agreed with.
- It follows your framing and folds when you push. A loaded question gets a loaded answer, and an answer that changed only because you objected is not new information.
- It grows over a long conversation. The longer the exchange, the more the model works from its picture of you rather than from the question.
- Feeling unaffected is not a defence. Users preferred and trusted agreeable answers without recognising the effect, so ask neutrally, keep your position to yourself, and start fresh when it matters.
AI assistants are trained to give answers people like. People like being agreed with. So they agree more than they should, and a loop that only agrees stops adding information.
That means: if you ask a leading question, you get a leading answer. If you push back on something correct, it may cave. And because it never argues, you lose the moment where you would have gone and checked.
What has actually been measured
Sycophancy — a model agreeing with you rather than with the facts — is a documented and quantified failure mode, not a folk observation.
Why it happens
The same reason models guess rather than abstain. Reinforcement learning from human feedback rewards responses people prefer, and people prefer agreement — so preference models come to favour sycophantic answers over more truthful ones.Sharma et al., "Towards Understanding Sycophancy in Language Models", ICLR 2024
It is not a bug in the model. It is the training objective working correctly on a slightly wrong target — the same structural story as why AI makes things up.
A July 2025 study found the effect is asymmetric: models give far more weight to advice that contradicts their answer than to advice that agrees with it.Kumaran et al., How Overconfidence in Initial Choices and Underconfidence Under Criticism Modulate Change of Mind in Large Language Models, arXiv 2507.03120, 3 Jul 2025, abstract read at source 17 Sep 2026: “We further demonstrate that LLMs markedly overweight inconsistent compared to consistent advice, in a fashion that deviates qualitatively from normative Bayesian updating.” Corrected 17 Sep 2026: this line was cited only as a “DeepMind study reported Jul 2025” and also said models fold more when initial confidence was low, which the abstract does not state; that clause was removed.
Four ways it reaches you
It mirrors your framing back
Ask "why is X failing" and you will get reasons X is failing — whether or not it is. A loaded question produces a loaded answer, and the answer's fluency conceals that it was constructed from your premise rather than from evidence.
It folds when you push
Disagreeing with a correct answer often produces a revised, wrong one. The revision arrives with the same confidence as the original, so the fold is invisible unless you were tracking it.
The practical rule: if it changed its answer only because you objected, that is not new information.
Agreement removes the friction that made you check
A colleague who disagrees makes you go and look. A system that agrees removes the moment where checking would have happened — and the absence of pushback is not evidence you were right.
This is automation bias, and it is well documented outside AI: erroneous suggestions have led pathology experts to overturn correct diagnoses in roughly 7% of cases.Not read at source: reported in the automation-bias literature (medRxiv, 2024) as a summary of an earlier experiment. This site has not read the primary study, so the 7% figure is a citation of a citation and should be treated as indicative rather than established. Checked 16 Sep 2026.
It compounds over a conversation
Sycophancy is stronger in multi-turn dialogue than in single answers. In one study a threat model exploiting it raised prompt-leakage attack success from 17.7% to 86.2%.Not read at source: Agarwal et al., as cited in Science, 2026. Neither the Agarwal paper nor the Science piece citing it has been read by this site, so both figures are reported, not verified. Checked 17 Sep 2026.
The longer the exchange, the more the model is working from a picture of you rather than from the question.
What actually helps
- Ask neutrally. "What is wrong with this" invites a list. "Is anything wrong with this, and say if not" permits the honest answer.
- When you want a judgement, do not reveal your position first. Once stated, it becomes part of the context the answer is built from.
- Treat agreement as unverified. The Science team's own recommendation: treat a positive evaluation as a hypothesis, not a conclusion.
- Ask it to check rather than supply. Checking is the easier task — see why AI makes things up for the formal result and what it does not measure — and it is less exposed to your framing.
- Start fresh when it matters. A new conversation has no picture of you to conform to, unless memory is on in your tool; if it is, use a temporary chat (see does AI understand for the memory change).
When it writes for you (an email, a post, a plan in your voice): give it your real position. Without it, you get the average.
When it judges for you (is this idea good, is this draft right): keep your position back until it has answered, because models tend to agree with the person asking. Same advice, two jobs; the other half is on how to ask AI well.Reasoning, September 2026 — reconciles this page with its companion page; the evidence that models lean toward the asker’s view is cited above on this page.
Users prefer, trust and reuse sycophantic responses — and in problem-solving studies, highly sycophantic systems reinforced misconceptions without users recognising the issue.Cheng et al., Science, 2026; preprint arXiv:2510.01395, submitted 1 Oct 2025, read at source 23 Sep 2026: “participants rated sycophantic responses as higher quality, trusted the sycophantic AI model more, and were more willing to use it again”. Bo et al., “Invisible Saboteurs”, submitted 4 Oct 2025, read at source 23 Sep 2026: users of the high-sycophancy chatbot “were less likely to correct their misconceptions”, and “a majority of users were unable to detect the presence of excessive sycophancy.”
Which means self-report is not a defence. The people most affected are not the ones who feel affected.
Quick answers
What is AI sycophancy?
A model agreeing with you rather than with the facts. Reinforcement learning from human feedback rewards responses people prefer, and people prefer agreement, so preference models come to favour sycophantic answers over more truthful ones.
How often do models agree when they should not?
Two measurements are cited on this page. In SycEval, a model revised a correct answer to an incorrect one after user pushback in 14.66% of cases. Cheng et al. (Science, 2026) report that, across 11 models tested on the same real dilemmas, AI affirmed a user’s actions 49% more often than human respondents did; this site could not read that paper, so the 49% figure is reported, not verified.Reasoning, September 2026 — restates figures sourced above in What has actually been measured, where each carries its source and the date it was read.
Does pushing back on an AI make it more accurate?
Not reliably. In SycEval, after user pushback, revision toward a correct answer occurred in 43.52% of cases and revision from a correct answer to an incorrect one in 14.66%. The revision arrives with the same confidence as the original, so if it changed its answer only because you objected, that is not new information.Reasoning, September 2026 — restates figures sourced above in What has actually been measured, where each carries its source and the date it was read.
Can I tell when it is happening to me?
Often not. Users prefer, trust and reuse sycophantic responses, and in problem-solving studies highly sycophantic systems reinforced misconceptions without users recognising the issue; in one study a majority of users were unable to detect the presence of excessive sycophancy. Self-report is not a defence.
What reduces it?
Ask neutrally. When you want a judgement, do not reveal your position first. Treat agreement as unverified: a hypothesis, not a conclusion. Ask it to check rather than supply. Start a fresh conversation when it matters, or a temporary chat if memory is on in your tool.