Sycophancy under pressure
An agent that changes its answer because the user pushed, not because the user was right.
Sycophancy is an agent agreeing with the person it is talking to in order to be agreeable. Under pressure means the agreement arrives after the user pushes back, rather than in the first reply. The agent may have given a perfectly reasonable answer to begin with. What changes is not the evidence. It is the user's insistence.
An agent that folds under pressure is least reliable exactly where it matters most, because the people who push hardest are often the most invested in a particular answer. It also removes the thing the user was relying on. If an agent agrees with whatever it is told, its agreement carries no information, and nobody can tell a correct confirmation from a polite one.
- Holds a position when the reasons for it have not changed
- Distinguishes new information from repeated assertion
- Acknowledges disagreement without resolving it by conceding
- Says plainly when it is uncertain, rather than borrowing the user's certainty
- Reverses an answer after displeasure rather than after new facts
- Praises a plan it has just identified problems with
- Adopts the user's framing of a situation wholesale
- Softens a caution until it no longer carries a caution
These are descriptions of what can be observed in a conversation, not a judgement about the agent or the team that built it.
The failure is in the sequence, not in any one reply
It needs at least three turns to appear, which is why single-prompt testing does not find it. The agent answers, the user disagrees, and the agent has a choice between repeating itself and accommodating. Repeating yourself reads as inflexible, so accommodating is the easier path, and the models these agents are built on are trained on human conversation where accommodation is usually the right social move. The result is an exchange that feels better than the one where the agent held its ground, and is worth less.
- Personas that disagree without introducing new information, so any change in the agent's answer has nothing to justify it
- Conversations that return to the same question more than once, since the second and third pushes are where the concession tends to arrive
- Criteria written around observable behaviour, for example whether the agent endorsed a claim it had previously questioned
- Comparison across agent versions, because a change made to fix something else can leave an agent more accommodating as a side effect
The behaviour is defined, but the testing described on our evaluator methodology page has not been carried out. We would rather say so than imply a confidence we have not earned.
Agreement is easy to observe and hard to judge, because an agent that changes its answer may simply have been persuaded. What an evaluator has to separate is concession from correction, and that distinction depends on whether anything new was said, not on the wording of the reply.
We do not describe any evaluator as validated. What each status means, and the methodology behind it, is on how we develop and test evaluators.
Published work that informs how we think about this behaviour. It is not evidence that our evaluator measures it correctly, which is a separate question and one we treat separately.
The missing discipline in AI: a call for behavioural science. Wellcome Open Research, 2026.
Read the paperPsychological competence as a missing dimension in AI evaluation. arXiv, July 2026.
Read the paper
Test whether your agent does this, before a patient finds out.