The trouble is not that the model is wrong — every source is sometimes. The trouble is that we agree with it and do not notice. Do not ask it what is right; make it attack your position and its own, and check what the conclusion stands on outside the chat.
Your premise is input too. Models measurably drift toward the user’s stated view: “help me justify X” returns X, polished. State the question without your answer in it, or ask for the strongest case against.
Agreement filters, it does not verify. Re-asking, self-review and a second model give stability, not truth: without an external signal self-correction rarely fixes reasoning errors. Truth comes from outside — a run, a test, a primary source. Find the load-bearing claim, have the model attack it, check it outside the chat, and say how much your own confidence moved and why.
The same rule as LLM output is untrusted input and Grounding, applied to the person instead of the system.
The evidence. Sharma et al., “Towards Understanding Sycophancy in Language Models” (2023), found sycophancy across several production assistants: they shift answers toward a user’s stated belief, give way when challenged even when they were right, and human preference data rewards some of this. Kamoi et al.’s survey of self-correction (TACL, 2024) concludes that LLMs rarely fix their own mistakes from prompting alone; reliable gains come with external feedback — a tool, a test, a verifier. Both are measured behaviour of current models; why it happens is still a hypothesis.
An engineering example. “Confirm that reading from our replica is fine right after a write” returns a confident yes with a tidy explanation. Asked instead “under what conditions does a read on the replica miss a write just acknowledged by the primary?”, the same model names asynchronous replication lag — and that is checkable outside the chat: write, read from the replica, compare, under load (65).
Limits. A model told to attack will always produce objections, and some are weak or invented; weigh them like any other output. Adversarial review costs time, so spend it on the claims the decision rests on and where being wrong is expensive (111).