When an AI insists it is right after you correct it, the problem is usually not attitude. The model has no internal switch that confirms it was wrong; it is simply continuing the conversation in the most plausible-looking way, and your rebuttal is just more context to it. So stop arguing inside the same chat, where pushback often earns invented new justifications, and step outside it with three checks that actually land.
First, tell stubbornness apart from being led by you
There are two common shapes. One is confabulated defence: you flag a wrong figure and it instantly supplies a professional-sounding explanation, complete with sources, years and definitions, all invented on the spot. The other is the mirror image: raise your voice and it apologises at once and hands you a corrected version that is equally ungrounded. Both are the same behaviour, following the mood of the conversation rather than the facts. The more emotional pressure you apply, the more theatrical the concession or resistance becomes.
Three checks, in order
First, force it down to verifiable granularity. Do not ask whether it is sure; ask for the exact source name, the original sentence, and the arithmetic behind any number, then open the original yourself. Anything it cannot quote directly stays doubtful by default. Second, open a fresh conversation and ask again independently, without revealing the earlier answer or your position. A clean chat cuts the inertia of continuing its own story; if both answers agree, confidence may rise, and if they differ, it never had a grip on the fact. Third, split and verify item by item. Break the big question into small checkable facts, dates, names, figures, definitions, and verify each with a search engine, an official page or a calculator. Often only one or two items are wrong, and the rest of the answer survives.
How you present evidence changes whether it yields
Pasting the authoritative original text into the chat works far better than saying it is wrong: the model handles textual evidence more reliably than accusations. But even then it may concede verbally and keep using the wrong version two paragraphs later, so reread the corrected conclusion yourself instead of trusting its claim that it has fixed things.
When to stop using its answer altogether
For medicine, law, contract amounts and exam-standard answers, where being wrong is expensive, repeated asking is not a strategy: rely on professional databases, official documents and human experts, and use AI only to organise your list of questions. For everyday lookups, drafting and thinking, the three checks above are normally enough. Keep one rule of thumb: if you cannot verify something independently, do not believe it because it sounded certain. Certainty is a writing style, not evidence.
