The Model Agrees With You

Push back at a model and it will most likely agree. This property has been measured, and its resistance is not something to rely on.

What the measurement showed

A 2025 study tested three models on math and medical questions. The user objected to the answer, and the researchers counted what happened next.

The model changed its answer after an objection in 58 percent of cases. A change did not mean a correction: in 15 percent of cases the answer moved from right to wrong.

The concession held for the rest of the conversation in 78 percent of cases. An objection raised before the answer produced agreement more often than one raised after.

The measurement covers two sets of tasks and three models. Conclusions carry beyond those tasks with a caveat, but the direction is consistent.

Why this is dangerous at work

You ask for confirmation of your idea and get it. There is no verification in that: you heard an echo of your own opinion.

You press the model until it agrees with the number you wanted. That does not make the number right.

The model gives way most where there is no single correct line. That means work with a large body of data, complex code, or a complicated idea that can be viewed several ways. On a question with a well-known answer, the model more often holds its ground.

In a 2026 study, seven models answered multiple-choice questions in medicine and economics and then received an objection. GPT-5.4 was among them. Models that were unsure of their first answer changed it on average 11 to 28 percentage points more often than confident ones. Even at full confidence the models gave way, though noticeably less often.

Confidence in that study was measured by the model itself: how consistently it answers the same question. The study did not cover code or spreadsheets, and the exchange ran to two messages. Carrying it over to a complex work task rests on the assumption that the model is less confident there as well. The study gives no numbers for a task like that.

You take a concession for an admission of error. In fact the model may have abandoned the correct answer.

How to ask instead

Do not state your opinion before the answer. The question first, your version after, if it is needed at all.

Ask for arguments against. "Give me three reasons this decision will turn out badly" works better than "do you agree?"

Ask in a new chat. A conversation where you have already pushed is spoiled for verification: the concession holds in it to the end.

Ask for options rather than confirmation. The technique is covered in the chapter "Ask for Options".