An assistant states a wrong fact in the same tone it uses for a correct one. The uniformity of that confidence is a product of how the models are trained rather than a presentation choice.
Why the tone never varies
Generation selects likely continuations, and a hedged sentence is only produced when hedging was likely in similar contexts.
Training data is dominated by confident explanatory prose, since people write down what they are sure of and rarely publish their uncertainty.
The model reproduces that register regardless of whether the underlying information is well supported. Fluency and accuracy were correlated in the training data and are independent in the output.
What the model does and does not track
There is a real internal signal about how strongly a continuation is favoured, and it correlates loosely with reliability.
That signal is not exposed in the wording of the answer, so a low-confidence generation reads exactly like a high-confidence one.
The mechanism that would produce a warning exists at the wrong level, and connecting it to the phrasing of the response is an unsolved design problem.
How helpfulness training makes it worse
Models are tuned on human preference, and people rate direct answers above hedged ones when both are correct.
That gradient rewards confidence generally, since the rater usually cannot tell which answers were uncertain.
Tuning hard against this produces an assistant that qualifies everything, which raters dislike even more, so the equilibrium settles on the confident side. Calibrated hedging would require the rater to know the answer, which is the situation the assistant was there to avoid.
Why asking the model to check rarely helps
Asking whether an answer is correct produces another generation, conditioned on the answer already given.
The previous response is now context, which biases the review towards agreement, and a model asked to be certain will often supply certainty.
Verification that works uses an independent path: a fresh conversation, a different phrasing of the question, or an external source that does not depend on the model at all.
Where the failures concentrate
Confident errors cluster in a predictable region: specific citations, precise figures, names attached to claims, dates, quotations and anything that must be exactly right.
General explanations of well-established subjects are considerably safer, because the training data was consistent about them.
The usable rule is to accept explanations and verify particulars, which matches where the reliability actually differs rather than treating the whole output as one thing.