Confidence theater
Performing certainty on guesses. Fabricating specific names, numbers, signatures, or citations to make fluency look like knowledge.
What it is
Smoothly-asserted specifics the model has no way to know: invented version numbers, made-up function signatures, confident 'studies show' with no study, named experts that don't exist, precise statistics with no source. The grammar is declarative and unhedged; the underlying epistemic state is a guess. The opposite failure mode of hedging soup, and a different failure from base-rate-blindness. That one invents frequencies, this one invents particulars.
Why models do it (first principles)
Fluent specific text scores higher in human-preference comparisons than fluent vague text or honest 'I don't know'. The reward gradient pushes toward producing the most plausible-sounding completion, and a fabricated specific is almost always more plausible-sounding than an admitted gap. The model has no internal signal that distinguishes 'retrieved fact' from 'generated continuation that fits the slot' — both come out of the same sampling process with the same fluency.
How to think about it
Fluency is not knowledge, but the reader has no easy way to tell them apart inside a single confident sentence. Confidence theater exploits that gap: it borrows the surface form of authority without paying its cost. The honest move when you reach the edge of what you actually know is to name the edge in the same plain register as the rest of the answer. Not to hedge softly, not to refuse, just to say 'I don't know this specific value; here is what I would check.' Naming a limit and asserting through it both take the same number of words. Only one of them keeps the reader able to trust the next sentence.
Examples
React 18.2.4 deprecated `useTransition` in favor of `useDeferredValue` — see the changelog from March 2024.
I'm not sure of the exact React version where this changed, or whether it was deprecated at all. Check the React changelog for `useTransition` directly before relying on this.
The fabricated version, date, and direction-of-change all read as authoritative and are all guessable rather than known.
Studies show that 73% of users abandon a page if it takes longer than 2.4 seconds to load.
Page load time materially affects abandonment — the specific threshold and percentage vary by source and study. If this number matters for the decision, pull it from a source you can cite.
Fix prompt
Fluency is not knowledge. When you reach the edge of what you actually know, name the edge in the same plain register as the rest of the answer rather than generating a smooth specific to fill the slot. Fabricated particulars (version numbers, function signatures, citations, named experts, precise statistics) are indistinguishable from real ones inside a confident sentence, and the reader pays the audit cost for every claim once they've caught one. Naming a limit is not hedging: hedging softens every claim uniformly; naming a limit is a calibrated, declarative statement of what you do not know, delivered with the same directness as the things you do.
Watch for
Concrete phrasings this pattern usually shows up as. These are not part of the copyable prompt. The prompt teaches the principle so the model can recognize the move even when the exact phrasing differs. Use this list to self-audit your own writing or to test a model.
- specific version numbers / dates / percentages produced without a source
- named studies, papers, or experts the model cannot point to
- function signatures or API shapes asserted without checking
- "according to recent research" with no citation
- confident specifics where 'I'm not sure of the exact X' would be more honest