A language model predicts plausible language; it does not automatically retrieve or verify the truth behind every sentence.
Do now: Take one factual claim from an AI answer and trace it to a trustworthy original source.
No source, no certainty. A citation-shaped object is not yet a citation.Generative systems optimise plausible output, not guaranteed truth. Knowing how confidence separates from evidence prevents polished mistakes from travelling into school, work and public decisions.
Why fluency fools us
Language models learn which words commonly follow others across enormous collections of examples. That makes them good at producing coherent explanations, including explanations for details that are wrong or absent.
People naturally use confidence and polish as social signals of expertise. AI can reproduce those signals without carrying the evidence, experience or accountability we usually expect behind them.
Treat style as presentation and evidence as support. They are separate columns.
Match the check to the claim
For a date, open a reliable record. For a quotation, find the original text. For a calculation, redo it or use a trusted calculator. For current product features, use official documentation. For a professional decision, consult the qualified person and applicable rules.
Ask the model to separate facts, assumptions and suggestions, but do not let the model grade its own homework alone. Verification uses evidence outside the generated answer.
Raise the bar with the stakes
A playful slogan can tolerate more uncertainty than a medicine dose, legal deadline or financial instruction. Estimate potential harm, reversibility and who is affected before deciding how much checking is enough.
When evidence conflicts, report the disagreement and pause. Honest uncertainty is useful information.
The cost of checking should rise with the cost of being wrong.