Every figure is a claim, not a fact
The dangerous output isn't the obviously wrong one. It's the one that looks exactly right.
When a model invents something, it doesn't usually produce gibberish. It produces the most plausible-looking version of the thing you asked for. Ask for a statistic and you get a clean percentage. Ask for a source and you get an author, a journal, a volume, a page range. The fabrication arrives wearing the costume of a fact — and the costume is the point. Fluency is the model's native register, and fluency reads as confidence to us.
So the instinct to trust "specific" information gets it backwards. Specificity is cheap to generate and expensive to check, which is precisely why a made-up citation tends to be more detailed than a real half-remembered one. The detail is doing persuasion, not verification.
The shift that fixes most of it
Treat every concrete figure and every citation not as something the model knows, but as something the model is claiming. A claim is a thing you check before you stand behind it. That single reframe — from "the AI told me" to "the AI claims" — does most of the work, because it changes what you do next by default.
A real-looking citation is not evidence the citation is real. Verify the source, not its plausibility.
Two quick tells worth internalising. First, internal inconsistency: numbers that don't sum, a date range that contradicts a later calculation, a total that drifts between paragraphs. The model optimises each sentence for plausibility, not the passage for arithmetic — so the seams show if you read for them. Second, the obscurity trap: the more niche the topic, the more confidently the model will confabulate, because there's less in its training to anchor it and no part of it that knows it's guessing.
Try this. Before any AI-sourced figure or citation goes into work you'll put your name to, spend two minutes tracing it to one primary source. Not a site that repeats it — the report, the dataset, the transcript itself. If you can't find it in two minutes, that's information too.
This isn't suspicion for its own sake. Over-checking the trivial is its own miscalibration — you don't need to verify a casual restaurant suggestion. The skill is reserving the check for the things that would cost you if they were wrong, and making it automatic there.