Confident Fabrication
Fluency Is Not Evidence
The failure mode people underestimate is not the model saying "I do not know". It is the model producing a specific, plausible, well-formatted, completely invented detail with exactly the same confident tone it uses for things that are true. A citation to a paper that does not exist. A statistic with a decimal place. A quotation attributed to a real person. A function in a library that would make sense but was never written. The tone is identical because the tone was never tracking truth in the first place. This is why "it sounded sure" is not a reason to believe something, and why the polish of an answer tells you nothing about its accuracy.
- Confidence is a property of the writing style, not a signal of reliability
- Fabrications are specific and plausible, which is exactly what makes them convincing
- Citations, statistics, quotations, dates and names are the highest-risk categories
- A more fluent answer is not a more accurate one
Where Fabrication Concentrates
It is not evenly distributed, which makes it manageable. Risk rises sharply with specificity — exact figures, page numbers, dates, version numbers, direct quotes. It rises in narrow or obscure areas where there was little to learn from. It rises when you asked for something that would exist if the world were tidy: the perfect citation for your argument, an example that fits your case exactly. And it rises when you push for detail the material does not contain, because a request for specifics tends to produce specifics. General explanations of well-established topics are comparatively safe. Precise claims about particular things are where you look first, every time.
- Higher risk: exact numbers, citations, quotes, dates, names, obscure specifics
- Lower risk: general explanations of well-established, widely-discussed topics
- Asking for a detail that "should" exist is a common way to generate one
- Pushing for specifics beyond the source material reliably produces invention
Verify in Proportion to Consequence
Checking everything is unsustainable and you will stop within a week. The workable rule is to ask what happens if this specific claim is wrong. If the answer is "the sentence reads slightly differently", move on. If it is "I mislead a client, make a bad decision, or repeat something false in public", verify before use. Check by going to the source itself, not by asking the model whether it is sure — self-checking is the weakest available method, because the same process that produced the error will happily produce a confident confirmation of it. Two useful habits: an unfindable citation is a fabricated citation, and anything you will repeat to another person deserves thirty seconds of checking.
- Ask "what happens if this is wrong?" rather than "could this be wrong?"
- Always check: figures, citations, quotes, legal, medical and financial specifics
- Rarely check: phrasing, structure, brainstormed options you will evaluate anyway
- A source you cannot find in a minute should be treated as fabricated
Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.