Language models write fluently, which makes their output feel reliable even when it is wrong. Dates, numbers, quotes, and citations are the parts most likely to be invented, because fluency does not require truth. Evaluation means checking the output the way you would check a colleague's draft: against sources, against logic, and against itself.
Checks that catch most errors
- Verify every factual claim that matters against a source you trust, especially numbers and dates.
- Test the logic: does the conclusion follow from the premises, or does it merely sound like it does.
- Ask the same question twice in different words and compare answers. Inconsistent answers signal guessing.
- Watch for confident vagueness, such as named studies with no verifiable details.
Never paste model output into work without reading it fully. The time saved by generating is lost twice over if an error reaches a reader.