The expensive AI mistakes are not the obvious ones. Obvious errors get caught. It is the plausible ones that reach your audience with your name on them.
1. Invented sources
Models generate citations that look correct — real journal, plausible authors, perfect formatting — for papers that do not exist. Lawyers have been sanctioned for filing them. The format is right, which is exactly why it slips through.
2. Confident wrong numbers
Statistics arrive with the same certainty whether they are accurate, outdated or invented. There is no tonal difference between a fact and a fabrication, so how it reads tells you nothing.
3. Trusting a summary of something you did not read
Summaries are usually good and occasionally invert the meaning of a key paragraph. If you are going to act on one, spot-check the part you are acting on.
4. Publishing the structure along with the content
“In today’s rapidly evolving digital landscape” tells readers how the piece was made before they reach your first real point. That is credibility spent in the opening line.
5. Letting it agree with you
Models are agreeable by design. Ask “is this a good idea?” and you will usually hear yes. Ask “what is wrong with this?” and you get something useful. The question shapes the answer more than the model does.
6. Using it outside your competence
This is the one that compounds. In your own field you catch errors automatically. Outside it you have no filter — and that is exactly where people reach for AI most.
7. Pasting what you should not
Client data, unreleased numbers, someone else’s confidential document. Once submitted you cannot unsubmit, and “I did not realise” is not a defence anyone accepts.
A review pass that takes four minutes
- Open every link and citation. Every one.
- Search one number from the text. If it is wrong, assume the others are too.
- Read the first sentence aloud. If it sounds like a press release, rewrite it.
- Ask which claim you would least like to defend, and check that one.
Common questions
Why does it invent sources instead of saying it does not know?
Because it is completing a pattern, not retrieving a record. A citation-shaped gap gets filled with citation-shaped text. Saying “I don’t know” has to be trained in, and it is imperfect.
Do newer models still hallucinate?
Less, but yes. Lower frequency makes it more dangerous, not less, because you check less often.
What is the single best habit?
Treat the output as a draft from a confident junior who has never been wrong in front of you before. Useful, worth reading, not yet trusted.
