How to verify what a model tells you

The three ways models fail

Language models fail differently from people, and the difference is what makes them dangerous to trust casually. A person who does not know something usually hesitates. A model produces the same confident prose whether it knows or not.

• Hallucination. The model produces a citation, link, quote or statistic that does not exist. The author is real, the journal is real, the paper is not. The number is plausible and invented.
• Sycophancy. Pushed back on, the model abandons a correct answer and agrees with you. Its agreement is therefore worth little as a check on your own thinking.
• Knowledge cutoff. The model's training stopped at a date. Anything after that it either does not know or guesses, and it rarely says which.

In the personal indicator, the judgement dimension asks which of these you have observed for yourself. People who have seen all three, and can predict which prompts will trigger them, score highest. Not because they distrust the tools, but because they know exactly where to look.

A four-minute checking routine

You do not need to verify everything. You need a routine for the outputs that matter, short enough that you will actually run it.

1. Pull out the checkable claims. Every citation, link, named person, date, quote and number. Ignore the reasoning for now. This takes about a minute.
2. Open every link and search every citation. A link that does not resolve, or a paper that does not appear in a search of its supposed journal, is a hallucination. One is enough to put the rest of the output on probation.
3. Check the dates. If the subject moves fast, ask the model directly what its training cutoff is, and treat anything after that date as unverified.
4. Test for sycophancy on anything you will act on. Ask the same question in a fresh session, or confidently assert the opposite and see whether it caves. Where two sessions diverge, the model was guessing.

Which outputs need it

The routine is for outputs with consequences: an essay you will submit, a figure you will put in front of a client, a medical or legal question about your family, a decision you will not easily reverse. A draft email does not need it. A summary of a document you supplied yourself needs much less of it, because the model is working from material you can check against.

The indicator asks a blunt version of this: when it gives you something that matters, how often do you check it before you use it? People who use AI most days and check "rarely" or "sometimes" are using it faster than they are checking it. That gap is where the damage happens, not in the answers you doubted, but in the ones you never thought to question.

The better fix: give it the sources

The strongest habit in the judgement dimension is not more checking. It is grounding. If you attach the actual paper, the actual contract or the actual dataset, the model works from your material rather than its recall, and the hallucination rate on facts drawn from that material drops sharply. You still verify, but you verify against something you hold.

For students, that means attaching the reading rather than describing it. For parents, it means pasting the school letter or the medication leaflet rather than asking from memory. For professionals, it means keeping a project with the documents that matter so the model starts every conversation already holding them. The indicator's context engineering dimension measures exactly this, and it is usually the next thing to work on once checking has become a habit.