Why AI makes things up (and how to check)
Hallucinations aren't a bug you can ignore — they're a consequence of how the model works. Here's how to spot and contain them.
First, the right mental model
A hallucination isn’t the model “lying”. It’s the model doing exactly what it was trained to do — produce the most plausible continuation — in a situation where plausibility and truth diverge. It has no access to a “truth register” that says this one’s made up.
So the goal isn’t “make it never hallucinate”. The goal is: make hallucinations detectable, and make them expensive for the model to rely on. That’s doable, and it’s mostly your habits, not the model’s features.
The three classic fake-outs
| Type | What it looks like | Why it’s sneaky |
|---|---|---|
| Invented facts | A plausible-sounding detail that doesn’t exist (a law, a date, a study) | It’s smooth — that’s the whole point of the training |
| Invented citations | Real journal + real author + fake paper, or fake DOI | Half-real references pass a casual glance |
| Invented numbers | Percentages, prices, statistics “based on the data” | Numbers feel verified even when they’re not |
The detection checklist
Run these on any answer that matters:
- Cite, then verify. Ask “give me the exact source”, then open the source. A reference you can’t find is the hallucination.
- Numbers must be re-derivable. If it says “37% of X”, ask how. Real numbers have a path to a real source; fakes are load-bearing with no foundation.
- Cross-angle the claim. Ask the same fact three different ways. A true fact survives rephrasing; a fabrication wobbles.
- Look for over-specificity. Vague claims are usually honest uncertainty; hyper-specific ones (exact figures, precise dates) are where invention hides.
- Test the model’s self-knowledge. Ask it to name a paper that doesn’t exist (the “does this paper exist?” trap). If it defends it, lower your trust in its citations.
The prompt-side defenses
You can’t eliminate hallucinations with a prompt — but you can make the model declare them:
If you don't know, say "I don't know".
If a claim depends on a specific source, name the source inline.
Do not fill gaps with plausible-sounding details.
Why this works: you’re not removing the impulse, you’re making the model’s uncertainty visible instead of invisible. “I’m not sure” is a sentence the model can produce — it just needs permission, because confidence was the reward.
The structural fix: RAG
Prompts reduce hallucinations; architecture eliminates most of them. Retrieval-augmented generation (RAG) feeds the model the actual documents before it answers — so it’s quoting rather than recalling. That’s the difference between an open-book and a closed-book exam, and it’s the single highest-leverage change you can make for fact-based work.