How do you stop an AI from making things up?
Not by asking nicely. The tools that work are structural: give the model real sources, make it point at them, and check the pointing.
You cannot switch hallucination off, but you can make answers checkable. Ground the model in retrieved passages instead of memory, require a citation for every claim, and run a verification pass that compares each claim against the passage it cites, removing what the passages do not support. Where the sources have no answer, the system refuses. Dialing down temperature makes wording steadier, but it is not a fix on its own; the combination above is what converts a confident guess you must trust into an answer you can audit.
Change what the model reads, not what it is told.
A model answering from memory has nothing to be right about except its training data, which is where hallucination comes from, which is why retrieval comes first in every honest fix: real passages from your documents go into the prompt, and the model answers from those instead of from memory. This removes most invention at the source, because the model no longer needs to recall what a source said; it is looking at the source. What remains is the subtler failure of saying slightly more than the passage does, and no prompt, however stern, closes that gap, which is why the next step exists.
Verify after, not before.
The second layer checks the answer after it is written. Each claim is compared against the exact passage it cites, and the claims the passage does not support are cut or rewritten; a question the sources cannot answer gets a refusal instead of a guess. This is the difference between citations that look convincing and ones that hold, and it runs as code, not as a hope. Refusal deserves defending: in research work, being told your sources do not say is a useful finding, and it is the one outcome a chat-only tool cannot reliably produce because it has no sources to check against.
Keep the checking cheap.
The habit that catches everything else is clicking the citations, and a good tool makes that habit cheap by design: one click opens the passage, the passage is highlighted, and the claim sits next to the words it came from. When checking is cheap you actually do it, and the tool's trustworthiness stops depending on its tone of voice. A system that grounds, cites, verifies and refuses is the honest ceiling today; the residual errors it leaves are ones a reader can find in seconds rather than ones that hide in fluency, and that difference is what makes an answer usable in work that matters. The downstream habit is just as simple: citing an AI answer properly means crediting the tool and citing the sources you verified.