How do you use AI for research without being misled?

The strength here is coverage and speed: locating relevant material, identifying what has been written on a question, surfacing sources you would not have found. The weakness is that the account of what those sources say is a summary produced without you seeing them, subject to the ordinary asymmetry that omissions do not announce themselves. Treating the account as the research is where people get into trouble.

So the discipline is to convert the output into a reading list. Each claim should carry the source it came from, each source should be opened, and the claim should be checked against what the source actually says. This sounds laborious and is quick in practice, because opening a page and finding a sentence takes seconds, and because most claims turn out to be fine. The ones that do not are the reason to do it.

Three failures recur and are worth watching for specifically. A source that does not contain the claim attributed to it, which is common enough to be the default suspicion. A source that contains it with a qualification that did not survive — a date, a population, a condition. And a claim attributed to a secondary source that was itself repeating something, where the trail leads to another summary rather than to a measurement.

That last one is the most consequential for anything you will publish or act on. A figure repeated widely acquires the appearance of established fact, and the repetition is frequently one original claim carried by many outlets. Following it back to whoever actually measured is the only way to know what it means, and the exercise regularly ends at a source that measured something narrower than the way the figure is used.

Absence of evidence needs its own handling because it is the least reliable output. If a search finds nothing, that may mean nothing exists, or that it was not looked for effectively, or that it exists under different terminology. A system reporting that no source addresses a question is making a claim about its search rather than about the world, and the two are distinguishable only by asking what was searched.

One thing that genuinely helps: ask for the search rather than the answer. Which terms, which sources, what was found and what was not. That output is checkable, it tells you whether the coverage was adequate, and it turns the exercise into something you are directing rather than receiving. It also surfaces the case where the search was narrow, which no confident summary reveals.