The awkward position with these tools is needing an answer in an area where you cannot evaluate one. Several checks are available that do not require the expertise the answer concerns.
Ask the same question in a different form
Pose the question again in a new session, phrased differently, and compare.
Well-supported material tends to come back consistently, because the underlying information is stable across many ways of asking.
Weakly supported material varies noticeably between attempts, and that variation is a usable signal even when you cannot judge either version on its merits.
Ask for the reasoning before the conclusion
Requesting the steps rather than the answer exposes structure you can inspect without domain knowledge.
Reasoning that skips from premise to conclusion, or that restates the question as a justification, is visible as a gap regardless of subject.
This also tends to improve the answer, since generating the intermediate steps changes what the final one is conditioned on.
Treat specifics as claims to verify, not as evidence
Named sources, figures, dates and citations are the parts most likely to be fabricated and the parts that make an answer feel authoritative.
The correct response to a specific claim is to look it up, and the correct response to one you cannot find is to assume it does not exist.
An answer that becomes vague when you ask it to identify its source for a particular figure has told you something useful about that figure.
Ask what would make the answer wrong
Requesting the strongest case against a position, or the conditions under which it fails, surfaces caveats that a direct answer omits.
Models comply with this readily, and the omitted qualifications are frequently the part that mattered for your situation.
A position that produces no serious objection when asked for one is either uncontroversial or being presented more confidently than it should be.
Use it to find the question, not the answer
The most reliable use in an unfamiliar area is establishing vocabulary and structure: what the field calls this, which distinctions matter, what a competent person would ask next.
That output is easy to verify, because correct terminology leads to real sources and invented terminology leads nowhere.
Having the right terms turns an unanswerable search into an ordinary one, at which point you are checking a human source rather than trying to audit a model.