To check whether an AI tool is telling you the truth, trace the claim back to a primary source, look for a date and an author, and test whether the statement could be proven false — if you cannot find the original source, treat the claim as unverified no matter how confident the AI sounds.
The confidence of the wording tells you nothing about the accuracy of the content. An AI can state a wrong date in the same calm tone it uses for a right one, because it is predicting plausible text, not retrieving a verified fact. So the checking has to happen on your side, with a repeatable procedure, not with a gut feeling about whether the answer "sounds right."
Start with the primary source, not another AI
A primary source is the original place a fact lives: the company's own pricing page, the published study, the court filing, the vendor's release notes. A secondary source — a blog post, a summary, another chatbot — is someone else's retelling, and retellings drift. When an AI gives you a number, a name, or a date, your first move is to find where that fact was first published.
If the AI says a tool costs a certain amount per month, open the vendor's pricing page and read the number there. If it cites a study, search for the study title and check that it exists, that the year matches, and that it actually says what the AI claimed. This is the single highest-value check you can run, and it catches most fabrications, because invented facts usually have no primary source to find.
Look for a date, an author, and a falsifiable claim
Three quick signals separate checkable claims from vague ones. First, a date: facts about tools, prices, and policies change constantly, so a claim without a date is a claim you cannot trust. Second, an author or institution: "researchers found" is weaker than a named team at a named organisation.
Third, falsifiability — can you imagine evidence that would prove the claim wrong? "This tool is the best" cannot be checked. "This tool's Standard plan costs $30 per month" can be checked in about thirty seconds.
When you get a claim that fails all three tests, do not try to verify it. Rewrite your question so the AI gives you something checkable, or drop the claim.
A worked example
Suppose you ask an AI which image tools are worth paying for, and it replies: "Midjourney's Standard plan is $30 per month, and it's rated 4.8 out of 5 for image quality." Both numbers are checkable. According to our AI tool database, Midjourney's plans are Basic at $10 per month, Standard at $30 per month, and Pro at $60 per month, and the editorial rating recorded for it is 4.8 out of 5.
So in this case the AI got it right — but you only know that because you compared the answer against a source that records pricing and ratings at a verification date. Now imagine the same AI told you Midjourney's cheapest plan was $5 per month. The database shows no such tier, so the claim fails, and you would correct it before repeating it.
The procedure is identical in both cases: pull the claim out, find the recorded source, compare. You are not judging the AI's tone; you are comparing two documents.
Know which claims are cheap to check and which are not
Names, dates, prices, plan tiers, and version numbers are cheap to verify — they live on a single page and take under a minute. Synthesis is expensive: when an AI summarises a whole field, or answers a contested question like whether a policy is good, there is no single page to check against, and you would need to read several sources and form your own view.
For those, use the AI to find the sources, then read the sources yourself and decide. And there are cases where you should not lean on AI output at all: legal, medical, and financial decisions where a wrong detail carries real cost, and any situation where you cannot find the underlying source.
In those cases the honest move is to say the AI output is unverified and go to a human expert or the original document. One practical tip: ask the AI to quote the exact sentence from its source and name the source. It will often decline or produce something vague — and that hesitation is itself useful information about how much weight the claim deserves.
For a gentler starting point on what these tools are good for, see What can I actually do with AI tools as a total beginner?.