Verifying AI-generated information means checking specific claims against independent sources. Asking the same AI the same question again is not verification. Verification happens outside the chat window, in peer-reviewed research, official databases, and reputable journalism.
Here is the five-step check, roughly how long each step takes, and a worked example of catching a citation that does not hold up.
Asking twice is not a fact-check
The most common mistake is asking the model "are you sure?" or opening a second chatbot for a second opinion. Because these systems are trained on heavily overlapping data, they tend to be confidently wrong in the same way. A model that is hallucinating a fact will usually stick to its story, and a competing model will often repeat the same error back to you.
A second AI is not a second source. It is the same source wearing a different logo.
The five-step verification check
- Separate the claims from the prose. (30 seconds) Most AI answers mix safe generalities with a few specific, checkable assertions: a number, a date, a name, a quote, a citation. Mark those. They are the only parts that need verifying.
- Confirm the source exists. (1 minute) Before you check whether a claim is true, check whether its source is real. Paste the exact title, author, or DOI into a search engine. Invented citations are the most common failure, and they collapse the moment you look for them.
- Go to the primary source, not a summary. (2 to 5 minutes) If the AI quotes a statistic, find the original study, filing, or government report. Summaries of summaries drift. The original number is often narrower than the claim built on top of it.
- Check the date against the training cutoff. (30 seconds) Models have a knowledge cutoff, and web search does not always compensate for it. Anything recent, anything that changed, anything described as "current" or "latest" deserves a fresh look.
- Escalate anything consequential to a human. (varies) For health, legal, tax, or financial questions, an AI answer is a starting point for a better conversation with a professional, not a substitute for one.
A worked example
Say you ask a model for research on context switching and it hands you this:
A 2021 MIT Press book by Harrow puts the average knowledge worker's context-switching loss at 2.1 hours per day.
The citation above is invented for this example, but the shape is real, and the shape is what fools people: a named publisher, a plausible year, an author surname, and a number precise enough to sound measured.
Run the check on it:
- The checkable claims are the publisher, the year, the author, and the figure. Everything else is filler.
- Does the source exist? Search the exact title, then the publisher's own catalogue. A real book from a major academic press is trivially findable. If two searches turn up nothing, stop here. The claim is dead.
- If the book is real but the number is not in it, you have hit the harder case: a genuine source with a figure attached to it that it never reported.
- Interrogate the precision. "2.1 hours" reads as measured. Ask what was measured, over what period, and for which workers. Suspiciously exact numbers often come from a small study about one narrow population, generalised into a claim about everyone.
The failure mode is rarely an outright lie. It is usually a real finding stripped of the conditions that made it true. That is exactly the kind of error a second chatbot will happily confirm for you.
Want this to become a habit rather than a checklist?
The AI Starter Kit walks you through seven days of building verification into the way you already work.
What to do when you cannot verify it
Sometimes the source is paywalled, the data is proprietary, or you do not have twenty minutes to spare. That is a normal outcome, and it has a normal answer: do not use the claim.
An unverified specific is worse than a verified general. "Research suggests context switching costs more time than people expect" is defensible. "A 2021 MIT Press book found 2.1 hours a day" is not, unless you have read it.
Dropping the number costs you a little authority in the moment. Being caught with a fabricated citation costs considerably more.
If you are still deciding whether an answer is even worth verifying, start with the red flags that suggest an AI is guessing.