The most dangerous thing about AI answers isn’t that they’re often wrong. It’s that a wrong answer arrives in exactly the same confident, well-organized tone as a right one.
The case everyone in law now knows is Mata v. Avianca. In 2023, lawyers in a New York federal case filed a brief that cited court decisions ChatGPT had made up, complete with fake quotes. When they had asked whether the cases were real, ChatGPT assured them they “indeed exist” and could be found in legal databases. They didn’t. In June 2023, Judge P. Kevin Castel fined the lawyers and their firm $5,000.
Tools have improved a lot since then, but the lesson hasn’t changed: you can’t ask the model to vouch for itself. This guide is the routine I’d recommend for checking an AI answer before you rely on it. It takes about five minutes, and it catches most problems.
Why AI answers go wrong
Knowing how errors happen tells you where to look.
- Plausible beats true. Language models produce text that fits patterns they learned. When they don’t know something, they can produce something that merely sounds right instead of saying “I don’t know.”
- Knowledge goes stale. A model may not know about recent events unless it’s searching the web, and even then it can fall back on older information.
- Sources get blended. Two studies, two people with similar names, or two versions of a policy can merge into one confident sentence.
- Citations get invented or misread. A link may not exist, may point somewhere unrelated, or may say something different from the summary built on top of it.
- Numbers drift. Percentages, dates, units, and currencies are where errors cluster.
Assistants that search the web and show their sources are much better than they were a few years ago. They still make every one of these mistakes some of the time.
First, decide how much checking it deserves
Not every answer needs the same scrutiny. Match the effort to the stakes.
| Stakes | Examples | What to do |
|---|---|---|
| Low | Brainstorming names, rewording an email, explaining a concept you’ll study anyway | Read critically. No formal check needed |
| Medium | Facts in a work document, a summary you’ll share, instructions you’ll follow | Run the five-minute check below |
| High | Health, legal, money, or safety decisions; anything you’ll publish | Run the check, then confirm with primary sources or a qualified professional |
The five-minute check
Minute one: list the claims
Skim the answer and pull out the specific things you could check: names, numbers, dates, quotes, citations, and cause-and-effect statements like “X led to Y.” Ignore the connecting text. Most answers rest on three to six claims, and those are what you need to verify.
If you can’t find a single checkable claim, notice that too. Vague answers are hard to prove wrong, but they’re also hard to rely on.
Minute two: open every source
If the assistant cited sources, click every one. For each, ask three questions. Does the page exist? Is it the kind of source it claims to be, such as an official page, a study, or a reputable outlet, rather than a content farm? And does it actually say what the answer says it says?
That last question catches the most errors. In practice, a real source paired with a wrong summary is far more common than a fake link, and it’s much easier to miss because the link looks fine.
If there are no sources, ask for them and check those. Just don’t accept the model’s word that its sources are real. That’s precisely what went wrong in Mata v. Avianca.
Minute three: read laterally
Fact-checkers don’t judge a source by reading it closely. They open a new tab and see what everyone else says about it. Digital literacy researcher Mike Caulfield turned this into a method called SIFT:
- Stop before you react or share, especially if a claim gives you a jolt of outrage or delight.
- Investigate the source. Who’s behind it, and what are they known for?
- Find better coverage. Are other trustworthy sources reporting the same thing?
- Trace claims back to their original context: the actual study, the full quote, the original photo.
Searching the key phrase in quotation marks, or a name plus a keyword, often turns up the original in seconds.
Minute four: check numbers and dates
Look hard at every figure.
- Is it current? Look for an “as of” date. Prices, laws, product features, and statistics change constantly.
- Is it the right measure? Revenue and profit, users and monthly active users, percent and percentage points all get mixed up.
- Does the math work? If something “doubled” from 40 to 70, it didn’t.
- Are the units right? Watch for currencies, millions versus billions, and metric versus imperial.
Minute five: ask what would change the answer
Ask the assistant to poke holes in its own answer. Something like: “Which parts of your answer are you least confident about, and what would I need to check to confirm them?”
This often surfaces the weak spots quickly. Treat the reply as a to-do list, though, not as verification. The model is still grading its own homework. For anything important, putting the same question to a model from a different provider is a cheap second opinion; our guide on how to choose an AI model covers when a second opinion is worth it.
Red flags worth a second look
| Red flag | Why it matters |
|---|---|
| Very precise numbers with no source | Precision can be invented as easily as anything else |
| Quotes from famous people | Misattributed quotes are everywhere in training data |
| Citations with generic titles, or links that land on a home page | Common signs of a citation that doesn’t exist |
| Detailed descriptions of recent events from a tool without web access | It may be filling gaps from older information |
| An answer that changes when you regenerate it | The model is unsure, even if every version sounds certain |
| An answer that says exactly what you hoped | Models tend to go along with how a question is framed |
That last one is the sneakiest. If you ask “Why is X a good idea?” you’ll usually get reasons X is a good idea. Ask “What are the strongest arguments for and against X?” instead.
Special cases
Images, screenshots, and video
AI-generated and edited images are routine now, and a screenshot of a “post” takes seconds to fake. Run a reverse image search to find earlier copies and the original context. Check whether the file carries Content Credentials, the provenance standard from the C2PA coalition that some cameras, editing apps, and AI generators attach to images.
Regulation is starting to help. Since August 2, 2026, the EU AI Act’s transparency rules require providers of generative AI systems to mark synthetic content in a machine-readable way, and require those who publish deepfakes to disclose them. Systems that were already on the market before August 2 have until December 2, 2026 to add the marking. Our explainer on the EU’s AI content labeling rules covers the details. Labels can still be missing or stripped, so treat their absence as “unknown,” not “real.”
Health, law, and money
Use AI to understand your situation and prepare better questions, not to make the final call. Confirm anything you’ll act on against a primary source, such as an official government page or the actual text of a law or policy, or with a qualified professional.
Breaking news
Early reports are often wrong, and AI summaries of them inherit those mistakes. For fast-moving stories, wait for confirmation from several established outlets, or just go to them directly.
Ask for answers that are easy to check
You can save yourself time by asking for a checkable answer in the first place:
- “Cite a source for each factual claim, with a direct link and a short quote from the page.”
- “Separate what’s well established from what’s uncertain or disputed.”
- “Include the date each statistic refers to.”
- “If you don’t know, say so instead of guessing.”
None of these guarantees accuracy. Each makes problems easier to spot. And if you’re pasting in documents for the model to check against, strip out anything sensitive first. Our guide to using AI at work without leaking sensitive data covers what to leave out.
The one habit that matters most
If you take only one thing from this guide, make it this: open the sources. Most AI mistakes that cause real harm would have been caught by someone who clicked through and read the page.
The same rule applies when an AI agent reports back on work it did for you. “Task complete” is a claim, not proof. Our guide to what an AI agent is explains why that matters more as agents take on longer jobs.
Sources
- LawNext, Court imposes sanctions on lawyers who filed bogus cases after relying on ChatGPT for legal research, June 22, 2023
- Wikipedia, Mata v. Avianca, Inc.
- Mike Caulfield, SIFT (The Four Moves), June 19, 2019
- C2PA, the Coalition for Content Provenance and Authenticity
- European Commission, Guidelines on transparency obligations for providers and deployers of certain AI systems
- Morgan Lewis, EU AI Act’s transparency rules: what went into effect on 2 August, August 2026
