How to Fact-Check AI Answers and Catch Hallucinations

A while back I asked a chatbot for a court case to back up a point I was making. It gave me one. Real-sounding name, a year, a one-line summary of the ruling. I almost used it. Then I went to look it up and — nothing. The case didn't exist. The model had invented it, and it had done so in the exact same calm, tidy tone it used for everything else.

That's the thing that actually trips people up. It's not that AI gets things wrong. Everything gets things wrong. It's that AI gets things wrong with zero change in confidence. No nervous "I think", no hedge, no tell. A made-up citation and a real one come out formatted identically. So this post is about closing that gap — why it happens, the moments worth a second look, and a quick habit that lets you keep the speed without inheriting the mistakes.

Why it makes things up

A model like Gemini or GPT is, underneath everything, a very good next-word predictor. It learned the statistical shape of language from a mountain of text, and when you ask it something it generates the most plausible continuation. Not the truest one. The most plausible one.

Most of the time those are the same thing, because the training text was mostly accurate. But the model isn't looking anything up. There's no internal database, and — this is the important part — no built-in sense of "I don't actually know this." So when it hits a gap, an obscure fact or a specific citation or something that happened last week, it doesn't stop. It fills the gap with something that fits the pattern of a right answer. That invented-but-plausible output is what people call a hallucination.

A few things make it more likely:

• Precise details it can't recall — exact dates, statistics, page numbers, quotes, citations, version numbers.
• Niche or recent topics that barely showed up in training data, if at all.
• Leading questions. Ask "What study proved that X?" and the model may hand you a study rather than tell you none exists.
• Long, multi-step reasoning, where one wrong assumption early on snowballs into a confident wrong conclusion.

None of this makes AI useless. It makes it a knowledgeable friend who's usually right and occasionally, fluently, wrong. Treat it like that and you'll be fine.

When to get suspicious

You don't need to verify every word — that would defeat the point of using AI at all. You need to notice the moments that deserve a check. These are the ones I've learned to watch for.

Over-confident specifics. Be most skeptical exactly where the answer is most precise. "The law passed in 2017", "the study had 4,213 participants", "see page 84." That level of detail is where hallucinations love to hide, because precise-sounding things feel authoritative even when they're fabricated.

Citations, quotes, and sources. This is the big one. Made-up references are one of the most common and well-documented AI failures, and they're convincing — a real-looking author, a plausible journal, a believable year, all completely fictional. So I treat every citation, book title, URL, or direct quote as unverified until I've found it myself. No exceptions.

Anything time-sensitive. Prices, current office-holders, "the latest version", recent news, who won last season. A model's knowledge has a cutoff, and even inside that cutoff it can be stale. If the answer hinges on the present moment, assume it might be out of date.

Suspiciously tidy answers to messy questions. Genuinely contested topics rarely have clean, symmetrical answers. If you ask something that's actually debated and get back a confident one-sided verdict, that smoothness is itself the flag.

Math and counting. Models are better at arithmetic than they used to be, but they still slip on multi-step calculations. Anything load-bearing — a budget, a dosage, a deadline computed from a date — I re-check by hand.

How I actually verify

Here's the workflow. You won't run all five every time. Pick based on what's at stake.

1. Ask for sources, then go find them

Make the model show its work:

For each factual claim in your answer, tell me how confident you are and what kind of source would confirm it. If you are not sure, say so.

This doesn't make the AI honest by magic. What it does is surface the shaky parts and hand you a list to check. The check is the point — if it names a study or an article, search for that exact title. If it doesn't exist, you just caught a hallucination before it cost you anything.

2. Cross-check with a real search

For anything that matters, open a search engine and confirm it independently. The fastest version: copy a specific claim out of the AI's answer and search for it verbatim. Multiple reputable sources agreeing? You're probably fine. The only "source" being the AI's own phrasing echoed nowhere else? Careful.

A model is a great starting point — it tells you what to search for and which terms to use. It's just not the final word.

3. Make it critique itself

LLMs are often better at spotting flaws than avoiding them, which you can use. After you get an answer, push back:

Review your previous answer critically. What are you least confident about? What could be wrong, outdated, or made up? List anything a careful fact-checker should double-check.

You'll watch it walk back a claim it stated confidently thirty seconds earlier. That's not a gotcha — it's a free verification pass. Asking for two independent answers to the same factual question and comparing them works the same way. If they disagree, at least one is wrong, and now you know to look.

4. Poke at the specifics

When a precise detail looks shaky, interrogate it directly:

You said this happened in 2019. How sure are you about that exact year? Could you be confusing it with a similar event?

Real facts hold up under questioning. Hallucinations tend to wobble — they get revised, or suddenly come with a disclaimer that wasn't there before. The wobble is the answer.

5. Triangulate against what you know

You're a fact-checker too. If an answer contradicts something you're genuinely sure about, trust yourself and dig in. A surprising amount of the time, hallucinations fall apart against plain common sense or a single detail you happen to know cold.

A two-minute example

Say you ask: "What are the main provisions of the Smith-Hartley Data Act?"

The lazy version is to read the confident three-paragraph summary, paste it into your report, and move on. The problem is that the act might not exist at all — the model could've stitched together a plausible-sounding law out of fragments of real ones.

The version that takes two minutes:

1. Read the answer, but flag every specific in it — year passed, agencies named, penalties listed.
2. Search the exact name "Smith-Hartley Data Act" in a real search engine.
3. If nothing reputable comes up, ask the model directly: "Are you certain this act exists, or might you be confusing it with another law? How confident are you?"
4. Only use the details once an independent source confirms them.

Two minutes. And it's the difference between citing a law and publishing fiction with your name attached.

Where I don't trust it at all

Some categories don't get a quick check — they get a qualified human.

• Medical, legal, and financial decisions. Use AI to understand the terminology and frame your questions. Never as the final authority. The actual decision goes to a doctor, lawyer, or licensed professional.
• Anything you'll publish or send to others as fact. Your name is on it. Verify it.
• High-stakes numbers — dosages, tax figures, structural calculations, contract amounts.
• Live or local information the model has no reliable way to know.

This isn't anti-AI paranoia. It's the same standard you'd already apply to advice from a smart, well-read acquaintance who occasionally misremembers things.

Match the skepticism to the stakes

The goal isn't to distrust everything an AI tells you. That throws away most of what makes it useful. The goal is to match your skepticism to the stakes. Brainstorming, drafting, explaining a concept you'll sanity-check anyway? Let it run. Facts you're going to rely on? Spend the two minutes.

Better prompting helps on the front end, too — clearer, more specific questions tend to produce fewer wild guesses. If you want to sharpen that side, our guide on writing better AI prompts pairs naturally with all of this. Good prompts and good fact-checking are the two halves of using AI without getting burned.

You can try every one of these for free, no signup, on Smillee AI. Ask it something you already know the answer to and just watch how it behaves. Then run the self-critique prompt above and count how often it revises itself. Honestly, that one exercise taught me more about where AI's limits actually are than any warning label ever did.

Treat it as a fast, fallible assistant instead of an oracle and it turns out to be genuinely useful and rarely dangerous. The habit is the whole thing.

— Maya