Trust, but Verify: How to Sanity-Check Anything an AI Tells You

An AI's biggest weakness is that it's wrong in exactly the same confident, polished voice it uses when it's right. There's no wobble, no "um," no hedge. So you can't tell truth from fiction by tone. You have to build the check into your own habits. The good news: verifying doesn't mean re-researching everything from scratch. It means learning which answers deserve a second look and how to give them one quickly.

First, sort the answer by risk

You don't need to fact-check that it turned your email into three tidy bullet points; you can see with your own eyes whether that worked. Verification is for claims about the world, not for wording or formatting. Before you spend any effort, ask what type of thing you just got:

• Self-evident output (a rewrite, a summary of text you pasted, a list of ideas): check it by reading it. No outside verification needed.
• A factual claim (a date, a statistic, a law, a medical detail, "studies show," a quote, a price): this is the stuff that gets invented. Verify before you rely on it.
• Advice with real stakes (health, legal, money, safety): treat the AI as a starting point only, and confirm with a qualified human or an official source.

That triage alone saves you hours, because most of what a chatbot produces is the first kind.

The red flags that should make you pause

Certain things are statistically where AI lies most. Train yourself to twitch when you see:

• Specific numbers and statistics. "73% of people" or "the study of 1,200 participants." Precise figures are easy to hallucinate and feel authoritative.
• Named sources, studies, books, or quotes. AI is notorious for fabricating real-sounding citations. If it names a study, author, or article, assume it might not exist until you find it yourself.
• Anything time-sensitive. Current prices, who holds an office, whether a policy or product still exists, opening hours, latest versions. Its knowledge may be out of date.
• Small, local, or niche facts. A specific clinic's hours, a minor local ordinance, a little-known person. Less common information is where it guesses most.
• Anything that sounds too neat. A perfect quote perfectly attributed, a clean historical anecdote that's a little too on-the-nose. Neatness can be a manufacturing sign.

Quick checks that take under a minute

You don't need to become a researcher. You need two or three fast reflexes.

The independent-source check

Take the specific claim and search it in a regular search engine or on the source that would actually know. Confirming a law? Look for the government or official site. A study? Search the exact title in quotes, or look on the journal or the university's page. A price or a business fact? Go to the company's own site. The point is to land on a source that isn't the AI. If you can't find the claim anywhere independent, that's your answer: don't trust it yet.

The made-up-citation check

When AI gives you a source, copy the exact title and paste it into a search. Real sources show up immediately with matching details. Fabricated ones return nothing, or return a different real thing the AI mashed the name onto. This takes fifteen seconds and catches one of the most common failures cold.

The self-consistency probe

Ask the same question again in a fresh conversation, or word it differently. If you get a meaningfully different answer the second time, at least one of them is unreliable, and now you know not to trust either without checking. Consistent answers aren't proof of truth, but flip-flopping is strong proof of a guess.

Make the AI help you check itself

You can turn the tool's own habits against its weaknesses. Try these follow-ups on any answer that matters:

• "How confident are you in this, and which parts are you least sure about?" It will often flag the shaky bits itself.
• "What would I need to verify this independently? Point me to the kind of source that would confirm it."
• "Is this current as of today, or might it be out of date? Did you actually look it up?"
• "Give me the strongest argument that this answer is wrong."

These aren't magic; a determined hallucination can survive them. But they surface a lot of soft spots, and they cost you one line of typing.

One trap to avoid: the confidence loop

Here's a mistake I see constantly. Someone doubts an answer, asks the AI "are you sure?", and the AI doubles down with even more detail and reassurance. That extra confidence is not extra evidence. The model is generating a more persuasive version of the same possibly-wrong claim. An AI reassuring you about an AI's answer is not verification. The check has to come from outside the conversation.

Match your effort to the stakes

Don't exhaust yourself. Calibrate:

1. Low stakes (trivia, casual curiosity, a draft you'll edit): trust it, move on, fix it if you notice something off.
2. Medium stakes (facts going into your work, plans for guests, a decision that costs some time or money): spend the sixty seconds on an independent check of the key claims.
3. High stakes (health, legal, financial, safety, anything public or hard to undo): use the AI to understand and to prepare questions, then confirm with a real professional or an official primary source. Never let it be the last link in the chain.

The habit worth building

You already do a version of this with people. You take a chatty acquaintance's restaurant tip at face value and you double-check a stranger's claim about your tax deadline. AI deserves the same instinct, tuned a little tighter because it never sounds unsure. The whole discipline fits in one sentence you can keep in your head: if being wrong would cost me, I confirm it somewhere that isn't the AI. Get that reflex, and you can use these tools heavily without getting burned by them.