Most advice about AI errors stops at "be careful, it makes things up." That's true and useless. What you need is a repeatable routine — something you run in two minutes, the same way every time, so you don't have to decide from scratch whether an answer is trustworthy. This page gives you that routine: claim → source → date → action. Four steps, in that order, applied only to the parts of an answer that actually matter.
Want the same path as a short course? Open AI Earn on Google Play — that listing is the official app page.
Step 1: Pull out the claim
An AI answer is usually a mix of three things: reasoning you can judge yourself, wording you don't care about, and specific factual claims. Only the third category needs checking. So start by rewriting the risky parts as short, standalone statements a stranger could verify — "the deadline is 31 January", "this law applies to companies with more than 50 employees", "this tool has a free tier that includes image generation". If a sentence can't be turned into a checkable statement, it's an opinion or a suggestion, and fact-checking it is a waste of time. Vagueness is a signal in itself: an answer full of "many experts", "studies show", "generally around" is often a sign the model has no specific fact underneath and is smoothing over the gap with confident filler.
Step 2: Find a source that isn't the AI
Ask the assistant where the claim comes from, but treat its answer as a lead, not as proof. Models can produce citations that look perfect and don't exist — right-sounding title, plausible author, invented link. Two rules make this manageable. First, open the source yourself; never accept a citation you haven't clicked. Second, check that the source actually says what was claimed, not just that it's about the same topic. A real page about tax deadlines is not evidence for a specific date on that page.
Prefer sources where the claim originates: the vendor's own help pages for tool features and pricing, the government or regulator's site for rules and deadlines, the original publisher for numbers. For anything about ChatGPT, that means help.openai.com; for Gemini, Google's Gemini help. A blog summarizing a vendor announcement is one degree removed and often outdated.
Step 3: Check the date on everything
This is the step people skip, and it's where most "the AI lied to me" stories actually come from. Language models are trained on text up to a cutoff, and a chatbot may or may not be able to search the live web in the session you're in. Anything that changes — prices, plan limits, model names, laws, opening hours, who runs what — can be described accurately as it was two years ago and be wrong today. The AI won't flag the difference, because it has no sense of what has changed since it was trained.
So: check the date on the source page, not just the claim. Then ask yourself whether this is a fact that expires. Names of capital cities don't expire. Subscription features expire fast. When the answer matters and it's the kind of fact that moves, the vendor's current page beats any explanation, including a very good one.
Step 4: Decide the action
Verification isn't binary. Once you've looked, put the claim into one of four buckets and move on — this is what stops fact-checking from eating your afternoon.
- Confirmed. Source found, source says it, source is current. Use it.
- Corrected. The real number or date is close but different. Fix it and continue; don't argue with the chat about it.
- Unverifiable. No independent source exists. Cut it, or rewrite so nothing depends on it.
- Doesn't matter. Wrong wouldn't cost you anything. Leave it, knowingly.
| Type of claim | What goes wrong | Minimum check | Rough time |
|---|---|---|---|
| Brainstorm, draft wording, ideas | Nothing — you edit it anyway | None | 0 min |
| How a tool works, plan limits, prices | Confidently describes an older version | Vendor's own help or pricing page, today | 2 min |
| Numbers, statistics, percentages | Plausible figure with no real origin | Trace to the original publisher | 5 min |
| Quotes and attributions | Real person, invented sentence | Find the exact quote in a primary source | 5 min |
| Legal, tax, medical, financial | Right in general, wrong for your country or case | Official source plus a qualified human | Longer — and don't skip the human |
| Recent events, "latest" anything | Knowledge cutoff or stale search results | A dated source from this month | 3 min |
| Code, formulas, calculations | Looks right, fails on real input | Run it; test one case by hand | 2–10 min |
Prompts that make the ritual faster
You can push some of the work back into the chat. These don't replace checking — a model auditing itself is still the same model — but they surface the weak spots quickly so you know where to look.
Prompt: "Go back over your last answer. List every factual claim in it as a separate line. For each one, mark it as: verifiable from a public source, an estimate, or something you're unsure about. Don't rewrite the answer, just give me the list."
Prompt: "For this claim — [paste the claim] — tell me exactly what I should search for to confirm or disprove it, and which type of source would settle it. Give me the search terms, not the answer."
Prompt: "Answer [question]. Then, separately, tell me: is this the kind of fact that changes over time? When would you expect your information to be out of date, and what should I check against the vendor's current page?"
Ask "what would make this answer wrong?" It's a better question than "are you sure?", because a model will apologize and agree with almost any pushback — including pushback that's mistaken.
What not to do
- Don't ask the same chatbot to confirm itself. "Are you sure?" produces either a confident yes or a needless retraction. Neither is evidence. Go outside the chat.
- Don't accept a link you haven't opened. A well-formed URL is not a working URL, and a working URL is not a page that supports the claim.
- Don't fact-check everything. If you verify brainstorming output, you'll quit the habit within a week. Triage first — most answers need nothing.
- Don't argue with the model to find the truth. Pressing it usually just changes its answer. Changed answers feel like progress and aren't.
- Don't treat "I searched the web" as verification. Search results can be old, wrong, or misread. Open what it opened.
- Don't rely on it for your own numbers. If a figure came out of your business, your bank, or your spreadsheet, the source is your record, not a chat window.
Where this fits with everything else
The routine gets easier once you understand why the errors happen at all: a chatbot generates text that fits the pattern of a correct answer, which is not the same as retrieving a correct answer — see how AI chatbots actually work. Clearer prompts reduce the number of claims you have to check in the first place, because a specific request produces fewer invented details than a broad one; writing better prompts covers that. Trusting confident output is also the single most common beginner error — it's on the list in common AI mistakes beginners make.
Two neighbouring habits matter as much: being careful about what you paste in (using AI safely) and recognizing pitches that use AI as a costume (spotting AI scams). And if a term in an answer is doing more work than you understand, look it up in the AI glossary before you decide whether the claim is even meaningful.
The two-minute version
When you're in a hurry: underline the claims you'd be embarrassed to get wrong. For each one, find a non-AI source, look at the date on it, and either keep, fix, or cut. That's the whole thing. It won't catch every error, and it doesn't need to — it catches the ones that cost you money, credibility, or a bad decision, which is the only category that ever mattered.
