AI answers in exactly the same confident tone whether it is right or making things up. That is the real trap: it does not fail like a broken search engine that returns an error, it fails like a very convincing colleague who hands you a round number, complete with a source, that never existed. Learning to verify what it tells you is not distrust: it is the skill that separates people who use AI with judgement from those who end up publishing a false figure with their name on it.
Why AI invents data (and why it sounds so good)
A language model does not query a database: it predicts the most plausible text. Ask it for a statistic on a topic where it has seen thousands of reports and it will produce something shaped like a statistic — a believable percentage, a well-known consultancy, a recent year — even if that exact combination does not exist. This is called hallucination, and the name is slightly misleading: it is not a rare, visible glitch, it is the same mechanism that makes the model work so well, applied to something it does not know.
That is why fabrications have a recognisable pattern: they tend to be too clean. “37% of companies” attributed to a big consultancy, a court ruling with a number and date, a study from a prestigious university. Real data is messier: it comes with caveats, ranges, debatable methodology and years that do not quite line up.
Five signs a figure might be invented
- A round, emphatic number with no range or nuance: “cuts time by 40%”. Reality is usually “between 20% and 60% depending on the case”.
- A prestigious but vague source: “according to McKinsey”, “a Harvard study”, with no specific report, title or link. If you cannot reach the document, treat it as non-existent.
- It fits what you wanted to hear a little too well. Models tend to please; if your question implied a conclusion, it will happily hand it back “supported”.
- Verifiable details you never verify: proper names, dates, product versions, prices. Exactly where it fails most and where checking is easiest.
- Links you do not open. An invented URL looks identical to a real one until you click it.
The three-step verification method
You do not need to check everything: you need to check what matters, and do it fast.
- 1. Separate fact from reasoning. An explanation of how something works can be useful even if it is general. A concrete data point (figure, date, name, price, legal citation) is a verifiable claim. Flag those.
- 2. Look for the primary source, not confirmation. The classic mistake is asking the same AI “are you sure?”: it will say yes, or invent an apology. Go to the product’s official page, the official gazette, the original report. If the fact exists, it lives at home.
- 3. If you cannot find it in two minutes, cut it. This is the rule that saves the most pain. A figure you cannot stand behind does not improve your text: it turns it into a time bomb. Rewrite it qualitatively (“cuts time noticeably”) and move on.
What to always verify, and what you can let go
Verifying 100% is unworkable and unnecessary. Our practical rule:
- Always: figures and percentages, legal citations and regulations, prices and plans, dates, names of people or companies, medical or financial references, and every link.
- Almost always: claims about what a specific tool does or does not do (they change monthly).
- You can let go: conceptual explanations, structure, summaries of your own text, ideas and drafts you will rewrite anyway.
Notice the pattern: what needs verifying is precisely what a reader could check and call you out on. If something can be dismantled with one search, check it before they do.
Prompts that reduce fabrication
They do not eliminate it — no instruction does — but they change the output a lot:
- “If you are not certain about a fact, say so explicitly instead of estimating.” It gives the model permission not to know, which is exactly what it usually avoids.
- “Do not include statistics or figures unless you can cite the exact source.” Kills decorative numeric filler.
- “Answer only from the information in this document.” Anchoring it to text you provide dramatically reduces invention; that is the idea behind RAG systems.
- “Mark every claim that should be checked with [VERIFY].” It half-does your review work for you.
And a tool with live search and citations (like Perplexity) does not excuse you from verifying: it gives you the link, but you still have to open it and confirm it says what the AI summarised.
Our experience verifying AI content
- What has bitten us hardest: “decorative” figures. Sentences that added nothing to the argument but sounded authoritative, with a percentage and a consultancy attached. When we went looking for the report, it did not exist. The rule now is simple: if a number has no findable source, it goes — or becomes qualitative.
- The signal that works best: discomfort. When a data point fits suspiciously well, something is usually off. That twinge is the best detector we have.
- What is not worth it: chasing absolute perfection. Verifying what is verifiable and staying qualitative elsewhere produces more honest — and more useful — writing than stuffing text with numbers nobody has checked.
Our recommendation: use AI to think, structure and draft fast, and save your human time for verification. That is the division of labour that works best. You can see how we apply it in our editorial methodology.
Frequently asked questions
Why does AI give me sources that do not exist?
Because it generates the most plausible text, and a well-formed citation is highly plausible. The model is not copying from a database: it is building something that looks like a reference. That is how you get a citation with a real author, a real journal and a title that was never published.
Does asking “are you sure?” help?
Not much. Sometimes it corrects itself, but it is equally capable of doubling down with more confidence or inventing an apology plus another equally false figure. Verification has to come from outside the model: primary source or nothing.
Do models with web search still make things up?
They fabricate less on current facts, because they read real pages, but they can misread what they find or cite a source that does not support the claim. You gain a lot, but opening the link is still your job.
How do I verify something technical I do not understand?
Go to the product’s official documentation or the regulatory source. If the topic is medical, legal or financial, the right answer is not to verify better: it is to consult a professional. AI can prepare your questions, not deliver the verdict.
Conclusion
AI is an extraordinary assistant and an unreliable witness. Treat it like a brilliant colleague who occasionally makes things up with total confidence: use its speed, but do not sign off anything of its without checking. If you take one idea away: a fact with no findable source is not a fact, it is a liability. And if you want to keep going, you will like the AI myths worth dropping and why AI detectors fail more than you would think.
