How to Check AI Answers: Hallucinations and Practical Controls

Anyone who uses AI every day at work knows the feeling well: the answer arrives written with total confidence, well formatted, in a professional tone — and sometimes it is simply wrong. This is called a hallucination: the AI generates a plausible but false piece of information, with no visible sign that distinguishes it from a correct answer. It isn't a flaw you fix by picking a better model: it's a structural feature of how these systems work, and it has to be managed with a verification method, not blind trust.

What a Hallucination Is and Why It Happens

An AI model doesn't consult a database of verified facts when it answers: it generates the most statistically likely sequence of words given the question. In most cases this produces correct answers, because correct information is also the most frequent in the data it was trained on. But when a question involves a very specific detail, a recent regulation, an exact number or a precise source, the model may 'fill in' the missing information with something plausible instead of signalling that it doesn't know. The result has the same confident tone as a correct answer, and that is exactly what makes it dangerous.

The 3-Step Check Before Using an Answer

There's no need to distrust everything the AI writes: what's needed is a quick check, proportional to the risk, applied before copying an answer into a document, an email to a client or a professional decision.

• Step 1 — Isolate the verifiable facts. Separate the parts of the answer that are opinions or general reasoning from the parts that are specific facts: numbers, dates, names of regulations, citations, references to external sources. Only the latter need checking.
• Step 2 — Ask for the source. Explicitly ask the AI where that information comes from. If it answers with a vague, generic source, or one you can't find anywhere else, that's a warning sign, not a confirmation.
• Step 3 — Verify outside the chat. For each specific fact identified in step 1, do an independent search: an official body's website, the text of a regulation, a supplier's page. If you can't confirm it within a few minutes, treat that data point as unverified.

The Questions Most at Risk of Hallucination

Not all requests carry the same risk. AI tends to make more mistakes on: precise regulatory references (a law's exact number and date, a specific article), bibliographic or case-law citations, very specific numerical data (percentages, amounts, deadlines), and recent facts that postdate its training. On the other hand, it's generally reliable on reasoning, rephrasing a text you provide yourself, summarising a document you've given it directly, and explaining well-established general concepts. The practical rule: the more an answer depends on something the AI has to 'recall' on its own, the more it needs checking; the more it depends on a text you gave it yourself, the more reliable it is.

Practical Tricks to Speed Up Verification

A few habits cut down the time spent checking without reducing safety. First: whenever possible, give the AI the document or source yourself and ask it to work only from that, instead of letting it 'recall' information on its own — the risk of hallucination drops sharply when the reference text is in the prompt. Second: for questions about regulations or deadlines, ask the same question twice in two separate conversations and compare the answers; inconsistent answers are a clear sign the data isn't solid. Third: use the AI to find out where to look, not as the sole source — often the most useful answer is 'check section X on the tax authority's website', not an exact figure stated directly in the chat.

What to Do When You're Not Sure

If doubt remains after the check, the simplest solution is also the most underrated: write 'to be verified' in the final document, or leave that specific detail out until you have a solid source. A text with one fewer fact but all of them correct is far more professional than a complete text with an error a client or colleague later discovers. Anyone who handles a lot of professional communication and wants to avoid starting from scratch every time will also find it useful to organise their verification prompts into a reusable collection, as described in our article on building a prompt library for your practice.

It's also worth remembering that fact-checking is only part of the caution needed: equally important is being careful about what information, even if correct, you share with a generic chatbot. Our deep dive on what not to feed to chatbots covers this complementary aspect. And if you need a ready-made starting point to correctly set up your requests to AI in daily work, you can generate one with the prompt generator instead of writing every prompt from scratch.

Frequently Asked Questions

Can AI hallucinations be eliminated completely?

No, they are a structural feature of current models. They can be greatly reduced by giving the AI sources directly in the prompt and always checking specific facts, but they can't be eliminated 100%.

Do newer or more expensive models hallucinate less?

They tend to hallucinate less on generic questions, but the risk stays the same on very specific or recent data: verification remains necessary regardless of the model used.

How much time does this check really take?

For most professional cases, just a few minutes: the time to isolate the specific facts and do a confirmation search. That's negligible compared to the risk of an error in a public document or one sent to a client.