☕ aban news
Updated: May 2026 · no panic

Spotting AI hallucinations

AI makes up facts — confidently and convincingly. That's not a defect, it's the flip side of how it works. Here are the warning signs and the 30-second check that protects you.

Why AI hallucinates at all

A language model is at its core a machine that predicts the next word. It doesn't compute the truth, but the probability. That explains almost everything: why it phrases things brilliantly and invents facts in the same breath, why it never notices its own error, and why the same question gives slightly different answers twice.

A hallucination, then, isn't a glitch but the logical consequence of that. An invented quote is statistically likely — even if it never existed. Once you grasp this, you treat every AI answer as a draft, not as evidence.

The AI is an intern with perfect memory and zero judgement. It always sounds certain — whether or not it's right.

Five warning signs of made-up facts

1 Overly specific details with no evidence

The more precise a claim sounds without a source attached, the sooner you check it. "Studies show that 73% …" with no study is a warning sign, not a fact.

2 Named sources, studies or statutes

Named sources look credible — and are especially often invented. Models generate something that looks like a source. Open every cited study, ruling and statute yourself.

3 Exact numbers and dates

Round, confident numbers and precise years sound reliable but are easily invented. Treat every number that matters as unverified until you've seen it.

4 Internal contradictions

If the answer says X in one place and the opposite in another, at least one place is guessed. Have the same point explained twice and compare.

5 Answers about things that don't exist

Ask about a book or a law that doesn't exist — many models will happily describe it anyway. If the AI never says "I don't know that one", suspicion is warranted.

The 30-second check

  • Verify at the source: whatever has to be correct, you check at the original — not with the AI and not in a blog post.
  • Demand sources: "Name the source and quote the passage verbatim." Invented claims often collapse here.
  • Allow uncertainty: say in the prompt: "If you're not sure, say so." That noticeably reduces inventing.
  • Cross-check: ask the same question rephrased. If the answers differ a lot, at least one is guessed.

Where hallucinations are especially expensive

Not every mistake costs the same. A clumsy headline you fix in seconds — a wrong number in a quote or an invented statute in a client letter you don't. Treat these areas as check-mandatory by default:

  • Law: statutes, deadlines, rulings — always cross-check against the legal text.
  • Health & finance: dosages, amounts, tax rules — never adopt unchecked.
  • Quotes & sources: anything you pass on as evidence, open it first yourself.
  • Numbers going outward: anything that goes to clients or into publications.
Rule of thumb: the more expensive a mistake, the more your own eyes belong on it. The AI prepares — you approve.

Frequently asked

What is an AI hallucination?
A made-up but confidently delivered answer. Not a defect, but the flip side of how the model works — it computes probability, not truth.
Why does the AI invent sources?
Because a plausible quote is statistically likely, even if it never existed. The model generates something that looks like a source. Open every cited source yourself.
How do I spot a hallucination?
Watch for overly specific details with no evidence, exact numbers, named studies and internal contradictions. The more confident without a source, the sooner you verify.
Are newer models free of it?
No. They hallucinate less often, but not never — which makes them more convincing. The checking routine stays the same.

Free eBook: “Anti-Hype”

This way of thinking runs through the whole eBook: the 3-question filter, your minimal AI stack and how to separate substance from show. No sign-up.

Read the eBook →

Every workday: AI without the bullshit, in 5 minutes

aban news brings you the AI developments that matter for your work, Mon–Fri — honestly assessed, without panic and without hype.

Get the 5-min AI briefing →

No spam. Unsubscribe in one click. GDPR-compliant.

Read on

How to write AI promptsBetter prompts in 4 steps ChatGPT vs. ClaudeThe honest comparison AI tools for the self-employedThe honest stack 2026 Free eBook “Anti-Hype”Use AI without the bullshit

Honesty note: this guide contains no affiliate links. Hallucinations can't be switched off, only contained — the checking routine described here doesn't replace professional verification in an individual case. For legally or medically sensitive topics: verify at the primary source or ask a professional.