Limits and hallucinations: what not to trust

You can see by now how smoothly an AI answers. Time for the flip side: sometimes it makes things up just as smoothly. Knowing that isn't a reason to stop using it — it's the condition for using it like an adult.

What a hallucination is

A hallucination is a confident, polished, but false answer: an invented quote, a link to nothing, a mixed-up date, a made-up fact. The nasty part is that it sounds exactly as convincing as the truth — no hesitation, no caveats. This is a shared trait of all three models, not one model's disease.

Why it lies so confidently

An AI doesn't "know" facts the way a reference book does — it predicts a plausible continuation of text. Most of the time plausible matches true, but where the data ran thin the model still completes a smooth sentence — only now it's wrong. It has no internal "I'm not sure" until you ask for one. Hence the rule: a confident tone is not evidence of being right.

The knowledge cutoff: why it "doesn't know about yesterday"

Every model has a knowledge cutoff — the point up to which it saw the world's text. Anything that happened later it doesn't know on its own and may invent. That's exactly why an all-rounder without search access shouldn't be answering "what's the price today" or "who won last night" — that's a job for Gemini or Perplexity with fresh search (remember the "task → AI" table).

Where all three fail the same way

Some topics trip up the whole trio, and knowing them is more useful than comparing the models with each other.

• Exact numbers and dates — statistics, rates, years: they like "roughly right".
• Links and quotes — they can invent a plausible source that doesn't exist.
• Law and medicine — it sounds authoritative, but the responsibility is yours; these are topics for a human specialist.
• Recent events — everything after the knowledge cutoff, without search.
• Big mental arithmetic — long calculations are better asked for step by step, or done in a calculator.

A case: the half-million-dollar report with invented sources

If you think only lone amateurs get caught, here's 2025. Deloitte — one of consulting's "big four" — delivers a report to the Australian government worth about 440,000 Australian dollars. A sharp-eyed university researcher starts checking the footnotes and can't find the sources: some references and a quote from a court ruling turn out to be invented, with the trail leading to an AI. A scandal, a corrected report, part of the fee refunded. Note who got caught: not students — consultants who are paid precisely for verified facts. AI is a draft, not a source of truth, and checking is a human's job — at any price.

Where hallucinations barely matter

It's important not to swing to the other extreme — "if it lies, nothing can be trusted". There's a huge zone where invention does little harm: drafts, ideas, phrasings, explanations of general concepts, brainstorming. If you ask for ten headlines or a rewritten paragraph, there are no "facts" to check — you'll judge the result with your own eyes anyway. Switch caution on where the answer travels beyond you as a fact: a number in a report, a link in an article, an argument in a debate, advice about your health or the law. The rule is simple: the more serious the consequences of an error, the stricter the check.

Three habits that save you

You can't remove the invention entirely, but you can bring the risk down to something acceptable. Here's how.

1. Ask for sources and open them. "Give me a link for every claim" — then click through. No link, or a dead one, means it's a hypothesis, not a fact. This is exactly what Perplexity was built for.
2. Cross-check what matters with a second model. Same question to two AIs: if they agree, it's more likely true; if they diverge, go dig yourself.
3. Give the model permission to say "I don't know". Add to your request: "if you're not sure, say so, don't invent." The model fills gaps less often that way — not always, but often enough to be worth the extra line.

Common myths

Three myths about hallucinations — the last thing to clear up before the checkpoint.

• "The paid version doesn't hallucinate." Top models err less often, but they don't stop: you still verify.
• "It answered confidently, so that must be how it is." Confidence of tone and being right are different things; the model doesn't doubt out loud unless you ask it to.
• "Hallucinations are a bug they'll fix soon." They're a property of the "predict the continuation" approach, not a breakage; they'll get rarer, but they won't retire your habit of checking.

Do this now

A small trust test, three minutes. Ask your AI something checkable from your own field where you know the right answer, and ask it for a source. Judge it: is everything accurate, does the link lead to a real page? One experience like that calibrates your trust better than ten warnings — this one included.