Hallucinations & when to trust AI

What a hallucination really is

When AI states something false with total confidence, we call it a 'hallucination'. But the model isn't lying — lying needs intent. It's doing the only thing it knows: predicting the most probable-sounding next words. Sometimes the most probable-sounding answer is factually wrong.

The danger is that a hallucinated fact arrives wrapped in the same confident, fluent prose as a true one. You can't tell by tone. The model doesn't know what it doesn't know — it will happily invent a reference, a date, or a quote and present it as fact.

When to trust it — and when not

Here's a simple, usable rule. Trust AI freely for form; verify it for facts.

• ✅ Safe to trust (form): drafting, brainstorming, summarising, rewriting, formatting, tone, structure, ideas.
• ⚠️ Verify every time (facts): numbers, dates, statistics, legal or medical claims, quotes, citations, anything you'll act on or publish.
• 🔴 Never trust: anything that could cost you money, health, reputation, or a client — without checking a real source.

Build the habit

The professionals who use AI well don't have better tools — they have a better reflex. They treat AI output as a first draft, always. Draft with AI, decide with your own judgement, and check the facts against a source you can actually open.

One trick: when an AI gives you a source or citation, actually click it. You'll quickly learn how often it's made up. That single habit — verifying — is worth more than every prompt hack combined.