How to Spot AI Hallucinations

The five tell-tale patterns

1. Invented citations. A real-sounding study ("a 2019 Stanford study found…") that does not exist. Click every citation — fabricated ones are the most common hallucination in professional use.
2. Too-perfect numbers. Round statistics attached to vague sources ("87% of users report…"). Real numbers have decimal dust and named sources.
3. Confident obscurity. The model is excellent about famous topics and unreliable about niche ones — the exact moment it should say "I'm not sure" is when it sounds most confident.
4. Phantom quotes. Attribution errors: the right person, a quote they never said. Search the exact wording in quotes.
5. Blurred details. Dates, roles and names drift when the model "remembers" past its context window or stitches sources together.

The two-minute verification habit

• Ask for sources up front — "cite where that comes from; say so if you are not sure" — then actually open them.
• Ask again, differently. Rephrase the question; hallucinations wobble while facts hold.
• Use a source-first tool for facts. Perplexity-style answers with citations, or RAG grounded on documents you trust, beat memory-only chat for anything load-bearing.
• Match effort to stakes. Brainstorm = trust and move on. Health, money, law, publication = verify like your reputation depends on it, because it does.