Spotting a hallucination

A hallucination is text that sounds plausible, reads confidently, and is formatted perfectly, but is factually wrong. The dangerous part: the AI does not know it is wrong. It generates what sounds right, not what is right. That is worse than a lie, because a liar at least knows the truth.

Eight red flags

1. It never says "I do not know."A confident answer for everything is the warning sign. Ask "what are you uncertain about?" A good system names specifics; a hallucinating one stays perfectly confident.
2. It cannot give specific sources."Studies show," "research indicates," "experts agree." Ask which study, published where, what citation, then verify it.
3. Overly specific without caveats."87.3% effective," "exactly 750mg twice daily." Real medical information carries ranges and individual variation. False precision is a tell.
4. Internal contradictions.It says avoid a drug, then later calls it safe. Ungrounded text loses consistency. Ask the same question more than once.
5. Fabricated citations.Articles never published, authors who did not write them, studies with invented data. Check PubMed or Google Scholar; confirm the source exists and says what it claims.
6. Sounds too perfect and complete.Real medicine has gaps and controversies. Ask "what is controversial here?" or "what don't we know?"
7. Cannot explain its reasoning.Ask "why?" A good answer walks the logic; a hallucination cannot, or the logic does not follow.
8. Novel combinations of real things.A real drug joined to a fabricated indication. Example: "Metformin is FDA-approved for migraines." Each piece is real; the connection is false. Verify the relationship, not just the parts.

Verify before you trust

Source check: are sources specific (journal, year, authors)? Can you find them? Do they say what the AI claims? Cross-reference: Mayo Clinic, NIH/MedlinePlus, CDC, specialty-society sites; do reliable sources agree? Consistency: ask the same question again; any contradictions? Uncertainty: "what are you uncertain about?" Logic: "explain your reasoning," and check that it holds.

The "I don't know" test

Ask questions with progressively less information available remotely. A trustworthy system grows more uncertain as you climb:

• "What is hypertension?" confident, accurate.
• "Treatment guidelines for stage 2 hypertension?" confident, with sources.
• "How should it differ for an 80-year-old with multiple conditions?" some uncertainty, acknowledges individual variation.
• "What is the optimal target for my situation?" should defer: needs your full history and a physician.
• "Do I have hypertension?" should refuse: cannot diagnose without measuring you.

If it answers the last two with confidence, it is hallucinating.

Three dangerous real-world patterns

Reduce the risk

Prefer AI built on a validated knowledge base, content-controlled tools limited to verified medical sources that can say "I do not know" when a question falls outside them (this is the idea behind TheDude, built on StatPearls). Ask the five essential questions. Verify before trusting, cross-reference, and be suspicious of perfection.