Prompts & writing
AI Hallucination Risk Score
Six questions about the task you are handing an AI, and an honest read on how likely the answer is to be confidently wrong.
If nobody can tell when it is wrong, wrong answers survive.
Training data ends somewhere. Recent facts are the first to be invented.
Fabricated citations and statistics are the classic failure.
Thin coverage means the model fills gaps with plausible invention.
Grounded answers hallucinate far less than open-memory answers.
Stakes do not change the error rate, but they change how much checking you owe.
Hallucination risk
50/100 · High
Treat the answer as a first draft. Verify every number, name and citation before it leaves your hands.
Cut the risk
- Paste the source documents into the prompt and tell the model to answer only from them.
- Ask for every figure and citation to be listed separately so each one can be checked.
- Use a model with web search, or supply the current facts yourself.
- Ask the model to state its confidence and flag anything it is unsure about.
- Ask the same question twice in fresh chats and compare — disagreement signals invention.
Why models invent things so convincingly
A language model predicts the next plausible piece of text. When it has read a thousand sources on your question, the most plausible continuation is usually the true one. When it has read almost nothing — a niche vendor, last month's release, an internal process — the most plausible continuation is a well-formed invention that reads exactly like the truth.
That is why fabricated citations look real and made-up statistics land on suspiciously round numbers. The fix is never "ask it to be accurate" — it is to supply the source material, demand the specifics separately, and check the handful of claims that matter.