PPROMPT WORKBOOK Build a prompt
Plausible is not the same as true

AI hallucinations: why models guess

A language model can produce fluent text without evidence. Prompting can reduce the risk, but it cannot replace verification.

What is a hallucination?

It is a relevant-looking, confident output containing something incorrect or unsupported: a fabricated source, a quotation that was never said, a wrong number or an inference presented as fact.

Four layers of defence

  1. Supply sources. Ask the model to work from bounded material rather than general model memory.
  2. Allow no-answer. Require it to say when the sources do not cover the question.
  3. Demand traceability. Require a source ID, page or direct link for each important claim.
  4. Verify outside the model. Asking the same model to double-check is not independent verification.
Answer only from the material I supplied.

For every important claim:
1. give the source and exact location
2. label it as directly documented or inferred
3. state what is missing if certainty is limited

If the material does not support an answer, write "not documented".
Do not invent links, quotations, numbers or names.

What does not guarantee truth

Self-reported confidence, a persuasive rationale or agreement between several models does not prove a claim. Structured output only guarantees shape. Use the original source as ground truth when the stakes matter.

For the complete workflow, continue with research and fact-checking and verify before you cite.

Sources