No prompt guarantees factual accuracy. Supply relevant source material, allow the model to abstain, request evidence, and verify consequential claims against the source. For repeated tasks, measure errors and unsupported answers with an evaluation set.
Verified Oct 18, 2026Review by Jan 16, 20272 sources
What it is
This question is about hallucination: confident but false output. It matters most in client-facing copy, reports and support answers.
How it works
These steps follow Anthropic’s published techniques, plus testing:
Ground it. Paste or connect the documents the answer should come from (RAG).
Allow abstaining. Tell the model it may say when the documents don’t contain the answer.
Ask for evidence. For long documents, have it extract word-for-word quotes first and answer only from them. Ask for a quote or source for each claim, and tell it to retract any claim it can’t support. Verify consequential claims against the original source; a generated quotation can itself be wrong.
Define the scope. State the specific question, permitted sources and what to do when the evidence is missing. Measure whether that change helps your task.
Test. For recurring work, build a small eval with known answers.
Benefits
A process for finding unsupported claims before publication.
Evidence beside each claim can make review easier; measure the review time.
A clear standard your team can follow.
Limits and risks
Grounding reduces errors but doesn’t eliminate them. Models can misread sources.
Fabricated citations can look real, so open them.
High-stakes topics (legal, medical, financial) need expert review regardless.
Include missing evidence, conflicting sources and outdated documents in the test set. Track both wrong answers and unnecessary refusals. A model checking its own answer is not independent confirmation.
Practitioner views are restated in our words and labeled as views, not measured results. Product facts come from official sources dated above.
Suggest an edit
IonBin's wiki isn't user-edited. Corrections go to our Directory Curator, who checks them against primary sources before anything changes. Spot something wrong or out of date on “How do I stop AI from making things up?”? Email us with a link to a primary source.