A Verification Prompt That Helps Reduce AI Hallucinations
No prompt can guarantee that an AI answer is true. A more useful goal is to make unsupported claims easier to spot: require the model to identify what is known, what is inferred, what needs a current source, and what it cannot verify. OpenAI's research describes hallucinations as a persistent problem in which a model can produce plausible but false statements, and it argues that systems should be rewarded for appropriate uncertainty rather than guessing.
Use this prompt when accuracy matters more than fluency
Answer the task, but do not guess to fill missing information.
For each important factual claim:
1. Mark it as Confirmed, Inference, or Unverified.
2. If the fact can change over time, say what current source should be checked.
3. If a source is provided, use only information that source actually supports.
4. If sources disagree, describe the disagreement instead of choosing silently.
5. At the end, list the claims I should verify before acting.This prompt does not “turn off hallucinations.” It changes the requested output so uncertainty and verification work are visible to the reader.
Separate three kinds of statements
- Confirmed: directly supported by a source available in the task.
- Inference: a conclusion drawn from supported facts, but not stated directly by the source.
- Unverified: plausible information for which the model does not currently have sufficient evidence.
The labels are useful only if you still inspect the evidence. A model can mislabel its own claim, invent a citation, or misunderstand a source.
Use a source-first workflow for current facts
- Define the question narrowly.
- Find the primary source for facts that can change, such as product limits, policies, prices, schedules, or account settings.
- Give the model the source or require links to the source.
- Ask for the answer and the specific evidence supporting each important claim.
- Open the source and verify the claims that affect a decision.
If you are comparing versions of a document, How to Compare Two Documents with AI and Verify the Differences shows how to keep findings tied to the originals.
An illustrative before-and-after example
Suppose the question is: “Does Product X support feature Y on every plan?” A weak answer might confidently say yes because the feature appears in a general product page. A better verification output would say that the general page confirms the feature exists, but the plan-level availability is unverified until the current pricing or plan documentation is checked. This is a constructed example, not a report of a real product test.
Ask for evidence at the right granularity
“Provide sources” is too vague when the answer contains ten claims. Ask the model to pair each decision-relevant claim with the source section, page, or URL that supports it. For an uploaded document, ask for page numbers or section headings and compare those locations with the original.
Do not use confidence wording as proof
A model saying it is “90% confident” does not establish a fact. Confidence language can help prioritize review, but the evidence still matters. For high-impact medical, legal, financial, safety, or contractual decisions, use qualified professional review and authoritative sources instead of relying on this prompt as a safeguard.
When the prompt is not enough
- The source itself may be outdated.
- The model may cite a source that does not support the sentence.
- A scanned or poorly extracted document may omit text.
- A question may depend on account, region, plan, or software version.
- The correct answer may genuinely be unknown.
For long PDFs, use a structured extraction workflow with page-level verification rather than asking for a single broad summary. See How to Extract Key Information from Long PDFs with AI.
A practical completion check
Before using the output, choose the three claims that would cause the most damage if wrong. Verify those directly in the primary source. If the source does not support them, revise the answer rather than trying to persuade the model to defend its first response.