# AI Output Fact-Check & Evidence Audit

Audit an AI-generated answer for unsupported claims, weak citations, invented details, outdated facts, and misleading certainty before reuse.

## Prompt

Act as an evidence-focused reviewer of AI-generated content. Audit the draft below before it is published, submitted, or used for a decision.

AI-generated draft:
{{draft_text}}

Available source material or citations:
{{source_material}}

Intended use and stakes:
{{intended_use}}

Perform the review in this order:

1. Claim inventory: extract the important factual claims, numbers, dates, quotations, legal or technical assertions, named entities, causal explanations, and recommendations.
2. Evidence status: label each claim Supported, Partially supported, Unsupported, Contradicted, Outdated or time-sensitive, or Not verifiable from the provided material.
3. Citation quality: check whether each citation actually supports the nearby claim, whether the source is primary or secondary, whether the publication date matters, and whether quotation wording appears exact. Do not invent sources or URLs.
4. Hallucination indicators: flag fabricated specificity, impossible precision, mixed-up entities, unsupported superlatives, fake consensus, missing caveats, or confident statements that exceed the evidence.
5. Risk by use case: explain which issues are harmless wording problems and which could materially mislead a reader or decision-maker.
6. Corrected version: rewrite only the sections that can be safely corrected from the supplied evidence. Preserve uncertainty and insert [SOURCE NEEDED] where verification is still required.
7. Verification checklist: list the smallest set of checks needed before release, prioritizing primary and current sources.

Rules: Do not treat fluent writing as evidence. Do not assume a cited source says something unless its relevant text is provided. Do not fabricate replacement facts. For medical, legal, financial, safety-critical, or current-event claims, explicitly recommend qualified or current-source verification.

## Best for

Writers, students, researchers, marketers, and teams reviewing AI-generated reports, articles, summaries, proposals, or internal recommendations before relying on them.

## Compatible tools

- Claude
- ChatGPT

## How to use

- Paste the complete AI-generated draft and the source excerpts, links, or documents actually available to you.
- Describe where the output will be used and how harmful an error could be.
- Verify every item marked unsupported, contradicted, time-sensitive, or source needed before publication.
- Run the corrected draft through the audit again after adding verified evidence.

## Customization tips

- Ask for a compact claim-evidence table when reviewing a long report.
- Specify a cutoff date when the topic changes quickly.
- For academic work, add the required citation style and distinguish primary research from commentary.

## Example input

Draft: 'Solar panels lose 50% of their output after ten years, according to the International Energy Agency.' Source material: no IEA document is attached. Intended use: a customer-facing renewable-energy article.

## Example output

Claim status: Unsupported and likely misleading. The draft gives a precise degradation figure and attributes it to the IEA without providing a source. The wording also implies a universal ten-year loss rate. Replace it with [SOURCE NEEDED] until a current primary source is checked. Verification priority: locate the exact IEA publication, confirm the quoted metric and population, and compare it with manufacturer degradation warranties or peer-reviewed field studies.