Fact-Checking AI Outputs: A Practical Workflow

A concise, repeatable method to extract claims, verify them with primary sources, add accurate citations, and publish AI-assisted content you can trust.

Summary

This guide turns AI drafts into verifiable, publishable copy. You’ll extract claims, check them against primary sources, document citations, and mark anything unverified. The workflow fits newsrooms, tech docs, and everyday content operations.

Who is this for

Editors, reporters, technical writers, analysts
Skill level: Beginner → advanced
Prerequisites: Web search basics, note-taking discipline, access to reputable sources

Key concepts (fast)

Primary vs. secondary sources: Primary = original documents/data (laws, filings, press releases, docs, datasets). Secondary = reporting/analysis about those primaries. Prefer primary for facts; use secondary for context.
Verification loops: Re-check high-impact claims with at least two independent sources or one primary source. Loop until contradictions are resolved or the claim is labeled unverified.
Quote limits: Keep direct quotes short and necessary. Attribute clearly; avoid over-quoting. Summarize where possible and follow your outlet’s fair-use policy.
Date precision: Use absolute dates with time zones; confirm publish date and event date. Normalize formats consistently across the piece.

Tools & setup

Link checking: Use a dead-link checker or manual test (open in clean browser profile). Revisit all links before publish.
Archive links: Create Wayback/Perma.cc archives for key citations; keep both live and archived URLs in notes.
Policy on block quotes: Define maximum length (e.g., ≤75–100 words), mandatory attribution, and when to paraphrase.
Source log template: Columns: Claim ID · Claim text · Source type (Primary/Secondary) · URL · Archived URL · Accessed date/time · Status (Verified/Unverified/Contradiction) · Notes.
Time zone standard: Choose one house style (e.g., ET) and note original time zone in parentheses when material.
Version control: Store drafts and fact tables in a shared drive; lock final after approval.

Step-by-step

Plan
Define scope, audience, and “facts that matter” (metrics, dates, prices, names, SKUs, policy IDs). List your non-negotiables (e.g., must have at least one primary source per core claim).
Generate
Use AI to draft. Instruct it to separate facts from opinions and to footnote claims (inline markers like [C1], [C2]).
Extract claims
Pull every factual assertion into a table (Claim ID, text, importance). Combine duplicates; flag high-risk items (legal, medical, financial).
Verify with primary sources
For each claim, search for official docs, filings, press releases, technical docs, datasets. Use secondary sources only to triangulate or add context. Record live + archived links and access time.
Add citations
Insert citations near the claim (house style). Prefer primary links; include archive link in editor notes or footnotes. Keep quotes minimal; paraphrase accurately.
Final review
Re-run a link check, confirm dates/time zones, ensure names/titles match sources, and label any unresolved items as Unverified with a brief reason.

Mini checklist: All key claims verified ▸ Dates normalized ▸ Quotes within limits ▸ Citations placed ▸ Unverified items labeled ▸ Links live + archived.

Prompt patterns (copy-ready)

1) Claim extraction

You are my claim extractor. From the draft below, list factual claims as a table:
- Columns: [C-ID] [Claim] [Category] [Evidence needed] [Priority: High/Med/Low]
- Exclude opinions and adjectives unless they contain measurable facts.
- Do not verify yet; just extract.
DRAFT:
{{paste text}}

2) Source triage (primary-first)

Act as a research lead. For each claim in the table:
1) Suggest the most likely PRIMARY source (doc/filing/press release/dataset/manual).
2) Provide a search plan (3 queries per claim) and expected validation signals.
3) Flag claims that cannot be primary-verified and propose credible secondary fallbacks.
INPUT (claims table):
{{paste table}}

3) Quote-with-citation (fair use)

Given the verified sources, produce paraphrased statements with short, necessary quotes only.
Rules:
- Use ≤1 sentence of direct quote per claim unless essential.
- Add in-text citation placeholders [S#] next to each claim.
- Output a source map at the end: [S#] → Title, Publisher, Date, URL (live), URL (archive).
INPUT:
{{claim → source pairs}}

4) “Unverified” labeling

Scan the claims and mark any that remain unverified or contradictory.
For each:
- Add a concise "Unverified:" note with why (no primary source, conflicting numbers, outdated).
- Suggest next-best action (contact PR, check SEC/EDGAR, request dataset).
Return an updated claims table with a Status column.

5) Date normalization

Normalize all dates/times to the house style: Month DD, YYYY, h:mm AM/PM TZ.
- Preserve original time zone in parentheses if helpful.
- Distinguish publish date vs. event date.
Return a corrected list and highlight any mismatches with sources.
INPUT:
{{list of dates/times with contexts}}

6) Link checker & archive fallback

Validate all citation URLs:
- Test reachability, HTTPS, and canonical form.
- If fragile or paywalled, create/archive an alternative and record both.
Output a table: [S#] [Live URL OK?] [Archived URL] [Notes].
INPUT:
{{source map or URL list}}

Pro tips & tricks

Verify names, titles, and org spellings last—typos slip through late.
For stats, record both the number and its context (unit, population, time window).
Screenshot key tables/figures for internal notes; never as a substitute for a link.
When two reputable sources conflict, prefer primary; if still unclear, state the range and why.
Keep a “high-risk claims” sidebar for legal/medical/finance; escalate early.
Use consistent anchor text for citations (clear, non-clickbaity).

Examples

A) News article (policy update)

Before (AI draft): “The agency will roll out the rules next month, affecting millions.”
After (fact-checked excerpt): “The agency finalized the rule on March 15, 2025, 10:00 AM ET (Federal Register, Docket ####). The effective date is April 30, 2025. The preamble estimates ~3.2 million affected accounts, derived from the FY2024 dataset (Table 2). [S1][S2]”
Notes: Dates separated (final vs effective). Numbers attributed to a primary table. Generic “millions” replaced with the cited estimate.

B) Technical how-to (software flag)

Before (AI draft): “Enable fast mode by toggling the speed setting.”
After (fact-checked excerpt): “Enable Fast Mode via Settings → Performance → Fast Mode (v2.1+). In v2.0 and earlier, the flag is --fast only at runtime. See User Guide v2.1, §3.2 and Release Notes v2.1.0. [S1][S2]”
Notes: Version-specific behavior clarified; exact path + flag cited; sources are official docs.

Common pitfalls & fixes

Vague quantities (“many,” “most”) → Replace with sourced numbers or ranges.
Mixing publish and event dates → Track both; state them explicitly.
Over-quoting → Paraphrase; keep the quote fragment that carries unique language.
Dead links → Add archive links and a brief note in the source map.
Secondary-only verification → Keep publishing blocked until a primary is found or label Unverified with rationale.