AI Fact-Checking & Hallucination Avoidance Checklist

Updated: 2026-08-12

Why this guide

Models write fluent sentences—including wrong numbers, outdated policies, and papers that do not exist. Fact-checking is basic hygiene for publishing and decisions. Use Claim → Evidence → Verify (CEV), fake-citation tactics, and a graded pre-publish checklist. Tools: ChatGPT, Claude, Gemini; docs: OpenAI Help. Pair with AI writing workflow.

Outcomes

• Turn “feels off” into steps
• Catch fake citations and mismatched links
• Match verification depth to risk
• One shared publish gate for common accidents

Why fluent ≠ true

Models optimize continuation and instruction-following—not your system of record. They may splice true fragments into a false conclusion, invent plausible DOIs/URLs, treat 2023 policy as “current,” or fabricate sources when you demand citations. Treat outputs as draft hypotheses.

Method: Claim → Evidence → Verify

Column	Meaning	Example
Claim	Checkable sentence	“Product X raised prices 20% in Jan 2026”
Evidence	Primary source type	Vendor pricing page, gazette, paper PDF
Verify	Your open-and-record result	Confirm / deny / partial / not found

Extract all checkable claims from the text.
Table: claim | type (number/date/attribution/causal/citation) | risk (H/M/L) | source types to check
Do not prove them true; do not invent sources.
Text: { }

For claim “{one sentence}”:
1) If unsure, say so
2) Name source types I should open (official site / statute DB / paper DB)—no forged URLs
3) Describe what a counterexample would look like
Ban invented links and paper titles.

Minimum human moves: open trusted domains (not mystery short links); check entity/date/scope; fix or delete; if no evidence → remove or demote to opinion.

Fake citations & links

Red flags: overly perfect titles with odd author spellings; DOI/ISBN that resolve nowhere; contradictory venue/year/pages; “top journal” with no stable link; statute numbers that do not match known structure.

Steps: search title/author in Scholar/publisher/official sites; resolve DOIs officially; check statutes in primary databases; cross-check news with ≥2 independent outlets.

Review this reference/link list. Flag: format anomalies; items needing human search (say so—do not invent results); claims propped only by weak blogs.
Output: item | why suspicious | where I should check
List: { }

Risk tiers

Tier	Examples	Minimum check
R0	Private jokes, fiction	None
R1	Opinions, subjective reviews	Spot-check numbers; no fake cites
R2	How-tos, product features	Walk official docs
R3	Pricing, SLA, medical/legal/finance advice	Full claim table + expert/compliance review
R4	Press, regulatory, contract-adjacent	Dual review + saved evidence links

Any actionable advice others might follow is at least R2.

Pre-publish checklist

• Claim table run; high-risk rows verified
• Every link opened; content matches
• No “studies show” without a study
• Time words (“current/latest”) bound to dates/sources
• Benchmarks state conditions + date
• Knowledge-cutoff claims refreshed with 2026 primary sources
• Quotes/screenshots not misleading
• Sensitivity pass (privacy)

You are a pre-publish checker. List: claims/links needing human opens (urgency 1–5); hype sentences; absolute claims missing caveats.
Do not invent live-check results.
Body: { }

Prompt patterns that reduce hallucination

Pattern	Example
Grounding	“Use only the text below; no outside facts”
Uncertainty tags	“Unknown → [VERIFY]”
Reasoning over authority	“List assumptions and chain”
Split creative vs factual	Outline first; fact blocks verified separately
No citation coercion	Don’t demand “three papers” unless you have a corpus

Answer using ONLY the material. If missing, say “not in material.”
Output: conclusion | short supporting quote | unanswered questions
Material: { }

Also useful in study: AI study workflow.

Team rollout

Define default R tiers; share a claim table; author fills Claim, reviewer Verify, publisher ticks checklist; monthly sample three live posts for link rot; incidents get a correction plus a note on which CEV step was skipped.

Common incidents

Incident	Fix
Fake paper shipped	Delete cite; sweep author names; publish correction
Stale pricing	“Per site today” + update date
Wrong feature attribution	Check official release notes/version
Causation overclaim	Soften to correlation/possibility + conditions
Skewed translation of EN sources	Re-translate key sentences from primary text

FAQ

Does browsing/plugins remove the need to check?

No. Summaries can be wrong or poorly sourced. Open primary pages for critical claims.

Fast fake-citation heuristic?

The flashier the venue/title/year, the more you must search. Real papers survive retrieval; fakes often don’t.

Is “I’m not sure” good?

Yes—safer than confident error. Tag [VERIFY] and assign an owner.

Can RAG still hallucinate?

Yes—wrong chunks or stitching errors. R3+ still needs CEV against retrieved passages.

Check English primaries for translated claims?

If the claim originates in English official/paper sources, verify there—don’t trust model paraphrase alone.

Related reading

• AI writing workflow
• AI privacy
• Prompt tips
• Getting started

Next reading

• Choose an AI assistant
• FAQ
• China access

Action path

Today: run claim extraction on your latest AI-assisted draft; verify ≥5 high-risk rows. This week: pin the publish checklist; agree R tiers with the team. Ongoing: numbers only from sources you opened—model drafts, you own truth. Help: OpenAI Help; non-sensitive practice: domestic entry.