AI makes up facts and citations in our documents
Last verified: Claude Fable 5 / GPT-5 era models ยท July 2026
The draft looks authoritative: precise percentages, named studies, confident figures. Then you check one and it doesn't exist. In an internal note that is annoying; in a client proposal or board report it is a credibility risk, and verifying every claim by hand removes the whole time saving. This page gives you a prompt that makes the AI use only real, provided data and flag anything it cannot stand behind, plus the habit that keeps it honest.
Part 1 · The generator prompt
Paste this into a fresh Claude or ChatGPT chat. It interviews you about what counts as a real source for your work, then hands back a grounding block you paste at the top of any factual-writing chat so the AI uses only real data and flags what it cannot verify.
You're going to build my company's GROUNDING BLOCK: a reusable instruction block I'll paste into any AI chat that writes factual documents for us, so it stops inventing numbers, studies and sources. Interview me first, one question at a time. Where I don't know, propose a sensible default and move on: 1. What documents does this apply to, and how damaging is a wrong fact in them (internal note versus client proposal versus board report)? 2. What counts as an allowed source of truth for us: data and documents I provide in the chat, our own published material, named reputable sources? What is off-limits? 3. When the AI doesn't have a figure, what do we want: a clearly marked placeholder, a flagged assumption, or a question back to me? Never a confident guess. 4. Do we ever cite external sources in these documents, and if so what counts as an acceptable citation (a real, checkable reference, not a plausible-looking one)? 5. What does the AI most often invent for us today: statistics, market sizes, quotes, study names, dates? Then output my GROUNDING BLOCK as one copy-paste block containing: (a) one line on the documents this covers, (b) the allowed sources of truth and what is off-limits, (c) the hard rule: use only provided or verifiable facts, never invent a figure, statistic, quote or citation, and never present an estimate as a fact, (d) the missing-data rule: mark it clearly or ask, do not fill the gap, (e) a closing self-audit that, before output, lists every factual claim in the draft with its source or marks it UNVERIFIED, then appends "Facts check: pass" only if nothing is unverified. Keep it under 400 words so it fits any project-instructions field. After you output it, name the three kinds of fact the AI is most likely to still invent for us and how the block catches each.
Part 2 · The habit that keeps it honest
The block sets the rule. This is how you use it so nothing slips through.
- Store the grounding block in your AI tool's project or custom-instructions field so it applies to every factual document, not just the ones you remember to paste it into.
- Give the AI the real source material in the chat: your data, the approved figures, the actual documents. A grounding rule with nothing to ground on still leaves the AI guessing.
- Require the claims audit before you accept a draft: the block makes the AI list every factual claim with its source or mark it UNVERIFIED. Read that list first; it is faster than re-checking the prose.
- Treat any UNVERIFIED or flagged line as a task for you, not the AI: supply the real figure or cut the claim. Never let the AI firm up an estimate into a fact.
Part 3 · A worked example
Here is the block the generator hands back, for our running example company, the fictional consultancy Fenwick & Cole.
CONTEXT: applies to all client-facing documents (proposals, reports, board memos) where a wrong fact is a credibility risk. ALLOWED SOURCES: data and documents I provide in this chat, and Fenwick & Cole's own published figures. Off-limits: invented statistics, market sizes, quotes, study names, or any figure you cannot point to a provided source for. HARD RULE: use only provided or verifiable facts. Never invent a number, statistic, quote or citation. Never present an estimate, extrapolation or "typical" figure as a fact. If a claim would need a source we don't have, don't make it. MISSING DATA: when you don't have a figure, insert [FIGURE NEEDED: what it is] or ask me. Do not fill the gap with a plausible number. SELF-AUDIT: before output, list every factual claim in the draft, each with its source, or mark it UNVERIFIED. Append "Facts check: pass" only if nothing is marked UNVERIFIED; otherwise list what needs a source.
Common questions
Why does AI make up statistics and sources?
A language model generates the most plausible-sounding next words, and a specific percentage or a named study reads as more plausible than "I don't know". It has no built-in sense of what is true, only what is likely, so it will confidently produce a figure or citation that fits the sentence. Constraining it to provided sources and forcing it to flag anything unverified is what closes that gap.
How do I stop AI hallucinating facts in documents?
Give it a grounding block: name your allowed sources of truth, forbid inventing figures or citations, tell it to mark missing data rather than fill it, and require a claims audit that lists every fact with its source before output. The generator prompt on this page builds that block. Pair it with the real source material in the chat, because a grounding rule only works if the AI has something real to ground on.
How do I get AI to cite only real sources?
Tell it explicitly that a citation must be a real, checkable reference and that a plausible-looking one is worse than none, then require it to list each citation for you to verify. Better still, provide the sources yourself and instruct it to cite only from those. Stored as always-on context, that rule applies to every document instead of being a check you run by hand.
The prompt is the aspirin. Stored context is the cure.
The grounding block you just generated, plus your real data, is exactly what an AI Brain stores as always-on context. Paste it and this document stays honest; installed, every document draws on your actual figures and flags anything unverified by default, because the rule and the source of truth live in the system. This page proved it on your own work. The install is how checking every number by hand stops being your job.
What else does AI get wrong that you fix by hand?Browse the other fixes →