How to Stop ChatGPT, Claude and Gemini From Making Up Numbers in Your Reports
Paste a locked numbers pack, ban unsourced figures, do the math outside the model, then run a number audit. Copyable prompts for reports, QBRs and decks.
You cannot make a language model incapable of inventing numbers. You can make invention much less likely, and easy to catch. Give it the only numbers it may use, tell it to write NEEDS FIGURE when a number is missing, do the arithmetic yourself, and run a second pass that traces every digit back to your data. That is the locked numbers workflow behind the Locked Numbers Pack Report Writer: Paste Only Approved Metrics, Ban Unsourced Figures, Precomputed Math, and a Number Audit Pass prompt and the wider PromptDig "no invented metrics" pattern.
Why AI invents numbers
Models generate plausible text. When a sentence wants a percentage and none was supplied, the model often fills a number shaped slot with something that sounds right. Search tools can help find real figures, but they can still misquote or mis-attribute them. The answer is not a pep talk. It is a workflow.
The locked numbers workflow (4 steps)
Step 1: Build a locked numbers pack
Paste a small closed table before any drafting instruction. Columns that work well: metric, value, unit, period, source label.
Sample data only (fictional company):
| metric | value | unit | period | source label |
| ARR | 4.2 | USD M | 2026-Q3 close | Finance close pack |
| ARR QoQ | 8 | percent | Q2 to Q3 | Finance close pack |
| Net logo churn | 2.1 | percent | 2026-Q3 | CS ops export |
| NPS | | score | 2026-Q3 | (blank) |
If a cell is blank, leave it blank. Do not ask the model to estimate it.
Step 2: Prohibit, with a placeholder
Add one hard line: Use only numbers present in the pack. If a claim needs a missing number, write NEEDS FIGURE instead of guessing. Never invent benchmarks or competitor stats.
The token gives the model something to do other than improvise.
Step 3: Do the math outside the model
Totals, deltas, and percentages belong in your sheet first. Put the finished math into the pack. Ask the model to narrate, not to calculate. This matches the pain people report when a chat "helpfully" derives a wrong margin.
Step 4: Run a number audit pass
After the draft, run a second prompt: List every numeric token in the draft, the pack row it came from, and match yes or no. Do not rewrite yet. Delete or rewrite unmatched sentences afterward.
The Locked Numbers Pack Report Writer: Paste Only Approved Metrics, Ban Unsourced Figures, Precomputed Math, and a Number Audit Pass prompt runs pack echo, draft, and audit in one structured job when you want a single tool.
Copy-paste prompts for common reports
Use short guards here, then open the full PromptDig prompt for the document type.
Weekly exec flash
Paste the pack, then: Write a one page exec flash for [audience]. Only pack numbers. NEEDS FIGURE for gaps. No new KPIs.
Full tool: Weekly exec flash from a closed metrics pack.
QBR deck outline
Paste the pack, then: Outline QBR slides from these KPIs only. No invented pipeline coverage or market share.
Full tool: QBR deck outline from locked KPI pack.
Investor FAQ
Paste diligence numbers, then: Answer FAQ items with pack figures only. Qualitative where blanks exist.
Full tool: Investor FAQ outline from diligence notes.
Newsletter or Substack stats
Paste the inventory of stats you actually measured, then: Draft a post outline with those stats only.
Full tool: Substack notes to post outline from draft inventory.
Client or campaign recap
Paste contracted metrics and delivery notes, then: Write a recap without inventing revenue or reach.
Related: Partnership one pager from a term sheet sketch and Looker Studio metric definition packet.
What to do when a number is missing
Keep NEEDS FIGURE in the draft. Ask the owner. Or rewrite the claim qualitatively ("customer sentiment data is not in this pack yet"). Never estimate in a client facing doc to sound confident.
Does web search or deep research mode fix it?
Search can surface real pages, and still misread a table or blend two years. You still audit. Do not claim tool specific accuracy rates you have not verified on publish day.
Before and after example (sample data)
Without the workflow, a prompt like "Write an exec flash: growth is strong and customers love us" may invent an NPS of 72 and a market growth rate of 14 percent.
With the pack above, the draft can say ARR 4.2 USD M and 8 percent QoQ, churn 2.1 percent, and NEEDS FIGURE for NPS. The audit table shows three matched numbers and zero unmatched digits. The sentiment sentence becomes a question to CS instead of a fake score.
How teams adopt this in a week
Day 1: pick one recurring doc (the weekly exec flash). Build a pack template in your sheet with the five columns. Day 2: add the prohibition line to your team prompt library. Day 3: require an audit table in the pull request or email draft before send. Day 4: delete any old prompts that say "include relevant benchmarks" without a pack. Day 5: retrospectively check last week's flash for unmatched digits and show the before after in standup without blaming people.
The cultural shift matters as much as the syntax. Praise people who leave NEEDS FIGURE visible. Treat silent invention as a process miss, not a clever draft.
Checklist (printable)
- [ ] Pack pasted with source labels
- [ ] Prohibition line and NEEDS FIGURE token included
- [ ] Math precomputed in the sheet
- [ ] Draft uses only pack digits
- [ ] Audit pass completed
- [ ] Unmatched sentences rewritten or removed
- [ ] Human sign off before send
For a related compliance style pass on job content, see Job prompts compliance pass.
FAQ
How do I stop ChatGPT from making up statistics?
Lock a numbers pack, ban unsourced figures, precompute math, and audit.
Can a prompt completely stop AI from inventing numbers?
No. It reduces and exposes them. Say that honestly to stakeholders.
Why does ChatGPT make up numbers?
It completes plausible text when a slot wants a figure and none was supplied.
Should I let ChatGPT do the math in my report?
Prefer sheet math. Put finished results in the pack.
What should the AI write when a number is missing?
NEEDS FIGURE, or a qualitative rewrite.
Does ChatGPT web search stop fake statistics?
Not by itself. Audit citations and digits.
How do I check AI-written numbers quickly?
Run the audit table: every digit beside a pack row.
Is this the same problem as fake citations?
Related. Numbers and sources both need closed inventories. A sister brief on fake citations may follow on PromptDig; until then, keep both packs closed.
Related prompts on PromptDig
- Locked Numbers Pack Report Writer: Paste Only Approved Metrics, Ban Unsourced Figures, Precomputed Math, and a Number Audit Pass (flagship for this workflow)
- Weekly exec flash from a closed metrics pack
- QBR deck outline from locked KPI pack
- Investor FAQ outline from diligence notes
- Looker Studio metric definition packet
- Partnership one pager from a term sheet sketch
Explore more in Business. Browse more prompts. If you ship a locked pack template your team reused, Share a prompt.