Free tools Windows power users keep installed
One-click scans. No signup required.
Treat any unsupported or contradicted number in AI-drafted copy as a release-blocking error. If a figure can’t be traced to a source that supports the exact claim and its context, delete it or rewrite the sentence so it claims only what you can verify. This is an editorial recommendation, not a standard issued by NIST. It is built on principles that NIST and others have published about accuracy and verifiability.
Why numbers deserve a hard fail
A wrong adjective can be softened in an edit. A wrong number reads as a fact, gets quoted, and is hard to retract. NIST’s Generative AI Profile (2024) calls the broader failure mode “confabulation”: generative AI systems “generate and confidently present erroneous or false content in response to prompts.” NIST also notes that such material can be persuasive when it is delivered confidently or comes with apparently logical reasoning or citations.
As an Amazon Associate I earn from qualifying purchases.
So a plausible figure, a tidy explanation, or a citation-shaped reference tells you nothing about whether the number is true. Only the source can.
What “hard fail” means in practice
A hard fail is a gate, not a warning. A draft with an unresolved numeric claim does not ship. Each flagged figure ends in one of three outcomes:
#1 Best Overall
- Verified: a source supports the exact value and its scope, and you have recorded where.
- Qualified: the copy is reworded to match what the source does support (a different period, a narrower population, “about”, or attribution to the source).
- Removed: the number is deleted, or publication is held until better evidence arrives.
There is no fourth option of “probably right, ship it.”
Checking the source, not the citation
Verification means opening the source and finding the figure. A reference that is merely present doesn’t count. NIST’s paper On the Evaluation of Machine-Generated Reports (2024) stresses completeness, accuracy and verifiability, and states that “evaluation of citations that map claims made in the report to their source documents ensures verifiability.” Apply the same idea to your own copy: each claim should map to a place in a source document.
Common ways a citation fails to support a number:
- The link exists but the figure doesn’t appear on the page.
- The page repeats the figure from somewhere else, so it is a secondary repetition rather than the origin.
- The figure is real but belongs to a different year, country, population or definition.
- The source gave a range, estimate or caveat that the draft turned into a flat statement.
What to preserve when you check a number
This checklist is practical editorial advice; the cited NIST pages don’t provide a standalone numeric-claim checklist. For each figure, confirm:
- Value and unit: percent versus percentage points, millions versus billions, per capita versus total.
- Denominator or population: “38% of what?”
- Geography: a US figure is not a global one.
- Time period: the year, the date of measurement, whether it is a snapshot or a trend.
- Definition: how the source defines the thing counted.
- Qualifications: margins of error, limitations, “estimated”, “among respondents”.
A review workflow you can run and audit
This is an editorial recommendation derived from the accuracy and verifiability principles above, not a procedure NIST prescribes.
Rank #3
- Mark every figure, percentage, date, quantity and comparison in the draft, including words like “double”, “most” and “a third”.
- Open the cited source and locate the original figure or underlying dataset. If the reference only repeats the claim, follow it to the origin.
- Check the exact value, unit, denominator, population, geography, period and definition against the sentence.
- Check whether the source’s limitations or uncertainty were dropped.
- Record the source and a one-line note showing how it supports the sentence, so another reviewer can reproduce the check.
- If support is missing, contradictory or out of scope, fail the claim: delete it, qualify it, or hold publication.
When the AI gave a citation that doesn’t support the number
Don’t ask the model to confirm itself. Treat the citation as a lead, not evidence.
- Source exists, figure differs: use the source’s number and rewrite the sentence around it.
- Source exists, figure absent: remove the number. Search for a primary source; if you find one, restart the check from step 2.
- Source can’t be found or opened: treat it as unverifiable and remove the claim.
- Source supports a narrower claim: narrow the sentence to match.
Comparing review approaches
If you’re choosing between review methods (a single editor, a second-reader check, an AI-assisted source finder), compare them on these axes. They are editorial criteria inferred from NIST’s emphasis on accuracy, completeness, verifiability and uncertainty-aware evaluation (reports, agentic AI probes, statistical evaluation).
| Axis | Question to ask |
|---|---|
| Source type | Is the source primary, or a secondary repetition? |
| Exactness | Are the exact figure and its denominator supported? |
| Context match | Do date, geography, population and definition match the draft? |
| Reproducibility | Can another reviewer repeat the check from the notes? |
| Uncertainty handling | Does the workflow record unresolved claims rather than silently passing them? |
Don’t use benchmark scores as an error rate
It is tempting to cite a model’s hallucination benchmark as the chance that a number in your draft is wrong. Don’t. Those results come from particular models, prompts, datasets and scoring methods. Examples:
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minute- OpenAI’s InstructGPT paper (2022) reports API-dataset hallucination scores of 0.414 for GPT, 0.078 for supervised fine-tuning and 0.172 for InstructGPT.
- OpenAI’s o1 System Card (2024), Table 3, reports SimpleQA accuracy of 0.38 for GPT-4o and 0.47 for o1, with hallucination rates of 0.61 and 0.44. For PersonQA, accuracy is 0.50 and 0.55, and hallucination rates are 0.30 and 0.20.
These are scores on those evaluations. They are not the share of AI-drafted numerical statements that are invented. NIST has also warned that benchmark analyses can rest on implicit assumptions, conflate performance concepts, or fail to quantify uncertainty (NIST, February 19, 2026). The sources we reviewed do not establish a general published rate for invented numbers across tools, topics and editorial settings, so any such figure you see quoted should itself be checked against its origin.
Best Value
The practical consequence: a better benchmark score on a newer model is not a reason to relax the gate. Every consequential number still gets checked.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
