AI wrote it. Did it also invent the numbers?
Ask a language model for a persuasive paragraph and you will usually get a statistic with it. Not because the model checked anything, but because the paragraphs it learned from contain statistics. The number arrives fully formed, plausible, precisely wrong. A fabricated figure reads exactly like a real one, which is why editing does not catch it and why the fix has to be mechanical.
Why models invent figures
A language model completes patterns. In its training data, confident claims travel with numbers: adoption rates, growth multiples, survey ratios. So when it drafts your landing page or your comparison post, it reproduces the pattern, and the pattern includes a figure. The gate's README opens with the canonical case: a draft asserting "studies show 73% of teams" when no such study exists.
The failure is not occasional carelessness you can prompt away. It is the default behavior of a system optimized to produce likely text. A prompt that says "only use real statistics" moves the odds; it does not create a guarantee. A prompt rule without a gate is a suggestion.
Why review does not catch it
A fabricated statistic carries no tell. It has the same shape, the same precision, and the same confident framing as a sourced one, so a human skimming the draft has nothing to snag on. Checking it properly means hunting for the underlying study, for every figure, on every page, forever. Nobody sustains that at publishing cadence.
Asking a second AI to review the first does not close the loop either: the reviewer shares the writer's failure mode. It, too, judges whether the sentence reads plausibly, and the whole problem is that fabricated figures read plausibly.
The mechanical fix: traceability, not vigilance
Flip the burden of proof. Instead of asking "is any number in here fake?", require that every number in the copy traces to an entry someone deliberately recorded: the claim, the exact value, the source URL. That record is the source-pack manifest, and the check against it is mechanical: extract every data-claim from the published files, and fail the build on any figure with no matching manifest value.
This is what no-fabricate gate does. It recognizes the five shapes a statistic takes in prose (percentages, money, multipliers, "N out of M" ratios, and scale words), exempts structural numbers like years and ordinals automatically, and exits non-zero the moment an unsourced figure appears. The gate does not judge whether your source is good. It guarantees something narrower and enforceable: no number ships without a citation next to it.
What this page practices
This page is about fabricated statistics, so it is a fair place to check the discipline. It quotes exactly one figure, the README's own worked example, and that figure is recorded in this site's source-pack manifest, cited to the README it comes from. Everything else here is stated without numbers, because we do not currently hold sources for stronger quantified claims, and the honest response to a missing source is a sentence without a statistic.
Keep reading, then get the source.
- no-fabricate gate: what it is
- Source packs: the manifest that gates your build
- Why it is safe to install
- Sister gate: de-AI gate, for machine-written style
- View the source on GitHub
no-fabricate gate is open source under the MIT license. Put a citation requirement between your AI drafts and your readers.
Get no-fabricate gate on GitHubFAQ
Can better prompting stop AI from inventing statistics?
It reduces the rate; it cannot guarantee the outcome. A language model produces likely text, and prompts shift likelihoods rather than impose constraints. The reliable arrangement is a prompt that says "use only the numbers in this source pack" backed by a gate that fails the build when the draft ignores the instruction.
Will a human editor catch a fabricated number?
Sometimes, but not dependably. A fabricated figure has no visible tell: it is shaped like a real one. Verifying it means locating the underlying source for every figure on every page, which does not survive contact with a real publishing cadence. Traceability turns that open-ended hunt into a mechanical lookup.
Does the gate check whether my sources are accurate?
No. It checks that every statistic in the copy traces to a manifest entry with a claim, a value, and a source URL. Choosing trustworthy sources remains a human judgment, made once when the entry is recorded, instead of a scramble after something ships.