Editorial standards
AI Truth Bench exists to tell readers whether an AI tool's answers can be trusted. That only works if our own figures can be trusted. These are the rules every page follows.
Sourcing
Every number on the site comes from a published source that a reader can open. We use five kinds of source:
- Studies: peer-reviewed papers, preprints and working papers, such as the academic tests of AI detectors.
- Public benchmarks and leaderboards: shared tests where many tools are scored the same way.
- News investigations: tests run by newsrooms and research centres on AI answers.
- Court records: orders and filings, used for real cases of AI errors.
- Vendor figures: what a company says about its own product.
We collect published results; we do not run our own tests. Each figure is kept in a single results register with the tool, what was measured, the value, the date and the source link, so a number reads the same on every page that uses it. We link to the original source wherever it is publicly available, not to a summary of it.
Dating
Every result carries a date: the publication date of the study, the release date of a benchmark entry, or the date we checked a live page or leaderboard. Live leaderboards and vendor pages change, so pages show the date each figure was checked. Older studies stay on a page when they are still the best evidence, and the text says which version of a tool they tested.
Vendor claims and independent results
A vendor's own figure is always labelled as a vendor claim, in tables and in the text. Where a public benchmark accepts results that vendors submit themselves, we say so. We never rank tools by mixing figures from different tests, because each test measures something different: a detection rate in one study cannot be set against a false positive rate in another.
Verdicts
Each tool page gives one of four verdicts: Reliable, Mostly reliable, Mixed or Unreliable. The verdict follows from the evidence cited on that page, and the sentence beside it names the figures it rests on. The method is set out on how we score AI accuracy.
How we write
Pages lead with the answer, then show the evidence. We use plain words, short sentences and numbers with dates, and we avoid hype in either direction. When a test measured the model behind a product rather than the consumer app people use, the text says so. Where studies disagree, the page shows both results and explains what each one measured.
Corrections
If a figure is wrong, out of date or misattributed, email editors@aitruthbench.com with the page address and a link to the source. We check the report against the source. When a correction is needed, we fix the page, change its updated date and add a short dated note at the foot of the page saying what changed.