Core & Trust

AI Use Policy

Where AI is used

Drafting and structuring. Assembling an outline, producing a first draft from researched material, and tightening prose.

Checking. Cross-referencing figures across pages, finding internal inconsistencies, spotting where a limitation has gone missing, and catching claims that drifted from their sources during editing.

Building. The site itself is generated from a data layer, and that code was written with AI assistance. That is engineering, not editorial.

Where it is not used, and cannot be

Verifying a price. A researcher opens the provider's page, reads the figure, and archives a screenshot against an evidence ID. A model cannot do this, and a model asserting a price is exactly the failure this publication exists to correct.

Confirming a licence. A human checks the issuing board's own record.

Reading a trial. A human reads the publication. Summarising a summary is how estimand confusion and dose conflation propagate.

Medical review. A named clinician, on pages they have read.

Why the distinction is the whole point

Language models reproduce what is repeated. In this market what is repeated is the advertised starter price, and trial figures attached to products those trials never tested. A publication that used AI to generate its facts would industrialise precisely the errors it should be catching.

So the rule is simple: AI may help express what a human established. It may not establish it.

What this means for you as a reader

Every figure carries an evidence status and a capture date, and every claim resolves to a ledger entry naming the human who captured it. If a ledger entry has no reviewer name, the claim is not verified — and the page will say so rather than presenting it as fact.

That is the check on this policy. It is not a promise about our process; it is a record you can inspect.

Questions readers actually ask