A reader judging a healthcare article needs to know how the text in front of them was made. A
publication that used machine assistance and mentioned it nowhere is asking for a trust it has
never described. So every article must declare aiAssistance in its frontmatter, and a declared
value is rendered on the article page.
The disclosure is on the page
aiAssistance sits in the article's own metadata block, beside the author, the reviewer and
validityAsOf — not in a site footer, and not in a sentence at the end of a policy page a
reader would have to go looking for. A disclosure placed where nobody reads it has been placed
there for that reason.
The field records what the assistance did and to which part of the work: a first draft, a
restructure, a summary of a source a person supplied, a translation of a revision. It cannot be
left out — an article that omits it does not build — and null is how a person records that no
assistance was used, so that an empty field cannot be told apart from a field nobody filled in.
What null does not do is put a line on the page: a reader sees the disclosure where there is
one, and no row at all where there is not.
This publication holds no marketplace data — no bookings, no professional records, nothing about any patient — so there is no such data in this repository to put into a tool in the first place. The separation is architectural, and that is the whole of what it controls: it says nothing about what somebody with marketplace access does elsewhere.
A machine may draft. It may not be accountable
Drafting, restructuring, summarising and translating are tasks whose output can be held against something a person already has: a source they read, an outline they wrote, a revision they published. That is the whole of what machine assistance is used for here.
Accountability does not distribute across that arrangement. authorIds and reviewerIds name
people, and those people are answerable for every published sentence, including sentences they
did not type. A machine holds no qualification, appears in no register, cannot be a reviewer,
and cannot be asked six months later why it wrote a particular clause. Assistance is not
authorship.
Fabricated citations are the specific hazard
Every other machine failure shows up in the prose. This one does not. A citation with the right shape — a well-formed DOI, a plausible journal, a plausible year, authors who genuinely publish in the field — reads exactly like a real one. An editor reviewing the sentence sees nothing wrong, because nothing in the sentence is wrong. The failure is that the referent does not exist, and no amount of reading the article can reveal that.
The control is therefore not review of the text. Every source in sourceIds must be opened by a
person, and that person writes retrievalRoute: prose describing how the source was actually
read — the register searched, the term used, the PDF downloaded, the page the statement appears
on. dateRetrieved records only a date. A successful HTTP status is not proof that a source
was read, and a DOI that resolves is not proof that it says what the article claims.
The hazard is worst at the top of the source hierarchy, because an article at HIGH or
CRITICAL risk may rest only on sources at or above PEER_REVIEWED_STUDY, and
SYSTEMATIC_REVIEW and PEER_REVIEWED_STUDY are precisely the types easiest to invent
convincingly. LEGISLATION, REGULATOR_PUBLICATION and OFFICIAL_REGISTER at least have a
public canonical location a person can be required to reach.
A correction caused by a citation that turned out not to exist is material: true by
definition. A person sets that flag; once it is set, the correction file is refused without a
reviewer and refused with dateModifiedChanged: false, and it records the previousClaim
verbatim. The publication does not get to quietly delete a reference and move on.
MACHINE_DRAFT cannot publish
translationStatus carries MACHINE_DRAFT, and it is deliberately absent from the translation
statuses a published article is allowed to be in. The value exists so that machine-drafted
translations can be held in the repository and worked on. It does not exist as a shortcut to a
locale.
The reason is that a translated clinical or legal sentence is a new claim in a new language,
read by people in a different jurisdiction, and the reviewer who approved the original read
something else. translatedFromRevision pins which revision was translated so that a later
correction to the source article can be traced into every locale that inherited it. A
translation whose source revision has since been corrected is a stale article wearing a current
date.
What a claim has to carry before it publishes
No build can see whether a person read a source. What it can do is refuse a claim that rests on
nothing. Each entry in claims carries a basis, and a jurisdiction-specific claim resting on
REASONED_INFERENCE, COUNSEL_CONFIRMATION_REQUIRED or UNKNOWN cannot be published — the
check refuses the file rather than warning about it. Machine assistance creates no new basis
and upgrades none. A statement a machine produced starts at UNKNOWN and stays there until a
person attaches it to a source they opened themselves.
Only PUBLISHED and UPDATED reach a page, a sitemap, a feed or the search index. Everything
before them — EDITORIAL_REVIEW, CLINICAL_REVIEW, LEGAL_REVIEW, READY — is a person
doing work that no tool performs on their behalf.
What this policy does not enforce
aiAssistance is self-declared. Nothing in the build detects an undeclared machine draft, and
no credible check for one exists. retrievalRoute is prose, and prose can be written by
somebody who did not open the source; the field raises the cost of pretending and records who
pretended, which is not the same as preventing it.
The controls that actually hold are the ones that refuse a file: the claim-basis check, the
source tier required at HIGH and CRITICAL risk, the freshness failure on HIGH and
CRITICAL articles and on any article carrying emergencyInformation, and the absence of
MACHINE_DRAFT from the publishable translation statuses. Everything else in this document is
a record of what a named person says they did, kept in a repository under their name, where a
later correction can contradict it. This is honest, not clean.