This is the standing disclosure for the AI in Education Knowledge Base: which AI systems are used to produce it, what they do, what human oversight covers, and what a reader can check. It is written for the repository, for the offline editions, and for anyone who needs to restate how the text was produced — a publisher, a reviewer, or a system that asks whether a work contains AI-generated content.
The short version: page text is drafted by large language models from source documents, a human editor reviews, corrects and publishes, the sources behind every page are recorded, and the citations are checked against the publisher record. No AI system is listed as an author or contributor, because no AI system can be accountable for what is published.
The people accountable for this knowledge base are recorded in site.config.json
under contributors, and every page names its own contributors in its frontmatter
contributors field, by id, so the names live in one place. When more than one person
works on the corpus, each page lists the people who produced or reviewed it, and the
record grows by addition rather than replacement.
AI systems are never listed there. This follows the position taken by COPE, ICMJE and
the major publishers: an AI tool cannot be an author, because authorship requires
taking responsibility for the work. Its contribution is recorded separately, in the
ai_assist field described below.
| Role | What it covers |
|---|---|
| drafting | writing a page's prose from the source document |
| revision | rewriting, enriching or restructuring existing page text |
| link classification | proposing inline wiki links between pages (proposals only; a human accepts or reverts each one) |
| summarization | condensing a source into the synthesis and key findings on an article page |
| translation | rendering source or page text across languages, where used |
| none | the page is human-written |
Different models have produced different parts of this corpus. The systems in use are
listed in site.config.json under aiDisclosure.models, and the entry that settles
which one produced the text now on a given page is that page's own ai_assist record
— not the model list, and not the page's created date, since concept pages are
rewritten wholesale as new research arrives.
The editor commissions pages, supplies or approves the source documents, reviews the
drafted text, corrects errors of substance and tone, and approves publication.
Corrections during review are not always visible in the page history, so the
reviewed_by field records which pages a person has actually read. It is left empty
where that has not happened. An empty field is not a claim that the page is unreviewed;
it means the record does not assert a review.
Pages carry a provenance block in their frontmatter. It is not displayed on the site; it exists so that anyone reading the markdown, the repository history, or a harvested copy of the corpus can see how the page was produced.
contributors— ids of the people accountable for the pagereviewed_by— ids of the people who have read and checked the pageai_assist— an ordered record of AI contributions to the current text, each withmodel,roleanddatesource_depth— whether the page was written from the full text of its source, from an abstract only, or from metadata aloneverified— which checks were run:citation(the reference resolves and matches the publisher record),numbers(figures in the page appear in the extracted source text),quotes(quoted wording matches the source)
ai_assist describes the text as it now stands, not every edit ever made to it, and it
records material contributions: prose that was drafted or rewritten. Mechanical passes
over a page — wiring a back-link, refreshing a count, fixing a spelling — carry no
substantive claim and are not recorded there. The record is not backfilled, so a page
with no ai_assist entry is one the record is silent about, not a page without AI
involvement.
The per-change record is in the repository history: each commit carries trailers naming
the model and role behind that change (AI-Model:, AI-Role:), and human
co-authorship uses git's own Co-authored-by: convention. Commit trailers are immutable
and survive page rewrites, which a frontmatter field alone cannot do.
- Citations are checked against the publisher's record — the DOI resolves, or the journal page carries the article. A citation is never fabricated, and a paper with no registered DOI is cited by its journal landing page instead.
- Numbers are checked mechanically: a gate compares every figure in a page against the extracted full text of its source, and flags any that cannot be found.
- Sources are stored as extracted text under
raw/(git-ignored) with the URL, DOI, extraction method and a SHA-256 hash of the supplied file, so a claim can be traced back to what the model was given. - Inline links are proposed by a classifier and then reviewed: the pipeline reports each proposal, a deterministic scanner rejects known bad pairings, and a human reverts anything remaining that reads wrong.
Nothing here is peer reviewed. An article page summarizes one paper; it is not a replication, and it inherits every limitation of that paper. Concept pages are syntheses written across many articles, so their claims are only as good as the corpus they draw on. AI-generated text can be wrong in ways that read fluently, and the disclaimer in the offline editions says so plainly. Where a page's evidence is contested, the page says so rather than resolving it.
Illustrations, diagrams and the book cover are produced by code from the site's own data
and geometry — for example tooling/gen-epub-cover.mjs draws the cover from the concept
map — not by an image-generation model. Where a figure is ever produced by a generative
image model, that will be disclosed on the page that uses it, since publishers and
platform disclosure rules treat generated images as their own category.
Anyone who republishes this corpus, in whole or in part, is welcome to reuse this statement and should keep the parts that remain true of what they publish: the models used, the role of any AI system, the extent of human review, and the fact that the text is AI-generated. The license is CC0, so no permission is needed.