A site about synthetic media owes you a provenance statement about itself.

Why imadethisup.org publishes an AI disclosure at all

This project exists to argue that provenance should be disclosed rather than assumed. It would be incoherent to make that argument while staying silent about how our own pages are made. Most sites are not yet expected to publish a disclosure like this one. We think that expectation is coming, and we would rather be early than defensive.

This page is written to be specific. “AI was used responsibly” is not a disclosure; it is a slogan. What follows is the actual division of labour.

Where AI is used

Research triage. Large language models are used to survey a topic, surface candidate sources, and summarise long primary documents, agency releases, final rules, court opinions, and peer-reviewed papers, so that a human editor can decide which ones are worth reading in full. The model narrows the field. It does not decide what is true.

Drafting and structural editing. Article drafts, headings, summaries, and the “In short” answer blocks at the top of each post are produced with model assistance from source material and editorial direction, then revised by a human editor. Where an answer block or FAQ summarises an article; it is composed only from claims the article already makes and already cites.

Site code and build tooling. The HTML, CSS, and the Node scripts in /scripts that generate this site's link-preview cards, structured data, and machine-readable indexes were written with AI assistance. The site is hand-maintained static HTML with no framework and no build step beyond those scripts.

Link-preview cards. The 1200×630 OpenGraph card for each article is produced by a deterministic template script that typesets the article's own headline, topic tag, and date. It is code, not an image model. No card contains generated imagery.

Weekly blog posts are published automatically. Once a week an automated editorial agent selects a topic, researches it, drafts it, verifies every citation against its primary source, and publishes it to this site without a human reading it first, provided every check passes. The pillar pages, glossary, references, and organisation pages are not produced this way; those remain hand-authored and human-reviewed. We disclose this because a site about synthetic media does not get to be vague about how its own content is made.

Where AI is not used

No synthetic photography, ever. This site publishes no AI-generated photographs, no photorealistic imagery, and no photographs of people at all. Every diagram is hand-authored SVG. A site that documents the harms of synthetic imagery does not get to decorate itself with it.

No unverified citations. Citations are never accepted from a model's output on trust. Every citation in an automatically published post is fetched from its primary source and checked, source by source, against the specific claim it is attached to. A post whose sourcing fails that check is not published at all, not trimmed and not published without the failing citation. For hand-authored pages a human performs the same check by opening every source. This is the discipline we describe at length in our writing on the cost of unverified AI citations, and it would be indefensible to describe that duty and then not perform it.

No synthetic quotations or attributed statements. Quotations from named individuals, agencies, and opinions are transcribed from the source document. A model is never asked to reconstruct or paraphrase what someone said and have it presented as a quotation.

No generated legal, investigative, or safety advice. The operational guidance on this site, the authentication protocol, the incident-response steps, the reporting workflows, derives from named federal guidance, published standards, and documented cases, all cited inline. It is not a model's improvisation. It remains educational material, not professional advice; see the disclaimer.

Editorial control over AI-published posts, without pre-publication reading

No person reads an automatically published post before it goes live. Editorial control is still exercised, and it is worth being precise about where, because "nobody read it" and "nobody decided anything" are different claims and only the first is true here.

Each post is written from a brief that a person prepared: the topic, the specific sources the post may draw on, the facts those sources were confirmed to state, and the angle. The agent cannot choose its own subject and cannot introduce a source that is not in the brief. The constraints it writes under, and the checks that decide whether the result publishes at all, were set by people. Global Cyber Institute, Inc. holds editorial responsibility for everything on this site, including posts no one read first.

When a post publishes, an email goes to the editor listing what went live, the claims it rests on, the verification verdict recorded for each, and every source with its link. The post stays live unless somebody raises a correction. Publication is the beginning of review here rather than the end of it, which is a weaker guarantee than pre-publication reading and is stated as the weaker thing it is.

We are deliberate about what that notification does and does not establish. Sending a message nobody is obliged to read shows that a message was sent, and nothing more. It does not mean a person read the post, and no page on this site will tell you otherwise. What it buys is speed: an editor who does read it can correct or withdraw a post quickly, and a reader who finds an error can trigger the same thing. If you find one, a broken link, or a citation that does not support the claim attached to it, write to info@imadethisup.org. Substantive corrections are noted on the page rather than quietly patched.

What "verified" means for an AI-published post

Five checks run before an automatically produced post reaches this site, and every one of them fails closed. A failure publishes nothing rather than publishing something weaker.

1. Citation verification. Each source is fetched and checked against the specific claim it supports. Any verdict other than "supports", including a source that cannot be reached, stops the run. An unverifiable citation is indistinguishable from a fabricated one.
2. Structural audit. Structured data must parse, the canonical must match the page's own route, no internal link may be dead, and the post must appear consistently across the sitemap, the feed, and the machine-readable index.
3. Editorial constraints. A repository-wide check enforces the site's editorial policies before anything is committed.
4. Duplication. A proposed topic that substantially repeats an existing post is rejected.
5. Cost ceiling. Each run is capped, and the cap is enforced before each step rather than measured afterwards.

A week with no post is an acceptable outcome. A week with a fabricated citation is not.

What a human checks on the hand-authored pages of imadethisup.org

The pillar pages, glossary, references, FAQ, and organisation pages are written and reviewed by people. Each passes the same four checks, performed by a person:

1. Every cited source is opened, read, and confirmed to say what the article claims it says.
2. Every statutory citation is checked for current status, enacted, pending, amended, or superseded. Where a bill has not become law, the article says so explicitly.
3. Every factual assertion is traced to a primary source: a peer-reviewed paper, an official statute, a federal agency document, or the originating news investigation.
4. The full text is read end to end for accuracy and tone, and the editor takes responsibility for it.

The limitation we cannot engineer away: language models fabricate confidently

Language models produce fluent, confident text regardless of whether the underlying claim is correct, and they can fabricate citations, dates, and metadata that look entirely ordinary. We have written about that failure mode in both file metadata and legal citations. Our verification process is designed on the assumption that the model will sometimes be wrong and will not signal it.

That process is human on the pages people write and automated on the posts the agent publishes, and neither kind of check is infallible. If you find an error, a broken link, or a citation that does not support the claim attached to it, please write to info@imadethisup.org. Corrections are reviewed weekly, and substantive corrections are noted on the page itself rather than quietly patched.

AI crawlers are welcome here

We do not restrict AI crawling. Our robots.txt explicitly permits GPTBot, OAI-SearchBot, ClaudeBot, Google-Extended, PerplexityBot, CCBot, Applebot-Extended and others, and we publish an llms.txt index plus a full-text export for machine consumers. The content is licensed CC BY-NC 4.0 and the purpose of the project is the broadest possible public reach. If an answer engine is going to summarise this subject for someone in a crisis; we would rather it summarised material that is sourced than material that is not.

We ask one thing in return, which the licence already requires: attribute the material and link back to the page, so the reader can reach the primary sources themselves.

Content provenance

Where we publish or republish media that carries C2PA Content Credentials, we preserve the manifest rather than stripping it. The reasoning behind that commitment is set out on the provenance page.

Changes to the AI disclosure

This page is versioned with the site and carries a last-updated date. If our use of these tools changes materially, this page changes first.