By Greg Nowak. Last updated 2026-09-22.
When AI disclosure decisions live in a spreadsheet, the evidence starts drifting away from the content almost immediately. A page changes, an asset is replaced, or someone publishes through an import—and the row that once looked reassuring no longer describes what is live.
This has become a practical publishing concern. Article 50 of the EU AI Act has applied since 2 August 2026. It covers areas including deepfakes and AI-generated or manipulated text published to inform the public about matters of public interest. For that text, the disclosure obligation does not apply when the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility.
The rules do not create a blanket requirement to label every sentence touched by AI. They do make the quality of your review, the identity of the responsible publisher, and the connection between evidence and the approved version much more consequential. Your CMS is the sensible place to manage that connection.
Start with the publishing decision
“Was AI used?” is too crude to drive a useful workflow. An editor might use AI to suggest headings, translate approved copy, produce most of a first draft, or create a realistic image. Those activities do not carry the same editorial, legal, or reputational implications.
Define the decisions your team needs to make before adding fields or installing a plugin. Which uses must be recorded? What constitutes a substantive review? Who may accept editorial responsibility? Which cases require a public label or specialist advice? The answers should reflect your content, audience, role under the law, and legal guidance.
| CMS record | What it should answer | Practical options |
|---|---|---|
| AI role | How did AI contribute? | Ideation, drafting, translation, transformation, generated media |
| Content risk | Does the subject or format require escalation? | Routine, public interest, sensitive claim, realistic synthetic media |
| Review | Who checked the substance, and when? | Reviewer, date, source notes, review outcome |
| Disclosure decision | What must the audience be told? | Not required, required and present, specialist review needed |
| Approved version | Which revision does the decision cover? | Revision ID, approval date, editorial owner |
Collect evidence editors can realistically maintain
Use an AI role field with plain, limited choices instead of a free-text essay. Add a short review record that identifies the reviewer and covers factual accuracy, source reliability, misleading omissions, and fitness for the intended audience. Grammar correction alone is not a meaningful substantive review.
Store the disclosure decision separately from the AI role. AI use does not automatically determine the outcome. Centralise approved label wording where possible, while still allowing an explanation when an unusual case needs legal or editorial judgment.
For generated or manipulated images, audio, and video, keep the source asset, available provenance metadata, intended context, and approval together. Image optimisation and digital asset management can otherwise break the chain between the original file and the published derivative. This deserves its own process alongside any visible label.
Finally, name an editorial owner. “Human reviewed” is weak operational evidence when nobody can identify the person or organisation that accepted responsibility for publication.
Make the workflow enforce the policy
A practical sequence is Draft → Content review → Disclosure decision → Ready to publish → Published. The CMS should prevent the final transition when required evidence is missing or a case remains unresolved. A material edit after approval should return the page to review rather than silently inherit an old decision.
Apply the same checks to every publishing route. Scheduled posts, bulk updates, API imports, translations, migrations, and agency handovers can bypass controls that exist only as prompts in the browser. Validate critical requirements on the server and record the final revision that passed them.
How this fits WordPress
In WordPress, register structured governance fields as post metadata with appropriate data types, sanitisation, and capability checks. Enable revision support for metadata that must remain tied to the reviewed version. Show the essential fields in the editor, then add useful columns and filters to the Posts screen so an operations lead can find missing decisions and stalled reviews.
A small team may be able to retain the familiar Pending Review status and add validation around publication. Introduce custom statuses only when they improve routing or permissions. Whatever interface you choose, make sure programmatic publishing receives the same validation as a person clicking Publish.
How this fits Drupal
Drupal already supplies much of the foundation through fields, revisions, Workflows, and Content Moderation. Add governance fields to the relevant content types, create only the moderation states your team will use, and restrict sensitive transitions to appropriate roles. A View can become the working queue for content awaiting review or disclosure decisions.
Revisions are especially valuable here: the approval must identify the exact version that was reviewed. A moderation state alone cannot explain why a disclosure decision was made, so keep the evidence in structured fields attached to the content revision.
Put search quality through the same gate
Google’s current guidance does not treat generative AI assistance as an automatic search violation. It asks publishers to focus on accuracy, quality, relevance, and useful context. Generating many pages without adding value can fall under its scaled content abuse policy.
Add four editorial questions before approval: Does this page answer a distinct audience need? Does it contain useful knowledge or judgment beyond a generic summary? Is it substantially different from existing pages? Can its important claims be supported? For a proposed batch, review representative pages and overlap before producing the remainder. A disclosure label cannot make thin content useful.
Roll it out without creating form fatigue
Begin with one content type and a representative sample of real work. Include normal articles, edited AI drafts, translations, generated media, and at least one difficult edge case. Watch where editors hesitate, which fields they skip, and whether reviewers can reconstruct the final decision without opening another system.
Then extend the model to other publishing routes and backfill the live content where the risk justifies the effort. Owners get a visible control point, operations teams get an actionable queue, and agencies can hand over a decision trail that remains attached to the website.
If your AI-assisted publishing process has outgrown informal checks, Greg can help translate the policy into fields, roles, review states, queues, and publication safeguards. Start with a focused CMS workflow review.
Related on GrN.dk
- AI Images Need a Chain of Custody, Not Just a Disclosure Label
- Google’s AI Search Toggle Is a Publishing Decision, Not an SEO Setting
- Before AI images reach clients, audit your media library
Need help with this kind of work?
Plan your CMS workflow with Greg Get in touch with Greg.
