An AI-assisted blog: what has to be disclosed

The blog has been running for two years. A draft comes out of a chat window, somebody from the editorial team reads it over, then the post goes online. Since 2 August 2026 Article 50 of the AI Act has applied, and the question now arrives in writing. It gets expensive because the answer is not in the text but in a piece of information nobody ever wrote down — and because judging 340 existing posts individually after the fact costs a multiple of what one maintained table would have cost.

How to go about it in JMX

1. Settings → Content origin. One row per area of the site, with the same patterns as the segments: path prefix, URL component, regular expression or host. Plus the origin: Written without AI, AI-assisted, editorially reviewed, AI-generated or Machine-translated. For the blog described here the second one is right.
2. Exceptions go on top. The first matching row applies. So /blog/press-releases belongs above /blog, otherwise the row never takes effect. Into the note field goes which tool is used and who signs off.
3. Crawl, then the Issues tab. Filter to the AI disclosure category. Ten rules live there, two of them switched off by default.
4. Read the three comparison findings. AI.ORIGIN.UNDISCLOSED (error) reports declared AI generation without a disclosure. AI.ORIGIN.NO_EDITORIAL_OWNER (warning) is the finding for your case: the exemption in paragraph 4 presupposes a named responsible person. AI.ORIGIN.UNDECLARED_DISCLOSURE (notice) finds the template note on pages it does not apply to.
5. Check the two artifacts. AI.TEXT.PROMPT_ARTIFACT and AI.TEXT.INVISIBLE_MARKERS apply regardless of any declaration. A forgotten sentence from the chat window is on the page and therefore in the Google index.
6. The Tasks tab, then the client report. There the findings turn into a work plan with estimated effort; the report has its own AI disclosure section, and AI.ORIGIN.COVERAGE stands in front of it as the tally — how many pages are declared, for how many a disclosure applies, how many carry one.

What to watch out for

This post deliberately does not quote the tell-tale sentences verbatim. Write them out and you have them on your page, and the rule does not distinguish whether what stands there is an example or a leftover — it reports what is in the source, and that is its strength and at the same time its only blind spot.

The series score is no proof of AI. What is measured is uniformity — narrow sentence lengths, paragraphs of equal length, no long sentences at all — from 250 words upward, in German and English only, reported as a notice. The site-wide rule only judges from 30 percent of flagged pages and at least ten pages. A hand-written text can trigger it. That text is then bulk-produced copy, “scaled content abuse” to Google since 2024, but no proof about its origin. As soon as something is declared for an area, this check goes completely quiet there.

The limit of the method lies in the declaration itself: JMX checks your statement against the page, not your statement against reality. Enter “editorially reviewed” and do not review, and you get a green list. Whether a responsible person is named is something JMX measures — in the text or in the structured data; whether they actually read it, no crawler can establish. The machine-readable marking from paragraph 2 is also owed by the provider of the AI system, not by you; JMX keeps the two apart.

Two special cases. The order of the rows decides their meaning, and they cannot currently be moved — a row filed in the wrong place is deleted and created anew, and a row completely covered by one above it is not reported by JMX. And the coverage figure is an info-level finding: the Tasks tab deliberately does not take up info, so you find the tally only under Issues and in the report.

Enter the declaration only once the enquiry has arrived and you are reconstructing two years of editorial routine from the memory of three people; enter it today and you print a section from the client report.