EU code signatories set up a marking and detection task force
Published 13 October 2026
The European Commission now says that on 30 September 2026 over 90 signatories of the Code of Practice on Transparency of AI-generated Content set up two signatory task forces, one on marking and detection and one on labelling, chaired by the AI Office[16]. The marking and detection task force gathers providers of AI systems and models and makers of marking and detection technology. Its stated objectives include interoperable marking and detection, detection that is neutral to the AI provider, public benchmarks, and red-teaming of state-of-the-art detectors[16].
What changed
The Commission's page on the code gained a link to a new page on the task forces; both pages carry a last-update line of 12 October 2026[2][16]. The code's page had said signatories would collaborate in task forces still to be set up[2]. The new page says they exist: the signatories agreed their rules and procedures, both task forces held plenary meetings on 30 September and 1 October 2026, and minutes are to be published[16]. A working group on interoperability, meant to let laypersons check whether content is AI-generated without trying several detectors in turn, is to launch in October or November 2026[16].
What it means
Nothing in the tracker moves. The three text watermarks it follows are separate schemes: Google's SynthID Text, Anthropic's version of it, and OpenAI's own textGrain[4][10][15]. No public detector exists for Anthropic's or OpenAI's mark, and Google's open library needs the watermarking key[1][15][4]. The task force's objectives name exactly that gap, but they are objectives, not results[16]. What Article 50 and the code require.
What is still not known
The page does not name the signatories that joined the task forces or say which providers sit in the marking and detection group. It gives no dates for deliverables beyond the working-group launches, and the plenary minutes are not yet published[16]. Whether any of this work will give the public a way to check text is not stated. The tracker.
Sources
- How Claude marks AI-generated content primary
- Code of Practice on Transparency of AI-generated Content primary
- google-deepmind/synthid-text primary
- How Claude's text watermark works primary
- Provenance signals in OpenAI-generated content primary
- Signatory Task Forces of the Code of Practice on Transparency of AI-generated Content primary
Last verified against primary sources: