Beyond the hard text of the AI Act, Brussels is betting on a Code of Practice on transparency of AI-generated content as a bridge between the rule and engineering. In September 2026, public registers and industry coverage place the signatory count near 190 organizations—model providers, platforms, provenance tools, and actors across the content supply chain.
Google has been among the most visible signatories: it confirms participation in the provider section and contributes the SynthID triad (watermarking for text, image, audio, and video), associated detection tools, and C2PA support for provenance metadata. For many European teams, SynthID is the practical reference when people say “machine-readable marking.”
OpenAI, Meta, and Microsoft appear with their own approaches—mixes of classifiers, metadata, product policies, and, to varying degrees, watermarks. The Code does not mandate a single algorithm; it mandates process, documentation, and cooperation commitments that authorities can use when assessing diligence.
Signing is voluntary, but not empty cosmetics. In the European logic, adhering to the approved code offers a structured path to show Article 50 conformity efforts—especially useful while technical guidance and national inspector capacity are still maturing. It does not immunize against fines if real practice fails.
Collective value lies in interoperability. If every signatory invents an incompatible detector, journalists and platforms face a forensic Tower of Babel. That is why the February 2027 horizon—interoperable detection mechanisms—keeps recurring in research notes and in the Board’s own debate.
Neural Pulse sees a pattern: hyperscalers sign early and publish blogs; European content-generation SMEs arrive later, often via associations or cloud providers. The risk is two-speed transparency where only the large can prove conformity with pretty dashboards.
For newsrooms and OSINT, the Code improves the map of interlocutors: there is a register to point to when a deepfake circulates and someone needs to know which provider might issue a detection verdict. Human chain of custody is still required; the code does not replace journalism.
Competitively, signing is also trust marketing toward governments and advertisers. After years of synthetic-content scandals in elections and fraud, being able to say “we are EU Code signatories” enters public-sector RFPs and brand-safety policies.
Honest limits: the Code does not cover every offshore malicious use, bind third countries, or defeat sophisticated adversarial scrubbing. It is market-governance infrastructure, not a magic firewall. Anyone selling it as a total solution is overclaiming.
Interaction with the 2 December grace deadline is direct: signatories under remarking pressure use the Code as a shared checklist—inventory, methods, tests, authority contact. The ~190 signatures are, in that sense, an implementation network as much as a reputation club.
Open-source and downloadable models complicate the picture: a weight on Hugging Face may be marked by the publisher and then fine-tuned by a third party who signs nothing. The Code pushes documentation of that boundary; it does not erase it.
Neural Pulse closes by recommending three actions for European enterprise deployments: check whether their model and CMS providers appear in the register; contractually require the concrete commitments signed (not just the logo); and join detector-interoperability trials before February 2027.
One hundred ninety signatures do not guarantee a labeled internet. They do indicate that transparency left the whitepaper stage and became a list with names—and, soon, with audit dates.
