The European Commission on Wednesday finalized its Code of Practice under the EU AI Act, setting providers and deployers a framework for marking AI-generated text, images, audio and video.
Key Takeaways
The European Commission finalized a Code of Practice for marking AI-generated text, images, audio and video under the EU AI Act
AI Act transparency obligations require synthetic-content providers to mark outputs as machine-generated in a detectable format
OpenAI is rolling out an invisible watermark for ChatGPT and Codex outputs, according to Mashable
The Code of Practice does not certify which technical methods meet the requirement for detectable marking
The code sits under the transparency obligations section of the AI Act, detailed on the Commission’s digital strategy site. The act states that providers of systems that generate synthetic content must “ensure outputs are marked as machine-generated in a format that is detectable.”
For OpenAI, compliance is already an operational question, it is rolling out an invisible watermark, a hidden signal embedded in generated text, for ChatGPT and Codex outputs.
Why Labelling Became Unavoidable
General-purpose AI models fall under their own tier of the AI Act, with obligations that scale based on the compute used to train them and the risk the system poses once deployed.
Also Read: Amazon’s Bedrock Fixes GPT Access Gap For 5 Asian Markets, No More Throttling
The high-level summary maintained by the EU AI Act project notes the rules apply not just to providers inside the bloc but to any deployer, including those in third countries, whose AI output ends up used within the EU.
That extraterritorial reach means a company training models in the United States or elsewhere still has to comply if its generated content, text, images or synthetic audio, reaches EU users. The new Code of Practice gives providers a common reference point for how that marking should work in practice, rather than each company inventing its own labelling standard.
The Detection Gap Still Open
Watermarking text is harder than watermarking images or audio, because small edits to a sentence can strip an invisible signal without changing the meaning. OpenAI’s rollout shows one major lab testing exactly that kind of hidden marker now, just as the EU’s labelling rules take their final shape.
For regulators, enforcement turns on whether watermarks like OpenAI’s satisfy the Commission’s requirement that outputs be marked “in a format that is detectable,” which is not yet settled. The Code of Practice sets the obligation to mark content, but does not yet certify which technical methods meet the bar, leaving that gap for regulators and providers to close next.
Read Next: Freehand’s AI Agents NOW Pay Invoices Alone, No Human Gate, Via Stripe