When six AI heavyweights sign a European transparency code, the real news isn't the endorsement itself, but the commitment's reach into locally run models. Anthropic, OpenAI, Google, Meta, Microsoft, and Mistral have all agreed to the EU Code of Practice on traceability of generative content — a structural shift that brings watermarking obligations right into the company server, the edge device, or the on-premise cluster. It is no longer a matter confined to cloud APIs.

The core issue is that even the open source and local models from these companies will have to embed marking techniques for generated text and code. Meta, for instance, has already opened up with the Llama family; the EU pledge now means that future releases — and likely existing distributions through updates — will incorporate watermarking at inference time. This is not an optional add-on, but an impending regulatory requirement, given the Code of Conduct’s role as a precursor to the AI Act.

Filigrana without the cloud: the technical hurdle and the local infrastructure burden

For those who chose on-premise deployment for data sovereignty, latency control, or operational cost management, the outlook is nontrivial. Applying a watermark to every generated token requires adding a software module — typically a shared library or a middleware layer in the serving stack — that alters the model’s output probabilities with subtle statistical patterns imperceptible to users but detectable externally. In a self-hosted setup, the system administrator will have to ensure the presence and integrity of this marking component, without unduly affecting inference latency.

The computational overhead may not be huge (it involves slightly shifting token distributions, a constant-time operation), but it introduces an additional attack surface and software dependency. It also raises a question: what happens when an organization fine-tunes the model on proprietary data? The watermarking mechanism must be resilient to further training, or it loses effectiveness. Signatories have committed to developing robust techniques, yet verification in local settings will be tricky.

The stakes go beyond mere compliance. Extending the rule to the local sphere signals that regulators intend to leave no safe harbors: generative AI is considered pervasive enough to demand transparency wherever it runs. For large enterprises, this moves the compliance boundary — it will no longer be enough to pick a certified cloud provider; traceability will have to be factored into proprietary data centers as well. Anyone creating open source forks that strip out the watermarking faces growing legal risk, effectively producing a bifurcation between compliant and non-compliant ecosystems, with the enterprise market inevitably gravitating toward the former.

At AI-RADAR we track the evolution of inference frameworks and deployment stacks closely: native integration of watermarking mechanisms into on-premise runtimes will soon become a key differentiator in TCO and architectural assessments. It is no longer just about GPUs and VRAM, but about how the serving software handles the full lifecycle of model accountability.