Why applying Anthropic's model-level watermark to source code is a software sovereignty issue (and not just EU AI Act compliance)

Anthropic is about to start embedding an invisible statistical watermark into text generated by Claude. For ordinary text, this may seem harmless. For proprietary source code, I think the issue is very different. The real question is not whether Anthropic can claim copyright over Claude-generated code today. The real issue is control over software assets . If a watermark is embedded into source code, a third-party provider is introducing a proprietary provenance signal directly into files that belong to your core software production chain. Anthropic states that the effect on code is “negligible” in its official announcement . However, ordinary customers cannot independently audit that claim or determine the actual extent of watermarking in their own codebase. Because Anthropic controls both the watermarking mechanism and the detection key—while access to its Detection API remains in private preview—this creates a clear information asymmetry: the provider can detect a signal in your code that you cannot independently inspect, disable, or reliably remove. Furthermore, current guidance from the European Commission on the AI Act (see EU AI Act Guidance ) considers source code to be outside the scope of the watermarking obligation. If watermarking still reaches code because it is applied indiscriminately at the model level, this is no longer just a regulatory requirement—it is an architectural choice made by the provider. Which raises a straightforward question: Who can reasonably accept a third-party provider embedding a proprietary provenance mechanism into their software assets when they cannot independently audit it, disable it, or reliably remove it? Today, this watermark grants Anthropic no ownership rights over output code. But laws, terms of service, and evidentiary practices around AI will inevitably evolve. Introducing a third-party-controlled traceability mechanism into commercial source code isn't a trivial technical detail—it's a risk-management decision. Personally, I would rather keep full control over my software production chain. How are your teams handling source code provenance, third-party markers, and vendor lock-in risks in commercial repositories?