Hoskinson Challenges Claude Watermarking as Anthropic Rolls Out Global AI Content Marks

Charles Hoskinson has raised concerns about Claude’s new content-marking system after demonstrating how human-authored material can receive an AI provenance signal after being processed by the model.

In a video published on August 11, Hoskinson used text he said he wrote in 2016 for Cardano and asked Claude to modify its presentation. He then questioned how persistent AI markings could affect future interpretation of authorship and digital provenance.

Anthropic Expands Claude Content Marks Worldwide

Article 50 of the EU AI Act applies from August 2, 2026 and requires providers of certain generative AI systems to add machine-readable marks that enable detection of AI-generated or manipulated content.

Anthropic’s implementation is not limited to Europe. The company says markings apply to output from supported Claude models wherever Claude is offered worldwide, including Claude, Claude Platform API, Claude Code, Claude Cowork and Claude Tag.

For text, Claude embeds an imperceptible watermark directly into generated content. Anthropic says the mark can travel with text when it is copied and pasted into another application and may survive some subsequent editing. Supported files can also carry signed provenance metadata based on the C2PA standard.

Anthropic also states that detection of a Claude mark does not establish that Claude originally authored the material. Text that was written elsewhere and later proofread, translated, summarized or otherwise processed by Claude can still carry the mark.

Hoskinson Tests Claude With His Own Cardano Text

Hoskinson demonstrated that distinction using material he said he had written himself a decade earlier.

Rather than asking Claude to create the underlying text, he gave the model an existing passage and requested formatting changes. His concern was that material with clear human authorship could subsequently carry a Claude provenance signal simply because it passed through the model.

Hoskinson then extended the argument to intellectual property, suggesting that persistent provenance infrastructure could create future disputes over who created content and how authorship is demonstrated.

That part of his argument is not Anthropic’s stated ownership policy. Anthropic’s current Consumer Terms say users retain their rights in submitted inputs and that Anthropic assigns to users any rights it may have in generated outputs.

A Claude watermark therefore does not currently give Anthropic ownership of marked material. It indicates that content may have been processed by Claude, while Anthropic itself cautions that such detection is not conclusive proof of full provenance.

Hoskinson Links AI Provenance to Decentralized Training

Hoskinson connected the dispute to a broader argument for open-source and decentralized artificial intelligence.

During the video, he contrasted closed AI platforms with open models whose underlying software can be inspected and modified. He repeatedly framed the issue around who controls the infrastructure used to generate and verify digital content.

Toward the end of the broadcast, Hoskinson said that after his work on Midnight City he intends to explore decentralized training of open-source AI models. He referenced an approach comparable to Folding@home, where computing resources from distributed participants could contribute to model pre-training.

No Cardano AI product, funded initiative or technical roadmap was announced in the video. Hoskinson described decentralized AI training as a direction he intends to investigate rather than a completed project.

Anthropic’s rollout has already changed one practical aspect of AI-assisted publishing: text processed by supported Claude models can now retain a machine-readable record of that interaction even after being moved into another application. Hoskinson’s response adds a separate development path to the discussion, with decentralized model training now identified as an area he wants to pursue after Midnight City.