Hoskinson Challenge Claude Watermarking as Anthropic Roll Out Global AI Content Marks

Charles Hoskinson don raise concern about the new content-marking system for Claude after e show how human-authored material fit still receive AI provenance signal once model process am.

For one video wey dem publish on August 11, Hoskinson use text wey e talk say e write for 2016 for Cardano and e ask Claude to modify the presentation. E come question how persistent AI markings fit affect how people go take interpret authorship and digital provenance for future.

Anthropic Expands Claude Content Marks Worldwide

Article 50 of the EU AI Act go start to apply from August 2, 2026 and e require providers of certain generative AI systems to add machine-readable marks wey go make detection of AI-generated or manipulated content possible.

Anthropic implementation no dey limited to Europe. The company talk say markings go apply to output from supported Claude models anywhere wey dem offer Claude worldwide, including Claude, Claude Platform API, Claude Code, Claude Cowork and Claude Tag.

For text, Claude dey embed one imperceptible watermark directly inside generated content. Anthropic talk say the mark fit follow text when person copy and paste am enter another application and e fit still survive some edits wey dem do later. Supported files fit also carry signed provenance metadata based on the C2PA standard.

Anthropic still talk say to detect Claude mark no mean say na Claude originally author the material. Text wey person write somewhere else and later proofread, translate, summarize or otherwise process with Claude fit still carry the mark.

Hoskinson Tests Claude With His Own Cardano Text

Hoskinson demonstrate this distinction with material wey e talk say na e write am by himself like ten years ago.

Instead make e ask Claude to create the main text, e give the model one existing passage and request formatting changes. E concern be say material wey get clear human authorship fit later carry Claude provenance signal just because e pass through the model.

Hoskinson come extend the argument reach intellectual property, suggest say persistent provenance infrastructure fit create future wahala about who create content and how dem go take prove authorship.

That part of e argument no be Anthropic stated ownership policy. Anthropic current Consumer Terms talk say users retain their rights inside submitted inputs and say Anthropic assign to users any rights wey e fit get for generated outputs.

A Claude watermark therefore no dey currently give Anthropic ownership of marked material. E just indicate say content fit don pass through Claude, and Anthropic self caution say such detection no be conclusive proof of full provenance.

Hoskinson Links AI Provenance to Decentralized Training

Hoskinson connect the matter to broader argument for open-source and decentralized artificial intelligence.

Inside the video, e compare closed AI platforms with open models wey people fit inspect and modify their underlying software. E repeatedly frame the issue around who dey control the infrastructure wey dem take generate and verify digital content.

Toward the end of the broadcast, Hoskinson talk say after e finish e work on Midnight City, e intend to explore decentralized training of open-source AI models. E reference one approach wey resemble Folding@home, where computing resources from distributed participants fit contribute to model pre-training.

No Cardano AI product, funded initiative or technical roadmap show for the video. Hoskinson describe decentralized AI training as one direction wey e intend to investigate, no be completed project.

Anthropic rollout don already change one practical aspect of AI-assisted publishing: text wey supported Claude models process fit now retain machine-readable record of that interaction even after person move am enter another application. Hoskinson response add another development path to the discussion, as decentralized model training don now show as area wey e want pursue after Midnight City.