Anthropic's Watermarking Solves Anthropic's Problem, Not Yours

The disclosure duty reaching South African companies has not moved, whatever Claude now embeds in the text it produces.

Sign up for my Substack daily AI newsletter here.

See my AI Training course portfolio for corporate Business Leaders here.

Article link: https://johanosteyn.substack.com/p/anthropics-watermarking-solves-anthropics?r=73gqa&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/

A vendor announcement is not a compliance event. When a supplier reports that it has met a regulatory obligation, the natural executive response is relief, and in this case the relief would be misplaced. Anthropic has done something significant this month, and it has done it for itself.

CONTEXT AND BACKGROUND

Anthropic has confirmed that Claude models launched on or after 2 August 2026 weave an imperceptible watermark directly into generated text, while supported files carry signed provenance metadata built on the C2PA standard. The mark is applied at model level, so it travels across the Claude applications, the developer platform, Claude Code, Claude Cowork and Claude Tag, and the company has switched it on worldwide rather than only for European users. The driver is Article 50 of the European Union’s Artificial Intelligence Act, whose transparency obligations became enforceable on 2 August 2026. South African commentary has concentrated on whether the mark can be trusted as evidence of authorship, and local experts have warned that it cannot be treated as definitive. That question matters. It is not, however, the question that should be occupying boards.

INSIGHT AND ANALYSIS

The Act does not impose one duty on everyone in the chain. It divides responsibility. Providers, meaning the firms that build and supply the models, must ensure outputs are marked in a machine-readable format detectable as artificially generated. Deployers, meaning the far larger population of organisations that use those systems to create and publish content, carry separate disclosure duties, and the territorial trigger is the location of use rather than the location of the organisation. Anthropic’s watermark discharges Anthropic’s obligation. It discharges none of yours. Where a deployer publishes a deepfake or AI-generated text on a matter of public interest, the requirement is a clear visual label at the point of first exposure, and the narrow editorial exception applies only where a person or organisation exercises genuine editorial responsibility, because routine or pro-forma review does not suffice.

IMPLICATIONS

Consider the timing. Systems already on the market before 2 August 2026 were given until 2 December 2026 to implement the marking requirement, so the provider duty was eased. The deployer disclosure duties were not postponed. The party with the least awareness of its obligation received the least accommodation. A machine-readable mark is also not a visible label, and turning one into the other falls largely on the organisation that publishes the content rather than on the model provider.

Anthropic has meanwhile not released the detection tooling the European code of practice calls for, so the signal exists while the means of testing it does not. Compounding matters, content can carry a mark where Claude was used solely to proofread, format or translate human-written copy.

I have previously written about this in a piece on charting Africa’s path to ethical and secure AI adoption, which argued that the Act’s extraterritorial reach makes adoption of its standards a matter of obligation rather than choice for African corporations.

CLOSING TAKEAWAY

Three steps follow. Establish, for every use of AI in the organisation, whether you are acting as a provider or a deployer, since the answer determines which duty applies. Name the person who carries editorial responsibility for AI-assisted content reaching the public, and ensure that review is real rather than nominal. Write into policy that a detected mark evidences processing and not authorship, before the first dispute rather than after it. None of this is exotic governance. It is the principle King IV and King V already require, which is that a named human being owns the output whatever tool produced it. Anthropic has solved its own compliance problem elegantly. Yours sits exactly where it sat on the first of August.

Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net