Anthropic confirmed on 11 August that Claude models launched on or after 2 August weave an imperceptible watermark into generated text, and attach signed C2PA provenance metadata to supported file types like .svg, .png and .jpg. The marking applies worldwide.
The trigger is Article 50 of the EU AI Act, enforceable from 2 August, which requires generative AI providers to mark outputs in machine-readable formats. Anthropic was one of around 190 signatories to the EU Code of Practice on Transparency of AI-Generated Content in July 2026. Fines can reach €15 million or 3% of worldwide turnover.
What went down
Watermarking covers models launched from 2 August, applied globally.
The mark travels with the text when it is copied and pasted elsewhere.
Marks cover the API, Claude Code, Cowork and cloud deployments.
Code and hard facts carry few marks, since the wording is fixed.
A detection API is coming, and older models follow over coming months.
Partner message: Get 50% off your first month of Hedra
Most agents stop at a document. Hedra Agent 2 keeps going and makes the thing.
It works in a Space, a persistent workspace the agent reads from and writes to as it goes. Give it a starting point and it asks clarifying questions, proposes a plan, researches the web, and writes the brief or script onto the canvas in front of you.
It pulls from Notion, writes into a Google Doc, runs code in a sandbox, saves repeat requests as Skills, and runs whole workflows on a schedule. Then it picks the right media model and produces the finished video, image or audio from the same session.
Try it at hedra.com/agentic using code agentic50 for 50% off your first month.
Why does this matter?
The watermark only attaches to words Claude chooses. Light proofreading of your own copy may leave too little to register, while a translation carries a full mark because every word is Claude's. For consultants and comms teams, the depth of AI involvement is now recorded inside the deliverable.
Only Anthropic holds the key, and the test answers one question. It cannot confirm text was human-written, cannot tell whether a different AI wrote it, and works poorly on short samples. Universities and employers will treat a positive result as proof regardless.
Our take
The policy itself is narrow. The watermark carries no information about the user, their organisation or their chats, and Anthropic cites DeepMind's SynthID-Text research showing no measurable quality difference. What it delivers is a compliance answer for Brussels and a signal institutions will over-interpret.
Social media has had endless hot takes. Reddit posters called it a conspiracy against Claude users, and Business Insider reported dozens of users on X claiming to cancel subscriptions. Google has watermarked AI images since 2023 and extended the approach to text, audio and video. OpenAI has held text watermarking technology for years without shipping it, with internal arguments about false positives, circumvention and losing users to competitors winning out.
Around 200 companies signed the same code, so the rest will arrive eventually. Anthropic moving first makes it the test case, and if detection proves reliable enough for HR departments and exam boards, the pressure lands on the key rather than the law.
Claude is the best LLM at writing to our knowledge, and a lot of agencies publishing web copy will have been using it without much concern. However, with watermarking, they may be more cautious of using Claude within workflows.
Another big thing… OpenAI pauses Astra over cyber risk
On 7 August, OpenAI told Axios it "cannot rule out critical cyber capabilities" in Astra, an unreleased model, and has halted internal work that does not meet strengthened security controls. Under its Preparedness Framework, the critical tier applies to a model that can find and build working zero-day exploits against hardened production systems unaided, or run novel attack campaigns from a stated objective.
Before any deployment, OpenAI says it will ask government agencies and independent safety organisations to evaluate Astra's cyber capabilities. Sam Altman wrote on X that the company still intends a wide release, saying he does not think restricting powerful models to a chosen few is a good strategy. A lab stopping its own work on capability grounds, before release and before any regulator demanded it, is worth watching for whether the pause holds.
