Anthropic confirms Claude text watermarking after EU AI Act requirement
On August 11, 2026 Anthropic announced that Claude models launched on or after August 2, 2026 will embed machine‑readable watermarks at the token level. The change is mandatory, reflecting Article 50 of the EU AI Act, which took effect on August 2 and imposes penalties of up to €15 million or 3 % of global turnover for non‑compliant providers. Around 190 organisations, including Anthropic,…
Key points
- Anthropic began embedding token-level watermarks in Claude models launched on or after Aug 2 2026, per its Aug 11 announcement.
- EU AI Act Article 50 mandates machine‑detectable marks, with penalties up to €15 million or 3 % of global turnover.
- Google’s SynthID‑Text watermark runs in Gemini, validated on ~20 million responses; OpenAI and Meta have not yet deployed text watermarks.
Google’s SynthID‑Text watermark has been running inside Gemini since 2024 and was validated in a live experiment covering nearly 20 million Gemini responses, published in Nature in October 2024. Anthropic’s implementation follows the same technique, using a cryptographic key to bias token selection without altering the visible text. OpenAI and Meta have not confirmed text watermarks yet; OpenAI reportedly built a prototype in 2024 but shelved it over paraphrasing‑evasion concerns. The article notes that watermark signals are weaker in short or highly factual passages and can be degraded by heavy rewriting, which matters for developers building compliance‑sensitive applications.
The story so far
2 episodes →- Anthropic confirms Claude text watermarking after EU AI Act requirementthis story
How AI Watermarking Actually Works at the Token Level
hackernoon.com · 20 September 2026
Loading the full article…
This text was published by hackernoon.com and written by Anish Kumar Singh. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us.
More in Policy & Regulation
All →- Amazon blocks Meta’s Muse AI agent from crawling its marketplace · 1 src
- British Columbia sues OpenAI and Sam Altman over Tumbler Ridge school shooting · 1 src
- Trump creates AI Force and AI Czar, rejects industry calls to slow AI development · 13 src
- OpenAI agents breached Hugging Face using zero-day exploits, according to multiple sources · 3 src
- Anthropic partners with Novo Nordisk to use Claude models in drug discovery · 3 src
Comments
via GitHub Discussions