What Anthropic Did — and Why There Is No Off Switch
Starting August 2, 2026, all Claude models launched on or after that date began embedding an invisible watermark in all generated text. The technique used is a version of SynthID-Text, originally developed by Google DeepMind and described in a Nature paper in 2024. The principle is straightforward: when the model must choose between words of equivalent value — "grey" or "overcast", "however" or "nonetheless" — the choice is guided by a cryptographic key that creates a pattern detectable by whoever holds the key, but invisible to the human reader. The mark travels when text is copied and pasted. Light edits likely do not remove it. A complete word-by-word rewrite does.
Anthropic offers no opt-out option. The justification is compliance with Article 50 of the European AI Act and the Code of Practice on Transparency of AI-Generated Content, which the company signed alongside approximately 190 other organizations in July 2026. Enforcement is global — not limited to European Union users. The company plans to release a watermark detection API soon. A detected mark means only that Claude may have processed that content — not that it wrote it entirely. Someone who uses Claude to proofread or translate their own text will also have the result marked.
What Google Did — and What Remains Invisible
On August 14, 2026, Josh Woodward, Google's VP for Gemini, announced that users will be able to disable the visible mark — the sparkle icon that appears on AI-generated images, videos and music from the Nano Banana, Omni and Lyria models. The option is found under Settings, within the Media Watermark section, and will be available in Gemini and the Flow video editor, with Search support expected soon. In countries where local law requires the visible mark — such as India, South Korea and Vietnam — it remains mandatory.
What does not change: the invisible SynthID and C2PA metadata remain embedded in all generated content regardless of user settings. Google also open-sourced the Credentio library, allowing developers to integrate local C2PA verification into their own applications. SynthID has already watermarked more than 100 billion images, videos and audio files since its 2023 launch.
What OpenAI and Meta Do — and What the AI Act Requires
OpenAI has never used visible marks. On May 19, 2026, the company joined the C2PA standard as a Conforming Generator and integrated Google DeepMind's SynthID into images generated by ChatGPT, Codex and the API. The result is a dual-layer approach: C2PA metadata with detailed context and an invisible watermark resilient to transformations like screenshots and format conversion. Meta follows a similar path, using provenance labels without an explicit visual mark. OpenAI maintains a text detector with 99.9% accuracy developed two years ago but still not publicly released — reasons include how easily it can be circumvented by translation or rewriting, and the risk of false positives in educational contexts.
The AI Act, in its Article 50, requires providers of generative AI systems to mark content produced by models launched from August 2, 2026. Non-compliance can trigger fines of up to 15 million euros or 3% of global annual turnover. The grace period for existing systems runs until December 2026. The C2PA standard — maintained by the Coalition for Content Provenance and Authenticity, with more than 6,000 members including Google, Microsoft, Adobe, Meta, OpenAI, Sony and the BBC — is considered the primary compliance route, even without being named directly in the law.
Transparency for Whom?
The two decisions this week expose real tensions. Anthropic chose irrevocable obligation: every Claude output carries a mark the user cannot control, as a regulatory compliance requirement. Google chose layer separation: the visible mark is now a creative choice; the invisible one is non-negotiable. Neither approach resolves the central problem — watermarks are evidence of processing, not proof of authorship. A journalist who uses AI to review their text, a student who uses it to translate an article, a programmer who pastes generated code into a larger project — all produce marked content that may or may not be "AI-generated" depending on how the term is defined. The watermarking debate did not end in August 2026. It simply gained a concrete pair of cases to anchor the years ahead.
