Anthropic, OpenAI, and Google to mark generated texts with invisible AI watermarks

Anthropic updated its documentation to state that Claude will embed invisible watermarks in generated text at the model level, which survive copy-paste and edits, and will also add C2PA metadata to generated files and images. This will apply to all Claude models launched on or after August 2, 2026, with a transition period for current models, seemingly to align with the European AI Act, and it will be deployed globally, not just in the EU. Article 50 of the AI Act, effective from August 2, 2026, mandates generative AI providers to mark output as machine-readable synthetic content and requires deployers to disclose deepfakes and generated texts on public-interest topics. The European Commission issued a code of practice in two sections (providers and deployers) with voluntary signing but mandatory compliance, with fines up to €15 million or 3% of global turnover. About 190 organizations have signed, including Google, Microsoft, OpenAI, Anthropic, Mistral, and Meta (banned in Russia as extremist) in the provider section, while deployers like Getty Images, Lufthansa, Bulgari, Lenovo, Iberdrola, and Greek news sites signed but not the big tech platforms. People cannot visually detect these watermarks; for example, Google’s SynthID hides watermarks in the pattern of word choices, not special invisible characters. The text also notes that AI-generated text may be exempt from labeling if it undergoes human editing with responsibility, and that current detectors like GPTZero are unreliable, as one test misidentified a Gemini output as human-written. OpenAI launched a public verifier with API for images and audio but not text yet, while Google provides SynthID detection only to select media and researchers. The article mentions that X automatically labels AI content, and LinkedIn has a button for content resembling AI slop.