Anthropic is watermarking text generated by Claude to comply with EU law
Anthropic is introducing invisible watermarking for Claude-generated text to meet the European Union’s new AI transparency rules. The system is intended to let authorised parties identify likely Claude involvement without visibly altering writing, adding hidden characters or reducing output quality.
Claude will use a secret key to guide otherwise plausible word choices, creating a detectable pattern based on Google DeepMind’s SynthID-Text approach. Anthropic plans an API for checking text, though short passages, light edits and exact-output code may be difficult or impossible to identify; it will also label Claude-generated images through cryptographically signed metadata. Watermarking will apply globally to outputs from models released after 2 August, with older models to follow.
- Anthropic will invisibly watermark Claude text to meet EU transparency rules.
- Detection indicates likely Claude involvement, not necessarily full authorship.
- The watermarking rollout applies globally, including image metadata.
New here? Start with this
Anthropic is a US artificial intelligence company whose Claude chatbot can write text, answer questions and help with tasks such as drafting documents or computer code. The European Union is introducing rules intended to make it clearer when material has been created by AI, particularly where people might otherwise mistake it for human work.
A watermark is a hidden signal placed in an AI system’s output so that a checking tool can look for signs of its origin without changing how the text appears to readers. Google DeepMind developed one approach, called SynthID-Text, which uses small patterns in word choices; such systems can be less reliable if text is very short or has been substantially edited.
The rules matter because AI-generated material is becoming easier to produce and harder to distinguish from writing, images and other content made by people. They also raise practical questions about how reliably AI output can be identified, who can carry out those checks, and how transparency requirements apply across different countries.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Supporters argue that robust provenance tools are a proportionate way to meet the EU’s transparency aims while preserving the usefulness of AI writing. Invisible watermarking can help platforms, researchers and regulators distinguish likely AI-generated material from human work, improving accountability around misinformation, impersonation and undisclosed automated content without imposing visible labels on every reader. Applying it globally may also provide a clearer, more consistent standard than maintaining different safeguards by region.
The case against
Critics argue that text watermarking offers only limited and potentially misleading assurance, since short extracts, ordinary editing and certain outputs may evade detection while false confidence in a checker could still affect people’s reputations or decisions. Secret-key systems also concentrate power in the provider or authorised verifiers, raising questions about independent scrutiny, due process and who may inspect private text. A worldwide rollout in response to one jurisdiction’s rules may impose a contested design choice on users elsewhere and could chill legitimate anonymous or assistive uses of AI.