A watermark is a clue, not a verdict

A label on a photo can tell a person that a file was edited. Text is harder. A few copied sentences may pass through several tools, be rewritten by a person and lose any obvious trace of where they began.

Anthropic says future Claude models will generate text with a watermark that can make Claude’s involvement statistically detectable. The company is clear about the boundary: the mark is intended to answer whether Claude was likely involved at some point. It cannot say whether Claude wrote the whole piece, whether a person rewrote it or who used the tool.

That distinction is easy to miss in a debate about AI detection. A provenance signal can help establish part of a history. It is not a machine that decides authorship, intent or responsibility.

How Anthropic says the mark works

Anthropic says its approach is based on the SynthID-Text method published by Google DeepMind. It does not add hidden characters or visible words. Instead, when several ordinary wording choices would fit, the system uses a secret key to influence the source of the randomness behind the choice.

Over a longer passage, those small choices can form a pattern that a detector with the key can test. Anthropic says the output should look and read the same to a person, and says its internal testing found no practical effect on quality, creativity or readability. Those are company statements rather than an independent audit of every use case.

The mechanism deliberately has less room to operate when there is one exact answer. Facts, precise names, equations and much of executable code offer fewer interchangeable choices. Anthropic says watermarks will therefore be sparser in those cases.

Short text and heavy editing are difficult cases

A detector needs enough choices to see a pattern. Anthropic says confidence rises with the length of a passage and that small samples may not contain enough evidence. Light editing may not fully remove a signal, while a complete rewrite can.

That makes this useful for a narrow job: checking whether a sufficiently long piece is statistically consistent with Claude having helped produce it. It is a poor fit for declaring that a tweet, a code snippet or a carefully revised document definitely came from one source.

Anthropic also says the watermark contains no information about the person, organisation or chat behind the text. That is important for privacy. It also means the technical signal cannot answer the social question people often care about most: who actually made the decision to publish the words.

The EU context is real, but the technical answer is still plural

The European Commission says Article 50 transparency requirements began applying on 2 August 2026. The rules cover machine-readable marking of AI-generated or manipulated content, alongside disclosure duties in defined situations. The Commission’s associated code of practice gives signatories one recognised way to demonstrate compliance, but it is voluntary and does not replace the underlying law.

Anthropic says it is applying the watermark globally at launch because it does not yet have a durable way to limit it by region. It also says older models have a transition period and will be updated over coming months. The company plans a detection API, but has not yet published its implementation details.

There will not be one universal detector. Each provider can use a different method and key. A Claude detector cannot prove that text was written by another model, and a negative result cannot prove a human wrote it. That is why provenance has to be interpreted alongside source records, editorial process and human judgement.

What is confirmed, what Anthropic says, and what is open

Confirmed: Anthropic published its explanation on 14 August 2026. It says future Claude models will use a text watermark, that it is based on SynthID-Text principles and that a detection API is planned. The European Commission says relevant Article 50 transparency obligations have applied since 2 August.

Anthropic’s claims: the watermark does not materially affect output quality, cost or speed; it does not identify a user or organisation; and it can indicate likely Claude involvement in a sufficiently long passage.

Open questions: the detector’s public accuracy and error rates, practical thresholds for a meaningful result, how it handles translation and editing in the wild, the implementation timetable for each model and how organisations will combine technical signals with fair human review.

Sources

  1. Anthropic — How Claude’s text watermark worksPrimary Anthropic announcement, published 14 August 2026. Source for the proposed watermark method, stated limits, implementation plans, privacy statements and detection-API plan.
  2. European Commission — Code of Practice on Transparency of AI-generated ContentPrimary European Commission policy page. Source for the Article 50 timeline, the legal transparency obligations, the voluntary code and its scope.
  3. European Commission — Guidelines on transparency obligations for providers and deployersPrimary Commission guidance. Source for the distinction between machine-readable marking by providers and disclosure duties for defined AI-generated public-interest content.