Anthropic is introducing an invisible SynthID-Text-based text watermark in new Claude models. No special characters, hidden spaces, or personal data are added to the text. The marking is formed directly during generation through a mathematically controlled choice of the next word or token.
However, this does not mean that any Claude text can already be identified without error. The technology works probabilistically, is harder to detect in short and factual responses, and may disappear after a complete rewrite.
Does Claude already label all its texts?
No. In a publication dated August 14, 2026, Anthropic says it is introducing the watermark in future Claude models. The company also plans to gradually add labeling to models released before August 2, 2026.
For such previously released systems, the European Commission has provided a limited transition period: the requirements of Article 50(2) of the AI Act on machine-readable labeling must be met from December 2, 2026.
Therefore, the claim that “all Claude responses already have a hidden signature” is not yet accurate. Whether labeling is present depends on the specific model, its release date, and the stage of technology deployment.
How a neural network leaves an invisible trace in text
A language model generates text sequentially. At each step, it estimates the probability of possible next tokens — words, parts of words, characters, or combinations of characters.
If only one exact option is allowed, there is practically nothing to choose from. But in an ordinary phrase, there are often several equally suitable continuations. For example, a sentence about cold weather can continue with “overcast,” “gray,” or another natural option.
Claude uses a version of the SynthID-Text method published by Google DeepMind in 2024. In simplified terms, the process looks like this:
- The model calculates a probability distribution for the next token.
- Eligible candidates are selected from this distribution.
- A secret key and the previous context generate pseudorandom values.
- The candidates go through a mathematical “tournament.”
- The selected token becomes part of a statistical pattern that continues throughout the text.
A single word proves nothing. The detector analyzes a sequence of tokens and checks how closely it matches the pattern that should emerge when a specific key is used.
Why readers do not notice the watermark
The marking does not force the model to insert meaningless words. The choice is made among options already allowed by the model's original distribution.
Anthropic reports that internal tests found no practical impact on the content, creativity, or readability of responses. The SynthID-Text study also found no statistically significant decline in user ratings in an experiment involving about 20 million Gemini responses.
This does not mean that every text watermark configuration is always neutral. The SynthID-Text authors explicitly describe a trade-off: increasing detectability may reduce text quality or diversity. For industrial use, a configuration designed to preserve quality with acceptable detectability is used.
Where labeling works less well
A watermark needs variation. The more natural ways the model has to continue a phrase, the more opportunities there are to form a statistically detectable pattern.
The signal becomes weaker in several cases:
- short response — the detector does not have enough tokens for a confident conclusion;
- precise facts — the correct wording may have no equivalent alternatives;
- program code — replacing a token can break the syntax or operation of a program;
- light editing — if Claude changed only a few words, there is almost nothing for the watermark to attach to;
- complete rewriting — a deep rewrite can remove the original statistical pattern.
At the same time, a translation created entirely by Claude should receive a label: in that case, the model itself chooses almost every word. But translating an already labeled text using another system, according to SynthID-Text research, can significantly weaken detection of the original mark.
What a watermark can actually prove
Detection of the mark means only that Claude likely participated in creating or substantially revising the text.
The labeling does not allow one to reliably determine:
- whether Claude wrote the entire material or only edited it;
- which user submitted the prompt;
- which account or chat the text came from;
- whether the content is true;
- who owns the copyright;
- whether the author violated the rules of an educational institution, employer, or platform.
This is a fundamental difference from ordinary AI detectors. Classifiers try to guess the origin of text based on style and statistical features of language. A watermark checks a pre-embedded mathematical signal using a key.
Even such a detector should provide a probability rather than an unconditional verdict. The SynthID-Text study includes a mechanism for declining to answer when there is insufficient data to maintain the specified error rate.
Watermark, AI detector, and C2PA: what is the difference
| Technology | What it checks | Where the signal is stored | Main limitation |
|---|---|---|---|
| Text watermark | Likely involvement of a specific model or system | In token selection statistics | Weakens with a short response or deep rewrite |
| Ordinary AI detector | Whether the text looks machine-generated | There is no pre-existing mark | False accusations and unstable results are possible |
| C2PA Content Credentials | How a file was created or modified | In cryptographically signed metadata | Metadata can disappear in a screenshot or when a file is converted |
Anthropic plans to use C2PA for supported files, including PNG, JPG, and SVG. Unlike a text watermark, this does not modify the file's content but adds a signed record to its metadata.
Why Anthropic is introducing labeling globally
Anthropic cites the requirements of the EU AI Act as the main reason. Article 50 requires providers of generative AI systems to ensure machine-readable marking of synthetic text, images, audio and video.
Anthropic says it will apply the watermark globally at launch, as it does not yet have a reliable way to limit it to the European region only.
But the AI Act requirements are more complex than the formula “all AI content must have a visible label”. The law distinguishes between two obligations:
- the provider of an AI system ensures machine-readable marking and detectability;
- the organisation publishing the content, in certain cases, must clearly inform the audience about the use of AI.
Visible labelling is required, in particular, for deepfakes and certain AI texts on matters of public interest. Such texts are exempt if the material has undergone meaningful human review, is subject to editorial control and a specific person or organisation is responsible for publication.
These requirements are examined in more detail in the article on AI content labelling in the EU. Current changes to deadlines and rules are collected in the overview of the AI Omnibus and AI Act.
What this means for businesses in Bulgaria and the EU
For companies, the key takeaway is not that AI texts now need to be “cleansed” of watermarks. Trying to hide content’s origin does not replace a proper editorial process and may only increase legal and reputational risks.
A practical workflow should look like this:
- Keep information about the model and tool used for generation.
- Document the purpose of the text and the publication channel.
- Check facts, links, dates, calculations and legally significant wording.
- Assign a person who can revise or reject the material.
- Determine whether visible labelling is required for a specific publication.
- Keep confirmation of editorial review and the responsible person.
A standard spelling check does not count as full human oversight. To be exempt from the obligation to explicitly label a text on a matter of public interest, an editor must assess its content, verify sources and have the actual right to revise or prohibit publication.
For an initial review of processes, you can use the checklist for preparing an AI project for EU requirements.
Can you check Claude text yourself?
As of 20 August 2026, Anthropic has not yet provided a publicly available detector for Claude watermarks. The company says it is preparing a separate API, but details and a launch date have not yet been published.
A third-party service without an Anthropic key will not be able to perform the same check. It may use its own classifier and assess the style of the text, but such a result cannot be presented as detection of an official Claude watermark.
Key takeaway
Claude’s invisible marking is not a secret user identifier or proof of “machine authorship” in the legal sense. It is a statistical mechanism that makes it possible to estimate the probability that a specific AI system was involved in creating the text.
For businesses, the technology does not eliminate the need for transparency, fact-checking and editorial responsibility. On the contrary, watermarks, C2PA and AI Act requirements are gradually turning content provenance into a separate element of corporate oversight.
If a company uses AI for its website, newsletters, customer support or automated publication of materials, you can discuss a process audit and the minimum set of measures needed to comply with EU requirements.
Sources and verification
Last reviewed: 20.08.2026