Anthropic’s Claude adds statistical text watermarks to supported models, but detection remains restricted

Models and surfaces. Vendor-reported: Anthropic says new Claude models launched in the EU on or after August 2, 2026 support machine-readable marking at launch; currently supported models include Claude Fable 5.1 and Claude Mythos 5.1. Anthropic says marks apply across Claude Platform/API, Claude, Claude Code, Claude Cowork, Claude Tag, AWS, Google Cloud, and Microsoft Foundry when supported models are used. How Claude marks AI-generated content | Claude Help Center

Text vs file marking. Vendor-reported: Claude uses two mechanisms: embedded text watermarks and signed provenance metadata for supported generated files such as .svg, .png, and .jpg. The text mark is not conventional file metadata; Anthropic says it is part of the generated text’s statistical pattern and can travel through copy-paste. How Claude marks AI-generated content | Claude Help Center

Detection access. Vendor-reported: Anthropic says file credentials can be checked with a content checker, but text watermark detection is in private preview, available to eligible organizations such as regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, EU civil-society groups, and certain enterprises with verification obligations. How Claude marks AI-generated content | Claude Help Center

The public technical anchor is not a Claude-specific paper; it is Anthropic’s statement that Claude’s watermark is a version of SynthID-Text, originally published by Google DeepMind. Anthropic describes the method as changing “the source of randomness” used to choose among words, rather than inserting visible marks, Unicode artifacts, or extra tokens. How Claude's text watermarking works \ Anthropic

SynthID-Text belongs to probabilistic generation-time watermarking. In simplified terms, when a model samples among plausible next tokens, the watermarking processor biases sampling according to a keyed pseudo-random signal. Detection later tests whether the observed sequence contains a statistically unlikely surplus of keyed choices. The video’s red/green-word explanation tracks that intuition, although Anthropic says its implementation is a SynthID-Text variant rather than a simple static green-list scheme. The video transcript’s simplified explanation appears at 00:00:57–00:02:32.

Google’s public SynthID-Text materials are reproducible to a point: Google DeepMind has a reference GitHub implementation with watermarking and detection capabilities, tests, notebook examples, and mean, weighted-mean, and Bayesian scoring functions. The repository explicitly says it is a reference implementation and not intended for production use. GitHub - google-deepmind/synthid-text · GitHub

For implementation teams, the important integration point is sampling. Google’s developer documentation says no additional model training is required; a watermarking configuration can be passed into generation to activate a logits processor. SynthID: Tools for watermarking and detecting LLM-generated Text | Responsible Generative AI Toolkit | Google AI for Developers This means downstream developers cannot assume “same prompt, same model, same temperature” behavior if watermarking changes the decoding path. For deterministic or low-temperature workflows, factual completions, code generation, or exact formatting, the strength of the signal may vary because fewer semantically safe token choices are available; Anthropic makes this limitation explicit for factual passages and proofreading. How Claude's text watermarking works \ Anthropic

Evaluation remains the weak point for Claude specifically. Google’s 2024 Nature paper reports a production-scale SynthID-Text evaluation in Gemini, but the paper is funded and performed by Google DeepMind/Google, and it is not an audit of Anthropic’s deployed Claude watermark. Scalable watermarking for identifying large language model outputs | Nature Independent work on SynthID-Text gives mixed signals: ETH SRI researchers found SynthID-Text detectable in black-box-style testing and more resistant to spoofing than some schemes, but also easier to scrub than other state-of-the-art schemes for naïve adversaries. Probing Google DeepMind’s SynthID-Text Watermark | SRI Lab Another arXiv robustness study reports that paraphrasing, copy-paste modifications, and back-translation can significantly degrade SynthID detectability. Robustness Assessment and Enhancement of Text Watermarking for Google's SynthID

Text watermarking predates Anthropic’s rollout. Anthropic itself traces the family of approaches to Scott Aaronson’s 2022 proposal and says Claude’s implementation is based on Google DeepMind’s SynthID-Text. How Claude's text watermarking works \ Anthropic Google deployed SynthID-Text in Gemini and published the approach in Nature in 2024; its public materials describe SynthID as embedding imperceptible watermarks in AI-generated images, audio, text, or video. SynthID — Google DeepMind

The regulatory driver is explicit. EU official materials say Article 50 transparency obligations apply from August 2, 2026, and the EU Code of Practice supports compliance for marking, detection, and labeling of AI-generated content. Code of Practice on Transparency of AI-generated Content | Shaping Europe’s digital future Anthropic’s global rollout appears broader than the EU minimum because it says marking applies wherever supported Claude models are offered. How Claude marks AI-generated content | Claude Help Center

The field also has a history of failed or contested AI-text detection. OpenAI discontinued its 2023 AI Text Classifier because of low accuracy, and OpenAI’s own limitations page warned against using such classifiers as primary decision tools. New AI classifier for indicating AI-written text | OpenAI OpenAI later said text watermarking could be less robust against global tampering such as translation or rewording with another generative model and could stigmatize legitimate writing assistance by non-native English speakers. Understanding the source of what we see and hear online | OpenAI

The main limitation is unverifiability of the deployed Claude system. An independent September 2026 paper, Watermarks Without Verification, argues that the central governance problem is not watermarking itself but the lack of public verification: no public tool can test deployed systems, and independent researchers must evaluate open-source SynthID implementations rather than Anthropic’s production configuration. Watermarks Without Verification: AI Text Watermarking After the EU AI Act

False positives and false negatives remain central risks. Anthropic says lack of a detected mark does not mean content was not AI-generated, because text may have been generated by older models, heavily edited, paraphrased, translated, mixed into other writing, too short, or routed through unsupported surfaces. Conversely, a detected mark may reflect proofreading or translation rather than original authorship. How Claude marks AI-generated content | Claude Help Center

The business and academic safety concern is misinterpretation: a watermark is a provenance signal, not a misconduct finding, copyright assignment, authorship proof, or complete chain of custody. Organizations that automate adverse decisions from watermark detection alone would be over-reading the evidence.

For enterprises using Claude: inventory which Claude models and routes are used in production, especially if outputs are externally published, submitted to regulators, used in marketing, or incorporated into code/documentation. Anthropic says supported models carry marks even via major cloud partners. How Claude marks AI-generated content | Claude Help Center

For developers: treat watermarking as part of the decoding stack. If your product evaluates exact output distributions, reproducibility, code correctness, or regression baselines, record model version, route, sampling parameters, and whether the model is a marked supported model.

For legal, compliance, education, and media teams: do not interpret “Claude watermark detected” as “Claude wrote this entire document.” Policies should distinguish drafting, editing, translation, summarization, and formatting. Anthropic’s own limitation language supports that distinction. How Claude marks AI-generated content | Claude Help Center

For vendors building on Claude: Article 50 obligations may attach at both provider and deployer layers. Anthropic advises builders to independently assess what Article 50 requires of their products and says further technical guidance will follow. How Claude marks AI-generated content | Claude Help Center