Anthropic confirmed on August 11, 2026 that Claude models released from August 2 onward will embed an invisible watermark in every piece of generated text, using a system derived from SynthID-Text developed by Google DeepMind. The announcement was not entirely new — the initial reveal came a few days earlier — but what emerged in the last hours, in the form of a detailed technical blog post from Anthropic itself, reveals the inner workings of a mechanism that will practically reshape how we read, verify, and trust text produced by artificial intelligence.
The move is framed as compliance with the EU AI Act, which enters a stricter enforcement phase and mandates, under Article 50, that all AI systems must mark their textual outputs in machine-readable format, detectable as artificially generated or manipulated. The EU Code of Practice on Transparency of AI-Generated Content supplements this obligation, creating a regulatory landscape without precedent. Anthropic, along with several other major AI providers, chose not to design its own system but to adopt the SynthID-Text approach, published by Google DeepMind in the journal Nature in 2024.
What SynthID-Text does is elegant in its simplicity. Rather than inserting hidden characters or visible metadata into the text — the kind of watermarking we have seen in images — it operates directly on the text generation process by manipulating the randomness source the model uses when selecting tokens. When an LLM generates text, it rarely takes the most obvious deterministic decision at each step. Even with low temperature, there is a probability distribution across the vocabulary, and the model samples which token to follow. Watermarking works by partitioning this token space into two groups — a "green" and a "red" one — in a deterministic manner using a secret key. Whenever the model's probability distribution allows it to choose a token from one group over the other, it is slightly nudged toward favoring the green group. The result is a subtle, statistical bias that does not alter the text's fluency, speed, cost, or perceptible quality. But in aggregate, it creates a detectable signature.
Anthropic confirmed that this watermark operates "at the model level," meaning that texts generated via the API, AWS, Google Cloud, Microsoft Foundry, and all other supported channels carry the same mark. For detection, the company launched a verification API that allows third-party developers to query whether a given text snippet contains Claude's watermark. The ambition is considerable: integrate it into audit tools, publishing platforms, and content verification systems. But the limits are real.
Watermark-based detection requires sufficient text volume for the statistical bias to manifest with confidence. Short passages, text densely packed with facts and proper names, or passages translated and paraphrased by a third party tend to lose the signal. And there is a particularly awkward paradox: the detection API, by disclosing whether a mark is present or absent, also functions as an "evasion oracle" — anyone can apply rewriting tools that strip the watermark for about four cents per pass, as technical community members have noted. It is akin to affixing an authenticity seal to a product and simultaneously selling the manual for how to remove it.
Even more troubling is the question of ambiguity. Anthropic makes clear that a detected watermark indicates only that Claude may have processed the text. That processing can range from generating an entire article to editing a single paragraph of originally human-written text. A journalist using Claude for copy editing will carry the same mark as a student using it to write a full essay. The watermark, therefore, does not prove authorship; it proves processing. And if you think generic AI text detectors are problematic, know that they do not read the watermark — they are statistical classifiers trained to guess whether text "looks" machine-generated, with false-positive rates that hit hardest at writers with unusual or formal styles.
What to watch in the coming months is the race between detection systems and evasion techniques. Google's SynthID-Text is already embedded in Gemini's text output since 2024, with the detector living inside Google's own apps. OpenAI has also adopted SynthID for image watermarking. The field is becoming an arms race between transparency and circumventability — and until a universal verification standard emerges that does not depend on the goodwill of the content producer itself, trust in online text will continue to be an exercise in healthy skepticism.
Sources: The Verge, The Verge, The Verge, Hacker News
✓ Independent sources cross-checked and verified before publishing
