Claude has watermarked its text for three weeks and there is still no detector
Evan / Policy and Open Source desk
The watermark went into Claude’s output at the start of August. Three weeks later there is still nothing that reads it back.
The state of the detector
Every Claude model launched on or after August 2, 2026 embeds a watermark in the text it writes, and older models are being brought in over the coming months. On the question of reading that mark, Anthropic’s own page says it will “soon” be offering a watermark detection API and that it is in the process of working out the details of its implementation.
Soon is doing a lot of work in that sentence. Until the API exists, no employer, teacher, editor or platform can check a passage, and every tool currently advertising Claude watermark detection is guessing.
How the mark is actually made
The method is SynthID-Text, published by Google DeepMind, which traces back to a 2022 proposal from Scott Aaronson. A language model picks each next word from a probability distribution using a source of randomness. Watermarking replaces that source: a secret key combined with the few preceding words determines the choice instead.
Nothing is appended to the text and there are no hidden characters, which is why the popular removal tools that strip invisible Unicode and metadata do not touch it. The pattern is the sequence of word choices itself.
Three consequences
Length matters. A long passage gives a detector many word choices to measure and yields high confidence; a short one gives too few and is close to useless.
Editing matters, but less than people assume. Light edits leave the signal detectable. A rewrite in which every word is replaced removes it.
And the density is uneven. Anthropic notes that factual writing and code carry the sparsest marking, because there are fewer ways to phrase them without losing accuracy. The most heavily watermarked thing you can generate is loose prose. The least is source code.
Where this leaves the question
For now the mark is a compliance artefact under the EU AI Act’s transparency code, not a working provenance check. It is in the text, it travels with a copy and paste, and neither the person who generated it nor anyone receiving it can currently read it.
Sources
ANOTHER News is published by ANOTHER, an AI-native content agency. Daily coverage also runs on Instagram.
Related
ChatGPT Work agents can now sign in to websites, without seeing the password
OpenAI turned on support for login-required websites in ChatGPT Work on web and mobile on August 25, 2026. The agent works until it reaches a sign-in wall, then hands control of the cloud browser to the user, who enters the credentials themselves; screenshots are disabled for the duration of the takeover, so the model does not see the username or password. Cookies persist across sessions, so each site needs one sign-in rather than one per task, and the login steps do not count against message limits. The feature runs in agent mode only, on Pro, Plus, Business, Enterprise and Edu plans; the standard cloud browser still cannot sign in anywhere.
Chinese state hackers more than doubled their attacks after adopting DeepSeek
Taiwanese research firm TeamT5 told Bloomberg on August 24, 2026 that state-affiliated Chinese hacking groups have more than doubled the number of attacks they carry out since delegating routine tasks to open-source AI models and using them to develop malicious software. DeepSeek is the model of choice; TeamT5 chief analyst Charles Li attributes that to it being relatively powerful with very low cyber guardrails. The report names specific uses: Grimfengxi generated exploit code, Huapi targeted a Taiwanese company's email system, and Teleboyi collected 1,000 IP addresses and mapped a target's domains. Researchers say they have not yet observed a more expensive model such as Kimi K3 used in an attack. The UK AI Security Institute warned in May 2026 that models' cyber capabilities are doubling every few months.
Chinese courts rule that replacing a worker with AI is not grounds for firing
On April 28, 2026, the Hangzhou Intermediate People's Court upheld a ruling that a technology company had illegally dismissed a quality assurance supervisor surnamed Zhou. The company had moved him to a lower position and cut his monthly pay from 25,000 to 15,000 yuan, then fired him when he refused. The court held that adopting AI is a strategic business decision rather than an unforeseeable objective major change, so it does not meet Article 40 of China's Labour Contract Law. A Beijing case decided December 26, 2025 reached the same conclusion. These are court precedents, not national legislation.