Anthropic details the labeling of texts generated by Claude and promises a detector accessible to everyone

Since August 2, the AI Act’s transparency obligations have required AI providers to make their content identifiable. Fines can reach €15 million or 3% of global revenue. Anthropic took the strong approach: an invisible watermark embedded in every text generated by its Claude models since that date, worldwide. The grumbling did not take long to surface, with tracking paranoia and concerns about quality.The company eventually published, on August 14, a post that methodically answers every question. Including the most awkward ones.

Your translations are watermarked, your copyedits are not

The chosen method comes from a direct competitor: SynthID-Text, developed by Google DeepMind, validated in Nature at the end of 2024, and building on an idea from researcher Scott Aaronson dating back to 2022. The principle operates at the very level of how the text is produced. When Claude chooses each word, it draws from several plausible candidates, and that choice involves an element of randomness. The watermark replaces this randomness with a sequence derived from a secret key. Anthropic offers an analogy of its own: a game of Monopoly in which the dice are replaced by the decimals of pi. The moves remain unpredictable to the players, but anyone who knows the trick can afterward verify that the game bears its signature.

The mapping of uses, detailed in the post, reveals a few surprises. A translation comes out watermarked, with every word still selected by the model. A light copyedit of your own prose leaves few or no traces, because too few words have been rewritten. Code is difficult to watermark, since instructions rarely leave several plausible candidates (after “Principia,” it is difficult to write anything other than Mathematica, an in-house example). Detection also remains collective: the mark indicates that a Claude model was used, without ever identifying an account or an individual. As for the global rollout, Anthropic embraces it with disarming frankness, because it has no reliable way of restricting the system to Europe.

The graveyard of AI text detectors is already well populated

The question of reliability deserves to be raised all the more because the precedents argue against optimism. OpenAI had launched its own detector in January 2023 before shutting it down six months later. The tool correctly identified 26% of AI-generated texts while falsely accusing humans in 9% of cases. GPTZero and Pangram, the sector’s two commercial benchmarks, track stylistic tells in AI models. Both regularly fend off criticism through cross-examinations of their false positives, a scourge that primarily affects people writing English as a second language. Anthropic also cites, among the tells of its own model, the construction “it’s not X, it’s Y” (self-awareness be praised).

Watermarking operates in a different technical category. Whereas detectors guess after the fact, the mark is applied at the source and can be verified using a key held only by the company. DeepMind ran SynthID on nearly 20 million Gemini responses without any measurable decline in satisfaction. The figure comes from the method’s designer, a caveat that applies (Anthropic is likewise both judge and party in its own blind tests). Above all, the ticket itself acknowledges the blind spots: a short text, a heavy rewrite, or a mixture with human prose dilutes the signal, and the absence of a mark proves absolutely nothing.

A detection API is promised soon, enabling the watermark to be tested firsthand.By then, the prose you entrust to Claude will come back signed without your knowledge, and history will tell whether this brand will go the distance better than the detector buried by OpenAI. See you in six months, the average lifespan of the kind.

👉🏻 Follow tech news in real time: add 01net to your sources on Google, and subscribe to our WhatsApp channel.

Source: Anthropic