Anthropic extends Claude's text watermarking, with a flaw on some models

Anthropic deploys text watermarking on several Claude models to meet European AI regulation requirements. However, on two older models, a simple generation setting adjustment bypasses it.

Anthropic extends its text watermarking to several Claude models, with deployment announced for November 3 and 16, 2026 depending on versions, to help meet the European AI Act deadline in December. The system is designed to identify text generated by the model without adding characters or changing response formats. One drawback remains: on Claude Opus 4.6 and Claude Sonnet 4.6, adjusting a single generation parameter is enough to disable it.

Un marquage discret, mais pas invisible aux développeurs

Anthropic describes this watermark as an “imperceptible pattern in word choice.” The idea is simple: the text remains readable, but its statistical footprint can be detected by a verification API in private preview. The company also specifies that the watermark adds no tokens or characters, does not affect pricing, and barely slows generation. On paper, everything looks very clean. In practice, this kind of neatness depends heavily on knob settings.

The watermark applies at the model level, not the application. This means it follows Claude on AWS, Google Cloud, and Microsoft Foundry, and a client cannot disable it. Anthropic uses a variant of Google DeepMind’s SynthID text, a method that slightly modifies word choice probabilities to produce a detectable signature. OpenAI launched textGrain in October 2026, but with default activation reserved for ChatGPT and Codex in the European Union; the API remains opt-in, according to the source.

Pourquoi certains anciens Claude restent contournables

The most sensitive point concerns Claude Opus 4.6 and Claude Sonnet 4.6. According to the email sent to API clients, the watermark activates only if the three sampling parameters — temperature, top_p, and top_k — remain at their default values. These settings control randomness in word choice. Changing any one of them is enough to avoid watermarking. Anthropic explicitly warns: “If you rely on the watermark for your own compliance, keep these settings at their defaults.”

For Claude Mythos 5 and Claude Opus 4.7, the email does not mention this caveat [to be verified]. The source infers that the watermark is applied regardless of sampling values. Anthropic has not publicly explained this difference, nor the delayed deployment of Claude Opus 4.6, finally postponed to November 16, 2026. For compliance teams, the nuance matters: a marker that disappears with the slightest setting change does not carry the same weight as mandatory marking without escape.

Un calendrier dicté par Bruxelles, pas par le marketing

Anthropic presents this deployment as a response to the AI Act timeline, whose December 2026 deadline pushes providers to more clearly trace model usage. Models released after August 2, 2026, already included watermarking at launch; older ones receive updates in waves. The choice is strategic: traceability becomes an infrastructure function, not just a communication argument. For clients, this changes the game whenever documentary, legal, or regulatory use requires proving text origin.

This logic is not isolated. Google opened SynthID for text and integrated it into Gemini. OpenAI chose a more flexible approach with textGrain. Anthropic locks marking at the model level. A fundamental debate remains, already visible among commentators like John Gruber and the specialized media Artificial Lawyer: a watermark can suffice to signal provenance, but it can also pose problems in sensitive uses where word precision matters as much as origin. AI text is never completely innocent, and sometimes a top_p setting is enough to remind us.

Key points

  • Deployment announced for November 3 and 16, 2026 depending on models.
  • Claude Opus 4.6 and Sonnet 4.6 remain bypassable by settings.
  • Anthropic targets the December 2026 AI Act deadline.
  • Marking applies on AWS, Google Cloud, and Microsoft Foundry.
  • OpenAI and Google offer competing text marking approaches.

By the numbers

  • 2 models — Claude Opus 4.6 and Claude Sonnet 4.6, affected by bypass, source Anthropic.
  • 4 models — new watermarking wave announced late October 2026, source Anthropic.
  • August 2, 2026 — date from which new Claude models were watermarked at launch.
  • November 16, 2026 — new announced date for Claude Opus 4.6.
  • December 2026 — deadline mentioned by Anthropic for the European AI Act.

To read