Claudeのテキスト透かし機能の仕組みHow Claude’s text watermark works

Anthropic has indicated that future versions of Claude will produce text containing a watermark, a signal intended to help determine the likelihood that the model was involved in writing a given passage. The change matters because it represents one of the more visible attempts to make AI-generated text detectable at the point of creation, and because Anthropic frames it not as a standalone product feature but as part of a coordinated response to European regulation.

According to the company, the move is being made to comply with the EU AI Act, and Anthropic is not acting alone: it states that several other major AI providers are implementing similar watermarking changes. That framing is significant. Reliable text watermarking has historically been difficult to deploy, so a shared industry approach tied to a common regulatory obligation suggests the shift is being treated as a compliance baseline rather than a competitive differentiator.

The technical challenge is worth unpacking. Watermarking images or audio is comparatively straightforward because those formats contain large amounts of redundant data that can be altered slightly without a human noticing. Text is far sparser, which makes hidden signals harder to embed and harder to preserve. A common technique, seen in systems such as Google DeepMind's SynthID-Text, operates at the level of token selection. As a language model generates output, it repeatedly chooses among many statistically plausible next words, and a watermarking scheme can subtly bias those choices according to a secret pattern. A detector that knows the pattern can then measure whether a passage's word choices deviate from random chance in the expected direction, yielding a probability that the text was machine-generated. Anthropic has not published the full details of its method here, but its description of the feature as estimating a likelihood is consistent with this kind of probabilistic, statistical approach.

That probabilistic nature also explains the feature's inherent limits. Because a watermark lives in patterns across many words, short passages carry less signal and are harder to classify confidently. Heavy editing, paraphrasing, translation into another language, or blending AI-written and human-written text can all weaken or erase the mark. Detection therefore appears likely to produce estimates rather than definitive verdicts, and any real-world system must balance false positives against false negatives. These constraints are widely acknowledged across the research community and are not unique to Claude.

The regulatory backdrop helps explain the timing. The EU AI Act includes transparency provisions requiring providers of generative AI systems to ensure that synthetic output is marked in a machine-readable format so it can be identified as artificially generated. Watermarking is one mechanism that can satisfy this expectation, alongside metadata and other labeling techniques. Because the obligation applies broadly to providers offering services in the European market, it creates a strong incentive for companies to adopt comparable measures on similar timelines, which aligns with Anthropic's statement that multiple providers are moving together.

It is useful to place this within the wider provenance and transparency ecosystem. Watermarking is one of several complementary approaches to content authentication. The Coalition for Content Provenance and Authenticity, or C2PA, promotes a metadata standard often surfaced as "Content Credentials," which attaches a tamper-evident record describing how a file was created or edited. Image and video watermarking efforts, including SynthID for images, address synthetic media in other formats. No single method is sufficient on its own; metadata can be stripped, and watermarks can be degraded, so many observers view them as layered defenses rather than guarantees.

For users and organizations, the practical implications will depend on how detection is made available and how accurate it proves in everyday conditions. A watermark is only useful if a corresponding detector exists and is accessible to the parties who need it, whether platforms, educators, or publishers. As the feature rolls out in future Claude models, key open questions include how robust the signal is to ordinary editing, whether detection tooling will be offered publicly or restricted, and how it interoperates with other providers' schemes. For now, the announcement signals a clear direction: text-level provenance is becoming a standard expectation for major AI systems operating under the EU AI Act.