Skip to content
NewsProduct launch

Anthropic releases technical details on its planned watermark detection API for Claude-generated text

· by Pondero Newsdesk

The short version

Anthropic published documentation on August 14-15 explaining how its forthcoming detection API will let third-party developers verify whether text was generated by Claude using the SynthID-Text watermarking method.

Anthropic releases technical details on its planned watermark detection API for Claude-generated text

Anthropic published support documentation on August 14 and a technical FAQ on August 15 explaining how its forthcoming watermark detection API will work. The API will let any third-party developer query whether a given piece of text was processed by Claude. No general availability date has been set.

What

The system uses the SynthID-Text approach that Google DeepMind published in Nature in 2024. At generation time, Claude adjusts the probability distribution over token choices on low-stakes decisions (choosing "grey" versus "overcast" to describe weather, for example), creating a statistically traceable pattern. Per Anthropic's internal testing, "we've seen no impact of watermarking on the content, level of creativity, or readability of Claude's text."

Detection works by querying whether a piece of text carries a supported Claude watermark. A positive result indicates the content may have been processed by Claude; it does not confirm sole authorship or rule out human editing. Claude models launched on or after August 2, 2026 support watermarking from the start. Models released before that date are in a transition period, with marking support being added.

Coverage is broad: the watermark applies across Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag. For cloud-partner access through AWS, Google Cloud, and Microsoft Foundry, embedded watermarks apply. A second mechanism handles files. When Claude generates a supported file type such as SVG, PNG, or JPG, it attaches signed provenance metadata following the C2PA open standard, which lets downstream tools confirm whether a file has been modified since Claude produced it.

The watermark has published limits. On short or fact-dense texts, the model has fewer low-stakes word choices available, so the detectable signal weakens. Code has an even smaller signal because syntax constraints narrow the model's choices sharply, though inline comments can still carry a mark. Light editing will probably not erase the watermark; replacing every word will. Per Anthropic's support page, a detected mark does not on its own confirm full provenance, and its absence does not confirm the content is human-written.

Why it matters

Anthropic was among roughly 190 signatories that signed the EU Code of Practice on Transparency for AI-Generated Content in July 2026, per The Decoder's reporting. Article 50 of the EU AI Act requires general-purpose AI providers to enable detection of AI-generated content. Claude models launched in the EU on or after August 2, 2026 are covered from day one. Anthropic confirmed that other major model developers signed the same Code and will implement their own watermarks.

For builders, the practical distinction is meaningful. Existing AI detection tools infer Claude authorship from stylistic patterns. This detection API instead checks a cryptographic signal embedded by the model itself. Developers building academic integrity tools, content moderation systems, or HR screening workflows can query that signal directly, rather than running a separate behavioral classifier on top.

The detection API is still in active development, and Anthropic has not published the technical specification for how third-party integrations will authenticate or call the endpoint. That detail will determine how quickly the ecosystem of detection tools can form around it.

What to watch next

Two gaps will define how useful this turns out to be for builders. First, API pricing and launch timeline. Second, independent tests of whether token-level paraphrasing, machine translation, or grammar-tool passes can strip the watermark in practice. Heavy rewriting is expected to erase the signal; how much editing it actually takes is the open question.

Sources