Anthropic explains how Claude’s invisible text watermarks will work

Anthropic has revealed that Claude will use Google DeepMind's open-source SynthID-Text technology to embed invisible watermarks in AI-generated text, helping the company satisfy EU AI Act requirements mandating machine-readable identification markers on synthetic content.

Anthropic has detailed its plan to embed invisible watermarks into text produced by its Claude AI models, adopting Google DeepMind's open-source SynthID-Text framework as the underlying mechanism. The technology works by subtly manipulating word-choice probabilities during text generation, creating detectable statistical patterns that identify content as AI-produced without visibly altering the output. Alongside this text-watermarking initiative, Anthropic is also adding C2PA metadata support for images that Claude processes. Both measures are direct responses to the European Union's AI Act, which mandates that synthetic audio, images, video, and text must carry machine-readable provenance markers. This move signals how major AI developers are beginning to operationalize compliance with Europe's landmark AI legislation, which is progressively rolling out enforcement obligations across different risk categories.

Anthropic has pulled back the curtain on exactly how it intends to mark Claude's AI-generated text as machine-identifiable, confirming it will adapt Google DeepMind's SynthID-Text system — an open-source framework that encodes hidden signals by subtly influencing the probability distributions used when the model selects words during generation. The result is text that reads normally to humans but carries a statistically detectable fingerprint that automated tools can identify as synthetic. Simultaneously, Anthropic is integrating C2PA (Coalition for Content Provenance and Authenticity) metadata standards into Claude's image-handling pipeline, creating a parallel layer of provenance tracking for visual content. Both steps are being taken to satisfy the EU AI Act, which requires all synthetic media — spanning text, audio, images, and video — to include machine-readable disclosure markers. Why it matters: This development is significant on several fronts. First, it represents one of the clearest examples yet of a frontier AI company translating regulatory obligations into concrete, technical product decisions. The EU AI Act has been criticized by some as vague; Anthropic's disclosure shows companies are now engineering specific solutions rather than waiting for further regulatory guidance. Second, the choice to build on Google DeepMind's open-source SynthID-Text rather than develop a proprietary system is notable — it suggests an emerging industry consensus around shared watermarking infrastructure, which could make cross-platform detection more reliable and interoperable. Third, watermarking AI-generated text has historically been controversial among researchers, with concerns about robustness: determined actors can strip or distort watermarks. Anthropic's adoption of a probabilistic, statistical approach rather than a simple tag-based method is designed to be harder to remove, though no system is foolproof. As regulators in other jurisdictions, including the United States, watch the EU's implementation closely, Anthropic's technical blueprint could become a reference model for how the broader AI industry handles content provenance and transparency mandates going forward.