Anthropic is preparing to add invisible text watermarking to Claude and, according to The Decoder, will soon offer a detection API that lets third-party developers check whether Claude was likely involved in creating a text. BleepingComputer separately reports that Anthropic has detailed how it plans to implement watermarking across Claude as AI providers respond to EU transparency requirements. Both outlets report that Anthropic is introducing the system to comply with the EU AI Act. The Decoder says Anthropic signed the EU Code of Practice on transparency for AI-generated content in July 2026 along with roughly 190 other signatories. BleepingComputer reports that Anthropic and several other major AI providers have agreed to comply with the EU’s Code of Practice, and that Anthropic is among the first major AI providers to share implementation details for how it will mark Claude-generated text. The rollout is not expected to be limited to Europe at launch. Both reports say Anthropic plans to apply the watermark globally because it does not yet have a durable way to scope the feature by region. That makes the EU compliance mechanism a worldwide product behavior, at least initially. The technical approach is based on Google DeepMind’s SynthID-Text method, which was published in Nature in 2024, according to The Decoder. Rather than appending visible markers or modifying text after generation, the system works during generation by changing the source of randomness used in some word or token choices. Anthropic says those repeated low-stakes choices create a statistical pattern that is not visible to a reader but can be detected by someone with the relevant key, BleepingComputer reports. Anthropic says the watermark does not practically affect the quality, content, creativity, or readability of Claude’s output, according to both outlets. The important distinction is that this is not the same as a conventional AI-text detector scanning for stylistic clues. The Decoder notes that outside detection tools do not have Anthropic’s keys and instead look for patterns common in AI-generated prose; watermark detection is a different mechanism because it checks for a generated statistical signature. The limits matter. The Decoder reports that the watermark is less reliable for short passages, fact-heavy text with few alternative phrasings, and code. It also says pure corrections where a human selected every word will not carry the watermark, while translations can because Claude chooses the words; heavy rewriting can remove the watermark. The same report says detection can only indicate that Claude was likely involved in producing a text, not whether Claude wrote the entire document, heavily edited it, or whether the text came from a human or another AI model. There is one material discrepancy in the reports: the model cutoff date. The Decoder says all Claude models released after August 2, 2025 support watermarking out of the box and that older models will receive the feature in coming months. BleepingComputer says future Claude models will generate watermarked text, while models launched before August 2, 2026 are covered by the EU transition period and will receive watermarking over coming months. Because the secondary reports differ, the cutoff should be treated as unsettled in those reports. The Decoder also reports that Anthropic uses the open C2PA standard for files, attaching provenance metadata without altering the file itself. The same report describes the text watermarking mechanism as changing generation-time sampling rather than adding visible labels or metadata to the text output. Who benefits: Third-party developers and platforms that want Claude-specific provenance checks could benefit if Anthropic’s detection API becomes available as reported by The Decoder. Compliance teams also get a clearer mechanism for marking AI-generated text tied to EU transparency rules. Who's exposed: Users and organizations relying on generic AI detectors may need to distinguish those tools from provider-keyed watermark checks. Anyone treating watermark detection as proof of full authorship is exposed to overclaiming, because the reports say it can only indicate likely Claude involvement.