Anthropic used its blog to expand on a policy it announced earlier in the week: Claude will start embedding invisible watermarks into text it generates, a move required to comply with the EU AI Act's Transparency Code, which obligates AI providers to make AI-generated content detectable. The mechanics matter for anyone building on top of Claude or evaluating AI-generated content in their own pipelines. Rather than appending a visible tag, the system works by nudging Claude's word choices in situations where multiple options are roughly equally good, for instance whether to write "overcast" or "grey" when describing weather. Across enough of these low-stakes token-level decisions, a statistical pattern emerges that is invisible to a human reader but recoverable by anyone holding the cryptographic key, using an approach modeled on Google DeepMind's SynthID-Text scheme from 2024. Anthropic says it will ship a public detection API so third parties can verify watermarked text themselves, and that the technique is distinct from stylistic AI-detection tools like Pangram, which infer AI authorship from writing patterns rather than a cryptographic signal. Two details are especially relevant to engineering teams: light edits to AI-written text likely will not strip the watermark, but a full rewrite will, since there needs to be model-chosen tokens remaining for the signal to attach to; and code will carry a much weaker or absent watermark than prose, because syntax leaves little room for arbitrary word-level choices, though comments could still carry a trace. For teams building content-moderation, plagiarism-detection, or compliance tooling around LLM output, this clarifies both the reliability and the limits of watermark-based detection as a control, and signals that other major model providers who signed the same EU Code of Practice will roll out comparable schemes, so watermark handling is likely to become a shared consideration across the LLM ecosystem rather than an Anthropic-only quirk.