Anthropic announced on August 11 that it is now embedding machine-readable, imperceptible watermarks into text and images produced by Claude models, positioning it as compliance with the EU AI Act's transparency requirements — the same Article 50 disclosure rules that came into force across the bloc earlier this month, requiring AI systems to identify themselves to the people interacting with them. The company says it has signed the EU's AI Act transparency code, and that new Claude models will carry the watermark from the moment they're released rather than it being retrofitted later. The technically interesting detail for anyone building on top of Claude is the durability claim: Anthropic says the mark is designed to travel with the content even after a user copies and pastes text elsewhere, which implies the watermarking works at the level of statistical patterns in token or pixel output rather than a metadata tag that gets stripped the instant content leaves the original interface. That's a meaningfully harder problem than the C2PA-style metadata provenance tagging other vendors have leaned on, and if it holds up under adversarial testing, it changes the calculus for anyone building content-authenticity or AI-detection tooling downstream. For developers integrating the Claude API into products that generate user-facing text or images, this is worth flagging to product and legal teams now: content flowing through your pipeline may carry an embedded signal you didn't ask for and can't easily strip, which has implications for use cases like ghostwriting tools, anonymized content generation, or any workflow where output needs to be indistinguishable from human-authored work. It also puts Anthropic ahead of where OpenAI and Google currently stand on watermark durability claims, and is likely to put pressure on both to match the "survives copy-paste" bar rather than relying on metadata alone.