Anthropic has announced Claude text watermarking—and plans to offer a detection API.

The API may be the more consequential development.

Under Article 50 of the EU AI Act, providers of generative AI systems must make covered AI-generated outputs machine-readable and detectable, subject to exceptions.

But the Act does not impose a general duty on social media platforms, universities, employers, clients or publishers to scan every text or penalise AI use—although deployers must disclose certain AI-generated public-interest text.

If detectors become widely available, however, a compliance mechanism could become a gatekeeping tool.

Platforms and institutions might use detection readings to label, downrank, reject or investigate content—creating a ripple effect beyond the Act’s direct requirements.

If detection carries consequences, users—including employees, students, freelancers and social media users—may:

• Rewrite or paraphrase AI-generated text
• Mix human-written and AI-generated sections
• Switch to unwatermarked models, including some open-source models
• Use another model or a “watermark removal” service

Likely outcomes:

• More concealment—not necessarily less AI use
• Growth in anti-detection tools and services
• Disputed readings and appeals
• Private enforcement policies the Act does not require
• Compliant, watermarked content being disadvantaged against older unmarked content or content from harder-to-regulate systems

A positive reading may indicate that a particular watermarked AI system was involved, but it does not by itself prove authorship or misconduct.

And no watermark does not mean human-written.

The critical question is not simply whether watermarking works. It is who gets access to the detectors—and what they are allowed to do with the results.