Anthropic expands Claude text watermarking for EU compliance

Anthropic is expanding its text watermarking to four additional Claude models, according to a report by The-decoder. This rollout includes Claude Sonnet 4.6, Claude Opus 4.6, Claude Opus 4.7, and Claude Mythos 5. The company is implementing these changes to meet the December deadline set by the EU AI Act. While newer models already feature this technology, the update for older versions is being released in stages. Users on Amazon Bedrock may experience a delay of several days before the watermark is active there.

Why it matters

For businesses integrating AI into content workflows, this update introduces a technical variable in how you manage compliance and output quality. The presence of a loophole in older models means that your detection methods must account for specific sampling parameters like temperature or top_p. If you rely on these watermarks to prove human-like text or meet regulatory standards, you cannot assume the mark is permanent across all model versions.

• Technical settings can inadvertently disable watermarking on specific Claude models.
• Regulatory compliance becomes harder to track if different models in your stack behave inconsistently.
• The precision of your text may change if watermarking affects word selection.

What to watch

You should audit your current API configurations if you use Claude Opus 4.6 or Sonnet 4.6 for regulated content. Because minor changes to randomness settings can bypass the watermark, your compliance strategy depends entirely on keeping sampling parameters at their default values. Monitor how these invisible patterns affect the semantic precision of your automated marketing or technical documentation.

Within a year, manual audits will become obsolete. This requires regulators to mandate standardised, machine-readable watermarking protocols that operate independently of specific provider API settings.