Claude output is already watermarked, and OpenAI API users have to opt in

The two companies have chosen different coverage and defaults for watermarking under the same rules. On October 5, OpenAI said it would add invisible watermarks to text responses from ChatGPT and Codex for European Union (EU) users within a few weeks. API customers worldwide could enable the feature on some models starting that same day, but it is off by default.

On August 14, Anthropic announced that it would watermark Claude output worldwide. It said it did not yet have a reliable way to apply watermarking by region.

The requirement comes from Article 50 of the EU AI Act. Since August 2, providers of systems that generate AI text have been required to add machine-readable marks to their output. Generative AI systems released before then have until December to comply.

OpenAI's method, textGrain, slightly shifts the probabilities used to choose the next word, leaving a statistical pattern. Anthropic's approach is based on Google DeepMind's SynthID-Text.

OpenAI reports detection rates of about 95% for 400-token texts and about 80% for 200-token texts at a false positive rate of 1%. A false positive means unmarked text is incorrectly identified as watermarked. Tokens are the units a model uses to process text.

In a separate editing experiment, replacing 10% of the words in 400-token texts with synonyms reduced detection from about 92% to 66%. Replacing 25% brought it down to 17%.

Both companies restrict detector access to qualified researchers and institutions. Microsoft Foundry documentation lists only Anthropic models as supporting text watermarking and says customers are responsible for assessing their transparency obligations.

For developers, the difference in defaults stands out. Output from a service built with the Claude API may already be watermarked, while a service using the OpenAI API has no watermark until the feature is enabled. Switching models can change whether a service's text is watermarked.

That makes watermarking defaults worth adding to model comparisons alongside cost and response quality. For pipelines that also translate or copyedit text, whether the watermark survives remains an open question. Teams without detector access have to record confirmed settings separately from whether the watermark actually survives.

Whether text is watermarked and what users must be told are separate questions. OpenAI also says the mark cannot establish who wrote the text, how much a person contributed, or whether the content is accurate.

Sources

1. OpenAI: Our approach to EU text provenance rules, October 5, 2026. The announcement was cross-checked against an OpenAI developer community post.
2. TechCrunch: OpenAI will start watermarking ChatGPT's text in the EU, October 5, 2026.
3. AI Matters: OpenAI to add invisible watermarks to ChatGPT responses in the EU, October 6, 2026.
4. Anthropic: How Claude's text watermarking works, August 14, 2026.
5. European Commission: Quick Facts, Transparency rules for AI systems.
6. Microsoft Learn: Content provenance, updated September 17, 2026.