AI Text Watermarking in the EU: What Businesses Should Do

What changes in your tools

On October 5, 2026, OpenAI announced text watermarking, describing it as a response to EU AI Act requirements for machine-readable identification of generated text. This is the provider’s position; the supplied source does not include the legal text.

Availability depends on how employees access AI:

• ChatGPT and Codex: OpenAI plans to watermark eligible EU outputs across all plans over the coming weeks. This is a rollout plan, not confirmed availability for every user.
• Direct API: OpenAI announced opt-in watermarking worldwide for select models accessed through its application programming interface (API). Watermarking is off by default.
• Cloud partners: Availability through their services is still being prepared. Do not assume direct API conditions apply to your cloud setup.

The textGrain technology adds an invisible statistical signal to word choices—not a visible “AI-generated” label. There is no public text detector at launch; initial access is limited to approved researchers and expert organizations.

Three actions for business owners

1. Check tools, not just subscriptions. Have someone list services, models, and access channels. For API-based support or document generation, check whether watermarking can be enabled. Watermarking provides an additional signal of origin. It does not replace quality control.
2. Assign responsibility for publishing. Designate a reviewer to check facts, quotes, and promises in ads, product listings, and customer replies. Watermarking does not verify accuracy or establish ownership. Tell contractors in advance which AI tools are permitted and how to disclose their use.
3. Review legal obligations separately. Before publishing, ask counsel which AI notices and content rules apply to your situation. Do not treat a watermark as proof of legality or remove existing disclosures because invisible watermarking is available.

Where caution matters

Detectors can falsely identify a watermark or miss one. According to OpenAI, replacing 10% of words with synonyms in 400-token passages reduced detection from about 92% to 66%; replacing 25% reduced it to 17%. Tests used English responses from the ELI5 dataset. These figures do not establish reliability for Russian-language documents.

The absence of a detected watermark does not prove human authorship. Short, translated, and edited texts are harder to check. Do not use a detector result as the sole basis for withholding contractor payment or accusing an employee.

Next step: Start with one process, such as publishing product listings. Document the AI tool, reviewer, approval process, and required disclosures. Then extend that process to other tasks.