OpenAI introduces textGrain watermarking for EU ChatGPT outputs to meet AI Act rules

Deployment of textGrain in the European Union

OpenAI announced on 5 October 2026 that it will begin embedding invisible watermarks in text and code generated by ChatGPT and Codex across the European Union. The proprietary technique, titled textGrain, introduces a subtle statistical pattern into the model next-word selection process during text generation. Because these markers exist within the chosen vocabulary rather than in external metadata, the watermark stays intact when a user copies and pastes the output into other applications. The rollout covers all subscription tiers for eligible EU users over the coming weeks, while remaining disabled as a default setting globally. Developers accessing OpenAI application programming interfaces outside Europe can manually enable watermarking on select models starting 5 October 2026. Internal benchmark tests cited by the company showed no measurable decline in output quality, linguistic style, or model generation speed.

Compliance with Article 50 of the EU AI Act

The implementation aligns OpenAI with Article 50 of the European Union AI Act, which requires generative AI systems to mark artificial outputs in a machine-readable format. Transparency obligations under the law took effect on 2 August 2026 for new artificial intelligence offerings, whereas established providers have until 2 December 2026 to implement compliant identification mechanisms. This deadline applies across industry operators including OpenAI, Anthropic, Microsoft, Google, and Meta. Anthropic began applying watermarks to Claude outputs globally in August 2026 using Google DeepMind SynthID technology. OpenAI developed its textGrain mechanism in partnership with academic researchers from Yale University and the University of Pennsylvania, relying on a cryptographic key to arrange candidate tokens during sentence completion.

In response to EU regulatory requirements, we are expanding our approach to content provenance to include text.

Detection thresholds and technical limitations

The textGrain detection software identifies artificial origin by aggregating hundreds of subtle vocabulary adjustments alongside the private key. Detection achieves practical reliability on passages containing at least 400 tokens, representing roughly 200 to 300 words. Despite matching or exceeding SynthID benchmarks in controlled environments, the company noted that real-world effectiveness declines under specific conditions. Shorter responses register an average detection rate of approximately 80%, while mathematical computations, translations, and heavily restructured texts present detection challenges. Manual text alteration significantly reduces system accuracy: replacing 10% of words with synonyms drops detection rates from 92% to 66%, and replacing 25% of words reduces accuracy to 17%.

Verification access and data privacy

OpenAI is restricting initial access to its detection software to approved research groups and expert organizations rather than publishing a general public detector. The company clarified that the presence of a watermark does not measure human creative input, nor does the absence of a watermark verify human authorship. The system does not record individual user identities, account details, prompts, or conversation histories during the watermarking process. OpenAI also stated intentions to release textGrain as open-source software to facilitate independent evaluation and adaptation across the technology sector.

Strong performance under ideal conditions does not guarantee reliable detection in everyday use.