OpenAI Introduces Text Watermarking in EU for ChatGPT

OpenAI is rolling out an invisible, machine-readable text watermarking tool called textGrain for ChatGPT and Codex, with the initial deployment restricted to users within the European Union. According to reporting by The Verge, the release is designed to meet the regulatory requirements established by the EU AI Act.

The company stated that its textGrain approach matched or exceeded competing methods such as Google DeepMind's SynthID, which Anthropic also adopted for text watermarking in August. OpenAI shared benchmark scores indicating that watermarked and unwatermarked text perform similarly. However, the organization cautioned that textGrain does not guarantee reliable detection, verify factual accuracy, determine text ownership, measure human contribution, or prove human authorship.

Alongside the regional rollout across all plans in the EU, OpenAI is introducing an opt-in watermarking option for global API customers to help manage transparency obligations. The company is also collaborating with cloud partners to extend watermarking access for model outputs through their services. Furthermore, approved researchers and expert organizations can apply for restricted detector access to evaluate text provenance on a case-by-case basis, though the detector is not being made publicly available due to risks involving missed watermarks and false positives.

Why It Matters

The regional deployment of textGrain highlights how regulatory frameworks like the European Union AI Act are actively reshaping product launch strategies for major artificial intelligence developers. By targeting the EU first while keeping the feature optional elsewhere, OpenAI can test the technical implementation of machine-readable provenance tools under real-world conditions without forcing a premature global standard. At the same time, the company's explicit warnings regarding detection limitations and false positives suggest that automated watermarking alone may face significant technical hurdles in definitively establishing content origin, leaving open questions about how effectively provenance tools can scale across the broader digital ecosystem.