OpenAI Adds 'textGrain' Watermarking for EU Users in Response to AI Act

OpenAI has begun deploying an invisible, machine-readable text watermarking system across ChatGPT and Codex, starting exclusively with users in the European Union.

The feature, named textGrain, is rolling out over the coming weeks across all ChatGPT and Codex plans in the EU, as reported by The Verge. OpenAI confirmed it is holding back from making the technology a global default at launch, citing a plan to gather feedback and evaluate real-world performance before considering broader availability.

The regional launch follows regulatory requirements under the European Union's AI Act. OpenAI joins competitors such as Anthropic, which introduced text watermarking based on Google DeepMind's SynthID in August to meet the same EU standards.

According to OpenAI, textGrain matched or exceeded the performance of alternative approaches such as SynthID. The company published internal benchmark results showing comparable output quality between standard generated text and watermarked text.

Despite those benchmark results, OpenAI noted that textGrain does not guarantee reliable detection. The company stated that text watermarks cannot verify factual accuracy, determine intellectual property ownership, measure human editorial contribution, or prove human authorship.

Outside the EU consumer rollout, OpenAI is enabling global API customers to opt in to watermarked outputs for select models starting immediately. Similar watermarking support is scheduled to reach third-party cloud partner platforms in the coming weeks.

OpenAI also opened an application process for vetted researchers and expert organizations to access its detection tool on a case-by-case basis. The tool identifies whether text contains an OpenAI watermark without exposing user identities, prompts, or conversation histories. The detector will not be released to the general public at launch due to the risk of false positives and missed watermarks.