OpenAI rolls out textGrain watermarking for ChatGPT

OpenAI is introducing an invisible, machine-readable watermarking system called textGrain to its ChatGPT and Codex platforms. The rollout will initially target eligible users across all subscription plans in the European Union over the coming weeks. While the feature is not yet a global default, the regional launch allows OpenAI to gather real-world feedback and comply with the transparency mandates of the European Union's AI Act.

For developers and enterprise users outside the EU, OpenAI is offering an opt-in path. Starting immediately, API customers worldwide can choose to enable textGrain watermarking for select models. This flexibility allows businesses to meet their own transparency standards and regional regulatory obligations. Additionally, OpenAI is partnering with cloud providers to integrate this watermarking capability into external hosting environments in the near future.

According to OpenAI, textGrain performs on par with or better than rival technologies, such as Google DeepMind's SynthID, which Anthropic also adopted in August. Benchmark tests released by the company indicate that watermarked text maintains similar quality and performance compared to unwatermarked outputs. However, the system is not foolproof. OpenAI cautioned that textGrain "does not guarantee reliable detection" and cannot verify factual accuracy, determine legal ownership, or measure human contribution.

To refine the technology, OpenAI is granting approved researchers and expert organizations case-by-case access to a detection tool. This utility will identify whether text contains an OpenAI watermark without exposing user identities or prompt histories. For practitioners, this development signals a shift toward mandatory provenance tracking in regulated markets, forcing developers to weigh the integration of watermarks against potential false positives and detection limitations.