OpenAI is rolling out an invisible, machine-readable watermark, called textGrain, in text output from ChatGPT and Codex. The feature is initially limited to users in the European Union. According to OpenAI, the approach “matched or exceeded” other methods such as Google DeepMind’s SynthID for text, which also underpins watermarking that Anthropic announced in August. Both OpenAI and Anthropic cite compliance with the EU’s AI Act as the reason for the change. OpenAI shared AI benchmark scores indicating comparable performance between watermarked and unwatermarked text, while also noting that textGrain “does not guarantee” — a caveat left incomplete in the available text. The addition was not universally welcomed.
Why it matters
Watermarking AI-generated text is tied to regulatory requirements under the EU’s AI Act, and OpenAI’s move follows a similar step by Anthropic, signaling a broader industry response to those rules.
Who should care
EU users of ChatGPT and Codex are directly affected by the initial rollout. Developers and organizations tracking AI content provenance and AI Act compliance may also find the change relevant.