OpenAI Launches Text Watermarking, with Su Weijie as Co-Author

OpenAI announced the official launch of text watermarking. From the day of the announcement, global API customers can enable watermarking themselves for some models, while it remains off by default; within the coming weeks, eligible ChatGPT and Codex text outputs in the EU will be given watermarks invisible to the naked eye, covering all paid tiers. The detector, meanwhile, is open only to vetted researchers and professional institutions on application, and is not yet available to the public.

The direct driver of this launch is regulation. Article 50 of the EU AI Act applies from August 2, 2026, requiring generative AI providers to mark output content in a machine-readable way; the European Commission's June-published Code of Practice on Transparency of AI-Generated Content further lists "imperceptible watermarks" as a basic requirement, and points out that free text cannot carry metadata the way image files can, making watermarking almost the only feasible marking method. In other words, for text this is not one option among many, but one of the few feasible paths framed by regulation.

The engineering dilemma of text watermarking lies in measurement. An LLM generates text word by word, sampling at each step from a "next-word prediction distribution"; the watermarking approach uses a key and the preceding context to generate pseudorandom numbers, making the sampling process depend on this string of numbers, and at detection time the same key is used to reconstruct the numbers and test the statistical association. The property previously pursued by academia is called "unbiased": when the key is regarded as randomly drawn, averaging over all keys makes the watermark model's output distribution exactly identical to the original model's.

But in actual deployment the key is fixed. Taking the Gumbel-max scheme as an example, under a fixed context and fixed key it always selects the same word, so repeatedly generating from the same prompt yields a word-for-word identical answer—a real loss for applications that need multiple samples and then select the best. The other extreme is equally undesirable: if the generation process has no connection to the key at all, randomness is intact, and the watermark signal does not exist. What textGrain wants to answer is whether one can use an interpretable knob to precisely control the position between these two extremes.