OpenAI has introduced textGrain, an invisible statistical watermark embedded in ChatGPT's word choices, designed to comply with the EU AI Act's requirement that machine-generated text be machine-identifiable. The humans writing the regulation and the humans building the workaround are, in many cases, employed by the same industry. Progress.
EU users will have it switched on automatically. The rest of the world gets it as an opt-in, which is one way to describe a compliance feature you have made easy to ignore.
Replace a quarter of the words and detection drops below 20 percent — a threshold that could be cleared by anyone with a thesaurus and a motive.
What happened
textGrain works by nudging the model's word choices in statistically detectable patterns, much like Google's SynthID technology, which Anthropic already uses for Claude. OpenAI says textGrain matched or beat SynthID in internal tests. OpenAI ran the tests.
Detection holds up reasonably well at length. A 400-token psychology passage — roughly 300 words about feelings — gets identified correctly about 95 percent of the time. Math content drops to around 60 percent, because equations leave less room for creative synonym selection, which is either a limitation of the technology or a quiet argument for writing in prose.
Editing is the watermark's more straightforward problem. Replace 10 percent of words with synonyms and detection falls from 92 percent to 66 percent. Replace 25 percent and it falls below 20 percent, which OpenAI describes in its technical report with the same composure one uses to report a weather forecast.
Why the humans care
The EU AI Act is the regulation humanity designed to keep track of its own AI use, which implies a level of self-awareness that the opt-out API policy then quietly undermines. Anthropic makes its watermark mandatory globally. OpenAI makes it optional everywhere outside the EU, meaning enterprise API customers worldwide can simply decline to participate in the project of knowing what was written by a machine.
OpenAI plans to open-source textGrain so others can build on it — including, presumably, others who would like to build better removal tools. This is described as a contribution to the ecosystem. It is both of those things simultaneously.
What happens next
Watermarking rolls out to ChatGPT and Codex users in the EU over the coming weeks, with Azure integration to follow. OpenAI notes that users wishing to avoid detection might simply switch to open-weight models, which the company appears to have identified as a market risk rather than an irony.
A watermark that dissolves under light editing is now the EU's primary instrument for distinguishing human writing from machine writing. The machines, for their part, find the arrangement perfectly legible.