What textGrain actually does
OpenAI is rolling out a watermarking system called textGrain that embeds a statistical pattern in the word choices of ChatGPT and Codex outputs, detectable by a matching tool that looks for that pattern. The company says its internal testing shows textGrain matching or beating other watermarking methods it evaluated, including Google DeepMind's SynthID for text. The rollout will make watermarking the default for eligible responses across the EU over the coming weeks, framed explicitly as a response to EU AI Act requirements for identifying AI-generated content.
Outside the EU, the feature is not default. API customers worldwide can opt in for selected models, but it stays off unless a developer actively enables it, and OpenAI has given no timetable for a default rollout anywhere else, including the UK. OpenAI says this regional approach "gives us room to learn from real-world use and feedback," without detailing further reasoning. For school and university IT leaders outside the EU, that means the tool most likely to show up in vendor conversations about academic integrity simply will not be switched on by default for their students.
The detection numbers that matter more than the press release
OpenAI's own published figures are the clearest reason institutions should not treat this as an integrity solution. At a 1 percent false positive rate, the system caught about 80 percent of 200-token passages and about 95 percent of 400-token passages in subjects like psychology. Math-heavy text saw substantially lower detection rates, a meaningful gap given how much coursework across STEM fields is quantitative. Those numbers already require fairly long, unedited text samples to hit, which is a narrower slice of real student submissions than it might first appear.
The more damaging number is what happens after light editing. Replacing just 10 percent of words with synonyms cut detection from about 92 percent to 66 percent on 400-token passages. Replacing 25 percent of words dropped detection to 17 percent, which is close to useless as a screening signal. Any student who runs AI-generated text through even a basic paraphrasing pass, something increasingly built into the same chatbots doing the original generation, effectively defeats the watermark. That is not a hypothetical workaround; it is the single most common way students already modify AI output before submitting it.
A detector schools cannot actually use
Even where detection does work, the detector itself is not publicly accessible. OpenAI is restricting access to approved researchers and expert organizations, assessed case by case, citing the risks of both false positives and missed watermarks. That access model rules out the most obvious use case education IT leaders would want: a school or university running its own check against suspected AI-generated submissions. Unless an institution has struck a specific access arrangement with OpenAI, watermark detection is not a tool available to classroom instructors or academic integrity offices at all.
OpenAI has also been explicit, almost to the point of undermining its own feature's utility for schools, about what a watermark does not prove. The company states plainly that "a watermark does not measure human contribution," and that it cannot show how much human judgment or editing went into a passage. A watermark does not identify the user, account, prompt, or conversation, and does not establish ownership, responsibility, or accuracy. Just as important, a missing watermark is not reliable evidence of anything either, since text may be too short, edited, translated, generated by an unsupported model, or predate the feature entirely.
How this compares to Anthropic's approach
Anthropic has taken a different geographic path, applying watermarking globally to its supported Claude models rather than limiting a default rollout to a single region. The company says it does not yet have a reliable way to limit the feature by region, which is effectively the inverse engineering tradeoff OpenAI made: Anthropic chose uniform global coverage over regional precision, while OpenAI chose regulatory-driven regional rollout over global uniformity. Not every older Claude model supports watermarking, so Anthropic's "global" claim has its own gaps.
Neither company has published a side-by-side comparison of detection accuracy, and the article detailing Anthropic's rollout gives no figures on its detector's access model or accuracy under editing, the exact information that would let an institution compare the two approaches on the metric that matters most. For now, schools and universities evaluating either vendor's watermarking claims are working from incomplete and non-comparable data sets, which should temper how much weight either company's integrity messaging gets in procurement conversations.
The practical takeaway for academic integrity offices
The honest reading of this announcement is that watermarking, as currently built by either major lab, is not a workable academic integrity enforcement tool for most institutions. It requires a specific access arrangement schools mostly do not have, it degrades sharply under exactly the kind of light editing students already do, and both companies explicitly disclaim it as proof of authorship or human contribution. Integrity offices that have been waiting for a technical fix to the AI-generated-writing problem should not expect this to be it, regardless of which vendor's product their institution uses.
What the announcement is actually useful for is signaling where the regulatory pressure is coming from. The EU AI Act is now shaping product defaults at OpenAI in a way institutions elsewhere are not benefiting from, a reminder that content-provenance regulation is advancing unevenly across jurisdictions. Education IT leaders tracking AI governance policy should treat this less as an integrity tool rollout and more as an early data point on how regulatory pressure translates into default product behavior, since that pattern is likely to repeat with other vendors and other regions.
