TLDR
OpenAI said on 5 October it will add textGrain, an invisible statistical watermark, to ChatGPT and Codex output for eligible EU users over the coming weeks, to meet Article 50 of the EU AI Act. Editing just a quarter of words drops detection from 95% to 17%, limiting its use in professionally revised copy. Australian users aren't covered, though API customers anywhere can opt in.
KEY TAKEAWAYS
OpenAI will hide a statistical watermark in text written by ChatGPT and Codex for users in the European Union, but its own tests show the mark fades fast once someone edits the copy.[1]
The company said on 5 October that the watermark, called textGrain, would be added to eligible output "over the coming weeks" for EU users on all plans. It is meant to meet Article 50 of the EU AI Act, whose breaches can draw fines of up to €15 million or 3% of global turnover.[1][2]
"Our text watermarking technology, textGrain, adds an invisible statistical signal to the model's word choices," OpenAI said.[1]
A quarter of the words, and it's mostly gone
At a 1% target false-positive rate, the detector found the mark in about 80% of 200-token passages and about 95% of 400-token passages.[1]
Editing changes that. In 400-token passages, swapping 10% of words for synonyms cut detection from about 92% to 66%, OpenAI said. Replacing a quarter of the words cut it to 17%.[1]
OpenAI was blunt about the limits. "A watermark does not measure human contribution," it said, and "the absence of a detected watermark does not prove human authorship."[1]
Nobody outside a small group can run the check yet. Access to the detector "will initially be limited to approved researchers and expert organizations", and OpenAI said it plans to release the technology as open source later. A technical report sets out how it works.[1][3]
textGrain detection rate on 400-token passages
Optional outside Europe
Australian ChatGPT users aren't covered by the rollout. API customers anywhere can opt in to watermarking for selected models from 5 October, but it stays off by default.[1]
That leaves Australian agencies that write for European clients, or build products on the API, with a choice to make about the setting. For anyone else, OpenAI's figures cut both ways: a clean detector result proves nothing about who wrote a piece, and a lightly edited AI draft will often pass as unmarked.[1]
SOURCES & CITATIONS
FREQUENTLY ASKED QUESTIONS
What is textGrain and why has OpenAI released it?
How accurate is the textGrain detector?
Can Australian agencies or brands access the detector?
Xaviery Malinao writes for Prompt the Market on how brands and agencies are adapting to answer engines, drawing on Bushnote's work with clients across search, AI search and content.
