OpenAI announced EU text watermarks without settling who wrote what

OpenAI announced on October 5 that it was introducing invisible watermarks in text generated by ChatGPT and Codex for eligible European Union users, across all plans, over the following weeks. Developers elsewhere could opt in for selected models through its application programming interface, or API. The announcement addressed European transparency requirements rather than establishing a worldwide default. [6][7]

The important distinction is that OpenAI offered a way to trace its tools, not a certificate of who wrote the finished work. Its watermark cannot establish ownership, verify accuracy or measure human contribution. That matters to people copying AI-assisted drafts into documents, developers choosing a setting, and publishers deciding what evidence of AI involvement actually tells them. [6][7]

The rollout began in Europe

The regulatory timeline came first. The EU AI Act’s transparency rules took effect on August 2, requiring AI-generated content to be marked so other systems could identify it. The Next Web reported that providers already on the market had until December 2 to comply. Anthropic began watermarking Claude’s text worldwide in August, with some exemptions, before OpenAI announced its narrower approach. [6][7]

OpenAI’s October 5 announcement separated consumer products from developer services. Eligible EU ChatGPT and Codex users were included in the planned rollout, while API watermarking remained off by default everywhere. Developers could enable it for selected models, and OpenAI said it was working with cloud partners on availability through their services. The company said a regional launch provided room to learn from actual use and feedback. [6][7]

The system, called textGrain, subtly shapes the model’s word choices rather than attaching a visible symbol. Those choices accumulate into a statistical pattern that a detector can recognize using the text and a secret key. Because the pattern resides in the words, it travels when a passage is copied and pasted. OpenAI published a technical report with researchers from the University of Pennsylvania and Yale, reported no meaningful benchmark difference on its Astra model, and said it planned to release the technology as open source. [6][7]

A signal is not authorship

OpenAI drew a boundary around what that pattern means. A detected watermark can indicate that an OpenAI system generated or processed part of a passage. It does not reveal how much human judgment, editing or creativity shaped the result. The detector also does not identify the user or reveal their prompts. A positive result therefore supplies evidence about a production tool, without supplying the information needed to divide credit between that tool and its user. [6][7]

An October 8 report showed why that distinction matters beyond text. Digital Trends described cartoonist Brendan Loper hearing from people who thought he had drawn a circulating cartoon. His pen name appeared on it, but ChatGPT had generated the image. Nieman Lab documented signatures belonging to more than 15 New Yorker contributors on AI-generated cartoons. Emily Flake said images bearing her name blended several artists’ styles, while the signature was recognizably hers. [4]

OpenAI’s public image verification tool could look for signals associated with images generated by its products. It could not confirm that the artist named on a cartoon had made or approved it. Digital Trends advised checking an artist’s portfolio or the publication’s archive before assigning credit. That is the practical difference between identifying a generating system and verifying an attribution: the tool check and the credit check answer different questions. [4]

Users and publishers face different choices

For an eligible EU user pasting a ChatGPT response into a working document, the immediate change concerns what travels with the words. The hidden signal accompanies the passage, but OpenAI said it does not measure the person’s contribution. [6][7] Some Claude users had already objected to Anthropic’s watermarking decision on precisely that ground: they argued that they supplied the instructions, context and decisions, while Claude served as their tool. Their objection concerned credit for the work, rather than simply whether AI had touched it. [7]

Developers had a concrete setting to consider: OpenAI’s API watermarking was optional, not an automatic consequence of the EU consumer rollout. [6][7] Software publishers also faced disclosure requirements separate from detection. On October 8, XDA reported that Flathub had reversed its ban on AI-generated submissions. Submitters instead had to identify affected material and its approximate extent. Humans still had to submit the work; agents could not generate commit messages, descriptions, review comments or replies. Maintainers retained the right to reject disclosed submissions over quality, review or maintainability concerns. For someone submitting an application, disclosure was an obligation, not approval. [3]

Musicians encountered a different version of the problem: what happens when a platform acts and the creator disputes its reasoning? On October 10, The Verge reported that DistroKid had removed recordings in response to claims by UMG, which sued in September alleging copyright infringement and deceptive trade practices. DistroKid disputed the allegations. Musician McGwire said six songs disappeared without notice, including a properly licensed cover, according to his account. He described navigating a chatbot before reaching a person for an explanation. These were copyright claims, not OpenAI watermark detections, but the reporting exposed a separate operational requirement: creators needed explanations and a way to challenge decisions, not merely a classification of their work. [2]

Detection leaves several questions open

How much rewriting can the signal survive? OpenAI’s tests supplied a warning against treating it as permanent. Replacing 10% of a passage’s words with synonyms reduced detection from about 92% to 66%; replacing a quarter reduced it to 17%. The company also identified short passages and translated text as difficult cases, and reported substantially lower detection for mathematics, where word choice is less flexible. A copied passage and a substantially revised passage could therefore present different detection problems, even when both involved the same tool. [6][7]

What does a detector result justify saying? At a target false positive rate of 1%, OpenAI reported detecting its watermark in about 80% of 200-token passages on subjects such as psychology, and about 95% of 400-token passages. Those results allowed both missed detections and some incorrect flags. OpenAI explicitly warned that an absent watermark did not prove human authorship: the text might be short, edited, translated or generated by another company’s tools. The open question for anyone judging writing is what other evidence accompanies the result. [6]

Who gets to check, and what remains unexplained afterward? OpenAI initially restricted detector access to approved researchers and expert organizations, with applications opening on October 5 and decisions made case by case. It linked that restriction to reliability and responsible interpretation, without supplying a broader-access date in the reporting. [6] The cartoon cases left a separate question unresolved: how did contributors’ signatures enter generated images? The New Yorker said it had never licensed its cartoons for AI training, and OpenAI declined to explain how the names appeared. A detector could not resolve that provenance question. [4]

What it adds up to

We read OpenAI’s announcement as an identification measure with deliberately limited claims. The regional rollout, optional API setting and restricted detector access were distinct decisions, not one universal system for judging writing. Its published tests also showed why the limits belong beside the announcement: editing weakened detection, absence proved little, and a positive result did not establish ownership, accuracy or the balance of human and machine effort. [6][7]

The surrounding cases make the distinction concrete. Flathub required disclosure while preserving human responsibility for submissions; cartoonists faced false attribution; DistroKid’s affected musicians sought explanations for removals they disputed. [2][3][4] These were different problems, but each required more than identifying AI involvement. A watermark can contribute evidence about how something was produced. Assigning credit, accepting a submission or removing someone’s work requires a separate judgment—and, as these reports showed, a person able to explain or contest it. [2][3][4][6][7]

Compass found this story, and the sources cited here are the articles it pointed to. Compass is a service we built at Enginerds that reads 140-plus technology publications every day and keeps score of what they cover: which companies, products and technologies are drawing attention, which way each one is heading, and the dated articles behind every number. This piece began as a signal on that scoreboard, a subject the press was suddenly writing about far more than usual, and the research started from the sources Compass had already filed under it. Anyone can look up a company, a product or a technology and read its record the same way. There's a free tier at enginerds.com/compass.