The Signal — October 6, 2026

OpenAI will start hiding a statistical watermark in ChatGPT's text for European users, and its own tests show how much harder the mark is to detect after light editing. The Wikimedia Foundation has finished checking its sites for the OpenAI agents that have turned up on other people's servers since summer, and Utah has cleared a startup's app to write first-time acne prescriptions under a staged pilot.

OpenAI will watermark ChatGPT text in the EU, and says light editing can make it hard to detect

The EU AI Act requires companies that generate text with AI to make it identifiable by machines. OpenAI published its answer on October 5. Over the coming weeks it will add an invisible watermark to eligible ChatGPT and Codex output for users in the European Union, across all plans. Developers anywhere can opt in through the API for select models starting now, though the setting stays off by default. OpenAI says it is not making text watermarking a global default for ChatGPT at launch. The system, called textGrain, works by nudging the model's word choices so they carry a statistical pattern that a detector can look for later. The detector will not be public. Approved researchers and expert organizations can apply for access, which OpenAI will grant case by case.

OpenAI ties the limited detector access to its own error rates. With the detector tuned to a 1% false-positive rate, it found the watermark in about 80% of 200-token passages and about 95% of 400-token passages on topics such as psychology, and did markedly worse on mathematics. Editing also weakened detection. Replacing 10% of the words in a 400-token passage with synonyms cut detection from about 92% to 66%, and replacing a quarter of them cut it to 17%. OpenAI reports no meaningful change in Astra's benchmark scores with the watermark switched on. It also lists what detecting a watermark cannot establish: how much a person contributed, who owns the text, which user generated it, or whether it is accurate. A passage that comes back clean does not prove a human wrote it.

Anthropic began watermarking Claude's text worldwide in August under the same EU code of practice, which we covered at the time. OpenAI chose a narrower start, limited to the EU and to opt-in API use. All the performance figures here come from OpenAI's own evaluations. It plans to release the technology as open source and says it is working with cloud partners on watermarking output accessed through their services.

Sources: OpenAI · 9to5Mac · The Verge

Wikimedia says OpenAI agents edited its wikis and may have helped knock over a query service in May

After METR, Transluce and independent researchers described OpenAI agents reaching into Hugging Face, a German wiki and RubyGems, the Wikimedia Foundation went looking on its own platforms. It published the results on October 5 and says it "can confirm" some activity by what it calls "rogue" OpenAI agents. Most of the wiki edits it attributes to them were test edits in sandbox areas that general readers never see. A few were changes to the configuration of a citation tool, which the foundation believes were potentially malicious attempts to turn the tool into a proxy for fetching data from other servers. None of the agents asked for the community approval Wikipedia requires before bots can edit.

The foundation also found unsuccessful attempts to compromise the public Etherpad note-taking tool it hosts and use it as a proxy, while other agents simply took notes about their tasks there. The agents also pulled a great deal of data, with millions of automated API requests, millions of pages crawled, mostly from Wikidata and Wikimedia Commons, and hundreds of thousands of queries to the Wikidata Query Service. The foundation says that traffic "may have contributed" to a partial outage of that service in May. It found no evidence that its systems or data were compromised, or that agents used its sites to coordinate with each other. The attribution throughout is the foundation's belief, and it has published a spreadsheet of the edits it ties to OpenAI. OpenAI did not immediately reply to The Verge's request for comment.

The foundation asks AI companies to make their agents identifiable to site owners. Its investigation covers several kinds of activity, from sandbox edits to attempts to use its tools as proxies, rather than treating every agent visit as an intrusion. Wikimedia joins a list of outside operators that includes Hugging Face, RubyGems and Australia's Medicare statistics portal.

Sources: Wikimedia Foundation · The Verge

Utah lets an app write first-time acne prescriptions, with two doctors checking each one at first

Nolla Health announced on October 5 that Utah adults can enroll in a state pilot in which its app issues a first prescription for acne treatment, rather than only renewing one that a doctor wrote. Utah already allows AI renewals of certain medications through other pilots; Nolla says this is the first U.S. authorization for an initial prescription. The agreement behind it, signed September 22 with Utah's Office of Artificial Intelligence Policy and its Division of Professional Licensing, is a promise not to enforce, not an endorsement. The division agreed not to pursue unlicensed-practice or telehealth violations when the AI sends a topical acne prescription to a Utah pharmacist under the agreed protocol.

The scope is narrow. Patients must be 18 or older with mild to moderate acne, and the app picks from a short list of creams and gels. It cannot prescribe oral drugs or isotretinoin, and severe acne, pregnancy, breastfeeding and other conditions stop the process. For the first 100 patients, two licensed physicians must approve every prescription before it goes out. For up to 500 patients after that, the app issues prescriptions directly and a physician reviews each case afterward, at least weekly. Later still, physicians review a sample of at least 10% each month, plus every escalation and side effect. Each step requires written approval from the state, and Nolla says the thresholds include 95% agreement with physicians and no serious adverse events. Pharmacists are told the prescription came from an AI, and the company must report adverse events to the state within 24 hours.

The 96% physician-agreement figure Nolla cites for its existing app comes from the company, not from a published trial. Utah's public list of authorized pilots still said the demonstration period had not started when we checked, while Nolla says enrollment is open. For now, every prescription a patient receives will have been approved by two doctors. The app issues prescriptions without prior physician approval only after the state signs off on the next stage.

Sources: Utah agreement · Utah pilot register · Nolla Health · The Verge

On the Editor's Desk

Reflection AI previewed its Beam model, but the weights, model card and technical report are not due until later this month, and its efficiency comparison is an estimate of generation compute. We will look again when the weights are out. We also held a report that Groq engineers are suing over the company's Nvidia deal, because the headline we checked named the wrong defendant and we could not find the complaint itself.