Under the AI Act's text standard of "as far as technically feasible", OpenAI shipped a watermark that survives a copy and not a rewrite. Monday's other change, 50% faster defaults, shows which defaults OpenAI wants.
In 58 days, a European deadline will apply to every provider of a text model that was on the market before August. Regulation (EU) 2026/1744, the Digital Omnibus, says such providers "shall take the necessary steps in order to comply with Article 50(2) by 2 December 2026." OpenAI has outlined its response: a watermark that meets the letter of the law but gives editors, teachers and platforms little they can rely on.
The legal text sets a modest bar. Article 50(2) requires outputs to be marked in a machine-readable form and "detectable as artificially generated or manipulated", with solutions that are "effective, interoperable, robust and reliable as far as this is technically feasible, taking into account the specificities and limitations of various types of content." That final qualification matters: a provider can comply by shipping the strongest marking current science allows, even if it rarely catches text in practice.
OpenAI's post on EU text provenance describes a narrower rollout than the headline suggests. API customers worldwide can opt in to text watermarking for select models, but it is off by default. Over the coming weeks, an invisible watermark will be added to eligible ChatGPT and Codex text output in the EU only. OpenAI says it is "not making text watermarking a global default at launch." The system, called textGrain, adds "an invisible statistical signal to the model's word choices". OpenAI says it matched or exceeded other approaches, including SynthID for text, and plans to open source the method.
Detection rates show the design's limits. At a 1% target false-positive rate, OpenAI reports about 80% detection for 200-token passages and about 95% for 400-token passages, using psychology content. Mathematics scored lower. Benchmarks on its Astra model were within a point or two on most tests with and without the watermark, so the quality cost is small.
Edited text is harder to detect. In 400-token passages, replacing 10% of words with synonyms cut detection from about 92% to 66%; replacing 25% cut it to 17%. Swapping a quarter of the words removes about four fifths of detections. A paraphrasing tool, another model or an editor working through a draft can reach those rates with little effort.
These results make the watermark better suited to compliance than everyday editing. It detects an untouched copy about nine times in ten, but ordinary edits quickly weaken the signal. Most text gets some editing before anyone reads it. Article 50 asks only for technical feasibility, so the watermark can meet the legal standard and deter casual copying. As a tool for editors, teachers or platforms, though, its signal vanishes at normal edit rates. OpenAI also says that the absence of a watermark "does not prove human authorship", leaving the tool able to confirm some copies but unable to clear anything else.
Access to the detector adds another limit. Researchers and expert organisations can apply, but OpenAI is restricting initial access because of the risk of false positives and false negatives. The detector will not identify users. That caution is defensible because a false accusation of machine authorship is costly. It also means that the people who most want to check a text cannot do so. A provenance signal available to only a small group will do little to change what a newsroom, school or marketplace does day to day.
OpenAI handled another default differently. On Monday, Tibo Sottiaux of OpenAI wrote on X that "we have optimized the default speed to be ~50% faster across GPT-6 Astra and GPT-6.1 Sol through the subscription across all our products and partners using Sign in With ChatGPT (including OpenCode, Pi, Amp, Devin, ...)". He added, "No changes needed on your end." The post quotes his earlier promise that over the next 28 days OpenAI would each day "either ship one thing that is a clear improvement ... or ship a full reset." This is Day 1.
The change applies to subscriptions and partner products that use Sign in With ChatGPT; the 50% figure comes from OpenAI. Speed is on for subscribers and developers using third-party coding tools, with no action required. Watermarking is off in the API and on only for EU users of ChatGPT and Codex. The speed improvement spends compute to win the people who choose which model their tools call. The watermark meets the legal minimum.
OpenAI's choices reveal its priorities. If the company wanted text provenance to work as a market feature, it could make the watermark the default for paying developers and open access to the detector. Instead, it has done the opposite on both counts. Speed helps win customers this quarter. A watermark required by law in one jurisdiction, and only as far as technically feasible, offers little commercial return and adds some cost.
Anthropic has taken a broader approach. On August 14, it said it is applying watermarking to Claude text globally at launch and offers a detection API to regulators, researchers, media and obligated enterprises. OpenAI's EU-only design and closed detector offer less. Any statistical signal in text faces the same limits around word choice. After December 2, regulators will have to decide whether a watermark whose detection rate falls to 17% after a 25% synonym swap was the best current science allowed.
