{"schema":"sanctity-signal/0.2","record_type":"signal","id":"openai-eu-text-watermark-e043","domain":"safety","drop_id":"signal-e043","occurred_at":"2026-10-05","how_to_use":"This is a structured record of one real-world action and its competing interpretations, from Sanctity Signal. To build content from it (a reel script, a blog, a thread): open on the hook, ground the viewer with plain and observed, walk the dated trail, present the interpretations as live tension between named readings, and close on the before/after shift and what_to_watch. Hard rules: items under established are fact; items under unknown are never presented as fact; the evidence status (verified or reported) is stated, not dropped; sources are cited.","actor":"OpenAI","action_type":"publication of a text watermarking plan, with watermarks switched on for European Union users of ChatGPT and Codex, global opt-in for API customers, and detector access withheld from the public","observed":"Monday October 5, 2026, the date on OpenAI's own post, titled \"Our approach to EU text provenance rules.\" The method is named textGrain, described as adding \"an invisible statistical signal to the model's word choices\"; a detector then looks for that signal. 3 rollout steps are given. Immediately, API customers anywhere may opt into watermarking for selected models, off by default. In \"coming weeks,\" invisible watermarks are added to ChatGPT and Codex output for users in the European Union only. Starting the same day, researchers and expert organisations can apply for detector access, case by case. The post states plainly: \"We are not making text watermarking a global default at launch.\" Public detector access is denied at launch. The accuracy table is the part worth reading. At a 1% false positive rate, detection reaches roughly 80% on a 200-token passage and roughly 95% on a 400-token passage. Editing degrades it: replacing 10% of the words drops detection from 92% to 66%. Mathematical content is described as \"substantially lower\" than flexible prose, because there are fewer wording choices to hide a signal in. OpenAI also lists what a watermark cannot do: it cannot measure how much a human contributed, establish ownership, identify which user produced the text, verify that the text is accurate, or prove human authorship from the absence of a mark. Not established anywhere in the post: how many models carry textGrain, whether competitors can read the signal, what the EU AI Office has said about whether an invisible statistical mark satisfies a marking duty, how long the watermark survives translation or paraphrase by another model, and whether the EU-only scope is a technical choice or a legal one.","corrections":[],"evidence":{"status":"verified","note":null,"sources":[{"role":"primary","label":"OpenAI: Our approach to EU text provenance rules, Oct 5, 2026","url":"https://openai.com/index/eu-text-provenance/"},{"role":"primary","label":"OpenAI news index showing the post dated Oct 5, 2026","url":"https://openai.com/news/"},{"role":"corrob","label":"Council of the EU: Artificial intelligence, Council gives final green light to simplify and streamline the rules, Jun 29, 2026","url":"https://www.consilium.europa.eu/en/press/press-releases/2026/06/29/artificial-intelligence-council-gives-final-green-light-to-simplify-and-streamline-rules/"},{"role":"corrob","label":"European Commission: regulatory framework for AI, the AI Act transparency obligations","url":"https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai"}]},"strength":{"cost":"Medium","commitment":"High","reversibility":"Moderate","persistence":"Years","rationale":"Cost is medium. Watermarking constrains a model's word choices, which costs some output quality, and publishing your own degradation numbers costs credibility you cannot get back. Commitment is high: this is a named method, a named rollout, a dated opening of detector applications and a public refusal to go global, all on the company's own site. Reversibility is moderate, because a watermark can be switched off as easily as on, and \"coming weeks\" is not a date. Persistence is years: a provenance obligation written into European law does not expire, and once 1 laboratory ships, the others are asked why they have not."},"shift":{"prior":"Text watermarking was treated as impractical, and no laboratory shipped it.","action":"Monday October 5, 2026, the date on OpenAI's own post, titled \"Our approach to EU text provenance rules.\" The method is named textGrain, described as adding \"an invisible statistical signal to the model's word choices\"; a detector then looks for that signal. 3 rollout steps are given. Immediately, API customers anywhere may opt into watermarking for selected models, off by default. In \"coming weeks,\" invisible watermarks are added to ChatGPT and Codex output for users in the European Union only. Starting the same day, researchers and expert organisations can apply for detector access, case by case. The post states plainly: \"We are not making text watermarking a global default at launch.\" Public detector access is denied at launch. The accuracy table is the part worth reading. At a 1% false positive rate, detection reaches roughly 80% on a 200-token passage and roughly 95% on a 400-token passage. Editing degrades it: replacing 10% of the words drops detection from 92% to 66%. Mathematical content is described as \"substantially lower\" than flexible prose, because there are fewer wording choices to hide a signal in. OpenAI also lists what a watermark cannot do: it cannot measure how much a human contributed, establish ownership, identify which user produced the text, verify that the text is accurate, or prove human authorship from the absence of a mark. Not established anywhere in the post: how many models carry textGrain, whether competitors can read the signal, what the EU AI Office has said about whether an invisible statistical mark satisfies a marking duty, how long the watermark survives translation or paraphrase by another model, and whether the EU-only scope is a technical choice or a legal one.","update":"It ships where a law demands it, and nowhere else by default."},"inference":{"established":["OpenAI will add invisible watermarks to ChatGPT and Codex output in the EU.","Detection reaches about 80% accuracy on 200-token passages at 1% false positives.","Detector access is limited to approved researchers, not the general public."],"unknown":["That it survives editing. Replacing a tenth of the words drops detection sharply.","That unwatermarked text is human. The absence of a mark proves nothing.","That this becomes global. OpenAI says it is not the default at launch."]},"narrative":{"hook":"OpenAI will watermark ChatGPT's text, but only for users in the European Union.","plain":"On October 5 OpenAI published its text watermarking plan: an invisible statistical signal in word choices, switched on in the European Union within weeks. Caveat: the detector stays closed to the public, and editing defeats it.","story":["A company that resisted text watermarking for years shipped it, and published the numbers that show how weak it is. It is on in Europe because Europe wrote a rule. That is the shape of AI governance now: the feature follows the jurisdiction, and everyone else gets the default."],"trail":[{"d":"JUN 29","t":"The Council of the European Union gives final approval to the digital omnibus simplifying and streamlining the AI Act, moving several compliance deadlines and keeping the transparency duties on machine-generated content."},{"d":"SEP 21","t":"OpenAI publishes its own design for global AI standards, with no licence attached to it. Edition 029's card. The pattern of publishing a method rather than submitting to one is 2 weeks old."},{"d":"OCT 5","t":"OpenAI publishes the text provenance post: textGrain named, EU-only rollout for ChatGPT and Codex, global API opt-in off by default, detector applications open, accuracy and degradation numbers published, public access refused."},{"d":"Reacts","r":true,"t":"The EU AI Office, which has to decide whether an invisible statistical mark satisfies the marking duty; Google, Anthropic and Meta, now the frontier laboratories without a shipped text watermark; the AI detection industry, whose classifiers OpenAI's own numbers undercut; schools and universities, told the detector is not for them; and publishers who wanted provenance they could check themselves."}],"glossary":[{"term":"Text watermark","def":"An invisible pattern deliberately pushed into a model's word choices, so software holding the key can tell the text came from that model."},{"term":"False positive rate","def":"How often a detector flags human writing as machine-written. At 1%, 1 innocent passage in every 100 is wrongly accused."},{"term":"Token","def":"The unit a model reads and writes, usually a word or part of one. 200 tokens is roughly 150 words of English."},{"term":"Provenance","def":"Where a piece of content came from. Provenance rules require machine-generated material to carry a mark saying so."}]},"judgment":{"prompt":"Which explanation do you believe?","interpretations":[{"id":"a","label":"Regulation now sets product features","text":"The feature exists in 1 jurisdiction because that jurisdiction required it.","abstain":false},{"id":"b","label":"A narrow compliance exercise","text":"A labelling duty met at minimum scope, with the detector kept shut.","abstain":false},{"id":"c","label":"Groundwork for a global default","text":"Ship in 1 market, measure it, then turn it on everywhere else.","abstain":false},{"id":"d","label":"The technology finally matured","text":"Watermarking became good enough to ship, and Europe simply went first.","abstain":false},{"id":"e","label":"Too early to call.","text":"The evidence cannot yet separate these.","abstain":true}],"distribution":{"status":"held back until you have made your call","calls":null}},"story_thread":null,"what_to_watch":[{"linked_interpretation":"a","direction":"strengthens","observable_event":"OpenAI extends text watermarking to another jurisdiction only after a law or regulator requires it, or a second laboratory ships text watermarking EU-first, by June 30, 2027.","window_end":"2027-06-30","state":"unresolved"},{"linked_interpretation":"a","direction":"weakens","observable_event":"OpenAI makes text watermarking a global default, or opens the detector to the public, by December 31, 2027.","window_end":"2027-12-31","state":"unresolved"},{"linked_interpretation":"b","direction":"weakens","observable_event":"The EU AI Office, a national regulator or a court states that textGrain does not satisfy the marking obligation, by December 31, 2027.","window_end":"2027-12-31","state":"unresolved"}]}