The Signature Only the Signer Can Verify: OpenAI's textGrain and the Gated Detector
On Monday, October 5, OpenAI announced textGrain, an invisible statistical watermark embedded in the word choices of ChatGPT and Codex output — mandatory for eligible EU users across all plans in the coming weeks, opt-in and off by default for API customers everywhere else.1 The move answers Article 50 of the EU AI Act, which requires machine-readable identification of AI-generated text and gives incumbent providers a December 2 compliance deadline.2 The accompanying technical report concedes the numbers that matter: roughly 95% detection under ideal conditions, falling to 66% when a user edits 10% of the words and 17% when they edit a quarter.3 Detector access opens by application, initially limited to approved researchers and expert organizations.1 Anthropic has already been here since August 2 — watermarking Claude worldwide, with no documented opt-out, and taking the user backlash for it.4 And the Wall Street Journal reported back in 2024 that OpenAI had built text watermarking years ago and sat on it, partly because users would defect to rivals that didn’t watermark.2 Four recipes on the ink.
First principles — a signature the signer must authenticate
What is a signature? Strip the ceremony away and a signature is a mark whose evidentiary force rests on public verifiability: anyone holding the public key can check it, and its power in a dispute comes precisely from the fact that the signer cannot monopolize verification. What is textGrain? A keyed statistical perturbation of token selection — detectable only with the vendor’s key and vendor’s detector.5 By the fundamentals, it is not a signature. It is a serial number stamped by the manufacturer, readable only by the manufacturer. The conventional answer — “watermarks let us identify AI text” — survives only by borrowing the prestige of the word signature while discarding its mechanism. OpenAI’s own fine print says the watermark does not verify accuracy, determine ownership, measure human contribution, or prove authorship.6 That is an honest disclaimer, and also a complete confession: the watermark cannot prove any of the things people will actually use it to prove.
Nietzsche’s ladder — transparency at zero marginal cost
The Camel carries the inherited burden with dignity: a regulation exists, Article 50 demands machine-readable provenance, incumbents have until December 2, and two of them have carried the duty early — Anthropic worldwide since August, OpenAI across the EU now.24 Compliance is real work; the deadline is real; the obligation is genuinely owed.
The Lion answers: you have carried the tablets well — now ask who wrote them, and when. OpenAI possessed this technology and withheld it for two years, because unilateral honesty was competitively punishing: watermark your output alone and users drift to unmarked rivals.2 The EU did not create transparency; it synchronized costs. It converted a prisoner’s dilemma into a compliance schedule, and only then did virtue become affordable. Note also what the regulation did not touch: the mark is compulsory, but the reading of the mark is rationed — the detector opens by application, to approved researchers.1 The power to accuse is the scarce resource in this system, and it is being hoarded by the party with the most reputational stake in every dispute.
The Child, answering the Lion’s no: good, the idol of voluntary transparency is broken — now build in the cleared space. The pieces for a real provenance commons already sit in the announcement. OpenAI plans to open-source the watermarking technology;7 the same logic, applied honestly, demands open detector access, published false-positive rates, keys held by third parties, and independent verifiers. Provenance can become public infrastructure — boring, auditable, boring again — instead of a private evidence locker with a public mailbox.
Inversion — designing the false-accusation machine
The goal: text provenance that serves the public. Invert it: how would you guarantee text provenance fails?
| # | Failure mode | Likelihood | Damage |
|---|---|---|---|
| 1 | A fragile mark treated as proof — 25% of words edited erases it, yet institutions read detector output as verdict3 | High | High |
| 2 | False positives on human text — non-native speakers flagged, a risk OpenAI itself raised in 2024 and has not declared resolved3 | Med | High |
| 3 | The interested party as sole arbiter — the vendor made the mark and controls the only reading of it | High | High |
| 4 | Jurisdictional laundering — provenance becomes a property of billing address, not of text; content routes through unmarked regions | High | Med |
| 5 | Compliance theater — a statistical gesture satisfies the statute while accuracy and authorship disputes go unresolved | High | Med |
The guards follow directly: detector output legally admissible only as corroboration, never as sole evidence; published false-positive rates audited by parties who do not sell the model; detector custody independent of the vendor; and a global default with a documented opt-out, rather than transparency that stops at the EU border.
Question forge — before the ink dries
The question coverage keeps asking — does watermarking work? — is operational, and its answer is already public: it works until someone edits a quarter of the words.3 The question is doing something else: it smuggles the assumption that the interesting variable is evasion, and it keeps the spotlight safely on the machines.
The forged question: when the detector says machine and the writer says human, who must prove what — and to whom?
Every load-bearing choice in this rollout is a private, pre-dispute answer to that question. Who holds the key, who may read the mark, what error rate is tolerable, whether a vendor’s denial of detection counts as a defense — each is an answer made in advance, by one party, before any student, journalist, or contractor stands accused. Living with the question means carrying it into every classroom policy, newsroom standard, and procurement clause that cites a detector — and noticing each time someone answers it on your behalf.
What the four frames reveal together
The mark is cheap to make and expensive to contest. Regulation’s real achievement here was not transparency but synchronization — it made honesty cost-competitive by making it universal, which is why the ink arrived two years after it was invented. And the architecture, as shipped, concentrates in one company both the power to mark the world’s text and the rationed power to read the mark back. The deepest risk of textGrain is not that machines will slip the mark. It is that humans will be convicted by it.
-
https://techcrunch.com/2026/10/05/openai-will-start-watermarking-chatgpts-text-in-the-eu ↩ ↩2 ↩3 ↩4
-
https://sea.mashable.com/tech/55527/openai-is-adding-an-invisible-watermark-to-ai-generated-text ↩ ↩2 ↩3 ↩4
-
https://www.theverge.com/ai-artificial-intelligence/1004880/openai-chatgpt-text-watermarks-eu-ai-act ↩ ↩2
-
https://pasqualepillitteri.it/en/news/20962/openai-textgrain-invisible-text-watermark ↩
-
https://finance.biggo.com/news/4eb99749-1844-4e81-8f29-6dcdd8a370aa ↩
-
https://9to5mac.com/2026/10/05/openai-details-new-text-watermarking-system-for-chatgpt-codex-and-the-api ↩