ChatGPT Now Fingerprints Its Own Writing. Edit Accordingly.

OpenAI has started stamping an invisible signature into the words ChatGPT writes. It is live in the EU for ChatGPT and Codex, and available worldwide as an API switch developers can flip. If you publish content for clients, the question is no longer whether AI-assisted copy can be identified — it is who gets to check, and when.

“OpenAI will watermark ChatGPT and Codex text in the EU to comply with the AI Act. Editing can make the invisible marks harder to detect, it says.”

Aditya Mehta, TechCrunch

Our take

The mechanics matter, so start there. TechCrunch reports the system is called textGrain, described in OpenAI’s technical report as “entropy-calibrated watermarking for language model text” and co-written with researchers from the University of Pennsylvania and Yale. It works by using a secret key to sort the model’s next-word predictions, nudging word choices hundreds of times across a passage until a detector holding that key can recognize the pattern. Because the signal lives in the word choices themselves, it survives copy and paste. OpenAI says it does not identify the user and that it saw no meaningful change in model performance.

Where it applies is more uneven than the headline suggests. For ChatGPT and Codex, the rollout is EU-only and reaches eligible users on all plans over the coming weeks — it is not a global default at launch. For the API, it is worldwide and available for select models from October 5, 2026, but off by default. The driver is the EU AI Act’s transparency rules, which took effect August 2 and require AI-generated content to be marked so other systems can identify it.

Now the asymmetry nobody is pricing in. A watermark that is present is strong evidence that an OpenAI system wrote the text. A watermark that is absent proves nothing at all — OpenAI says so itself, in plain language: its absence “does not prove human authorship.” That is a one-way ratchet. You can be caught shipping raw model output; you can never produce a clean scan as proof that a human wrote something. Any agency that was hoping detector technology would eventually let it certify human authorship should let that idea go.

The number to actually plan around is the degradation figure. OpenAI’s own tests found that replacing 10% of words with synonyms dropped detection from about 92% to 66%. Short passages, math answers, and translated text are also harder to detect. Read that honestly and it cuts both ways. As an enforcement mechanism, the watermark is fragile — a light editing pass, which any competent content process already includes, knocks a third off detection confidence. But as a differentiator, “we write by hand” becomes unprovable, because edited AI output and human output converge into the same unmarked bucket. The market cannot verify the claim, so the claim stops being worth much.

Detector access is the quiet risk. Right now OpenAI is limiting it to approved researchers and expert organizations, which means your client cannot run a check on today’s deliverable today. It does not mean they never will. Watermarks are durable and text is archived, so anything you publish this quarter is potentially checkable in a later quarter, under whatever access rules exist then. The only safe posture is to assume retroactive checkability and ship work you would be comfortable having scanned in two years.

There is also a supply-chain wrinkle specific to agencies. The API flag is off by default and controlled by the developer, not by you. If your content runs through a third-party writing tool, SEO platform, or chatbot vendor built on OpenAI models, that vendor decides whether your output carries a mark. You will not see it in the interface. That is a question for a vendor questionnaire, alongside data retention and sub-processors — and if a vendor cannot answer it, that is itself the answer.

Finally, keep the categories straight. A watermark is a provenance signal, not a quality signal. It says a machine touched the passage; it says nothing about whether the passage is accurate, useful, or worth reading. OpenAI makes the limit explicit — the mark can indicate that an OpenAI system “generated or processed part of a passage, but not how much human judgment” went into it. The exposure this creates is contractual and regulatory, not editorial. If your agreements promise original work, the words “original” and “AI-assisted” need a definition both sides signed.

What this means for your business

• Define “AI-assisted” in your client contracts this quarter. One paragraph stating what role AI plays in drafting, what a human does before delivery, and who owns the output. Vague language is what turns a watermark into a dispute.
• Ask every AI content vendor one question in writing: is watermarking enabled on your API calls? The flag is off by default and theirs to control. Add it to your vendor review next to data retention.
• Make substantive editing a required step, not a nice-to-have. Not synonym-swapping to defeat detection — actual restructuring, original examples, and verified facts. That is also what makes the content worth publishing.
• If you serve EU clients or audiences, log your AI disclosure position now. The AI Act transparency rules are already in force. A short internal note on what you disclose and where beats reconstructing it under pressure.
• Stop selling “100% human-written” as a guarantee. You cannot prove it, the absence of a mark will not prove it, and the claim creates liability you do not need. Sell accuracy, expertise, and results instead.

Can a client detect AI in content I deliver?

Not today for most people — detector access is limited to approved researchers and expert organizations. Assume that changes, and that archived text stays checkable.

Does editing remove the watermark?

It weakens it. OpenAI’s tests show a 10% synonym substitution cut detection from roughly 92% to 66%. Heavier rewriting degrades it further, but nothing guarantees removal.

The technical report details, rollout scope, and detection limits are in the original report: OpenAI will start watermarking ChatGPT’s text in the EU, by Aditya Mehta for TechCrunch.

---