AI Watermark Deadline Moved Up: Can ChatGPT's Watermark Survive an Edit?
OpenAI's EU watermark for ChatGPT text finds about 95% of unedited passages and about 17% after a quarter of the words change, and the EU just moved its labeling deadline up to December 2, 2026.
Key Takeways
- OpenAI's new EU watermark for ChatGPT and Codex text detects about 95% of unedited passages, but accuracy drops to about 17% once a quarter of the words are swapped for synonyms.
- The EU's AI Act labeling deadline for AI-generated text moved up to December 2, 2026, two months earlier than the European Commission's original proposal, even as several other AI Act obligations were pushed later.
- Watermark-detection access stays limited to approved researchers and vetted organizations, so most ChatGPT users and the general public cannot verify the watermark themselves.
- Enterprise and government teams that edit AI-assisted drafts before sending them should not rely on the watermark alone; a named human reviewer and a disclosure policy are the safer second layer.
Bring Alex to your organization: keynotes, workshops and advisory, in person or virtual. Check Alex's availability
OpenAI's new watermark for ChatGPT text in the EU finds about 95% of untouched passages, and about 17% once a quarter of the words have been swapped for synonyms. Those are OpenAI's own published numbers. And the EU deadline that makes this watermark necessary just moved closer, to December 2, 2026.
Anyone who has read a cover letter lately and wondered who, or what, wrote it already understands the question this rule is trying to answer. The EU AI Act requires generative AI providers to mark AI-written text in a machine-readable way, so that a platform or an investigator can eventually tell the difference. OpenAI announced its answer on October 5: a hidden watermark called textGrain, now rolling out to eligible ChatGPT and Codex outputs across all plans in the EU.
The EU moved this deadline closer while the tool is still maturing
Several AI Act timelines just moved later. In their Digital Omnibus on AI agreement, the European Parliament and the Council of the EU pushed a group of obligations back. High-risk rules for biometrics, critical infrastructure, education, employment, migration, and border control now apply from December 2, 2027. Rules for AI built into physical products such as lifts and toys apply from August 2, 2028.
The labeling obligation went the other way. The co-legislators set it to apply from December 2, 2026, 2 months earlier than the European Commission's original proposal of February 2, 2027. That was a deliberate choice, and a defensible one. Of everything in the AI Act, knowing whether a person or a machine wrote what you are reading is the protection I expect ordinary citizens to feel most directly, and the Parliament and the Council treated it that way.
The governing idea behind my writing on AI governance is rules written for a slower world: institutions governing something that changes faster than any rule-making cycle can. This story turns that idea around. The rule-makers set a fast clock on purpose. The hard part now sits with the technology built to meet it.
How the ChatGPT watermark works, and where it breaks
textGrain nudges which words and word-pieces the model picks at each step, leaving a statistical pattern that a compatible detector can recognize. A reader sees ordinary text, with no extra characters or spacing. Responses under 200 tokens are excluded from the text method, and code gets separate handling because it leaves the model fewer word choices.
On clean text, the method works well. Edited text is a different story. Here is what OpenAI published, measured at a 1% false-positive rate on 400-token passages:
| Share of words replaced with synonyms | Watermark detected (OpenAI's published data) |
|---|---|
| 0% (unedited) | About 95% |
| 10% | About 66% |
| 25% | About 17% |
A light edit takes detection from about 95% to about 66%. A heavier one takes it to about 17%. OpenAI also says translation and substantial paraphrasing can interfere, and that "a negative result is also inconclusive." When the detector finds nothing, that tells you nothing about who wrote the text.
Access is narrow for now, too. Approved researchers and vetted expert organizations can request the detector, and OpenAI grants it case by case. Ordinary ChatGPT users and the public cannot run the check at launch.
A watermark that fades after a 10% edit is a strong first step toward provenance and a weak basis for any single decision about a person.
OpenAI is doing the hard part in public
OpenAI published the weak spots of its own compliance tool about 2 months before the deadline. That candor is useful. It gives the AI Office and the co-legislators real numbers to work with as implementation guidance takes shape, and it gives every other provider a public baseline to beat.
The constraint is stubborn. Text leaves little space to hide a mark, and every human edit erases part of the pattern. Any serious team working on text provenance runs into the same wall, and OpenAI and the EU are now working through it on a fixed date. The company is already under scrutiny on other fronts this month, including the California attorney general's subpoena, which makes a public, measurable compliance effort worth more to it, and to everyone watching.
Enterprise and government leaders carry the practical risk
AI-assisted writing inside a company or an agency usually gets edited before it ships. A staffer drafts a memo in ChatGPT, tightens a few sentences, and sends it. By OpenAI's numbers, that light edit is enough to make the watermark unreliable. A compliance plan that counts on the watermark to do the labeling rests on a version of the text that almost nobody sends.
Watermarks will help researchers and investigators trace large volumes of machine-written content. Deciding whether one contract clause or one public comment came from a person stays a human call, made with context and accountability attached. Advising the California State University system on AI governance has taught me that the hardest provenance questions arrive one document at a time, with one person's name on it. I call that work Above the Algorithm: the judgment and accountability that stay human even as the tooling improves. It is the same discipline I wrote about in governing AI agents, where a control you can verify is worth more than a control you assume.
The question to bring to your next leadership meeting
December 2 is 6 weeks away. Before then, ask your team one question: if a regulator or a customer asked us to show which of our published text came from AI, what would we hand them? If the answer is "the watermark," the plan needs a second layer, such as a disclosure policy and a named reviewer who signs off.
I am taking that question into my own advisory work this month, and I expect the answers to vary more than anyone would like. If you want a second set of eyes on yours, I'm easy to find.
Can you tell if ChatGPT text has a watermark?
OpenAI's detector can check ChatGPT and Codex outputs in the EU for its textGrain watermark, but access is narrow. Only approved researchers and vetted expert organizations can request a check, granted case by case. The general public and most ChatGPT users cannot run this check themselves at launch, and OpenAI says a negative result does not prove a human wrote the text.
Does editing AI text remove the ChatGPT watermark?
Editing weakens the watermark fast, according to OpenAI's own published numbers. Detection on a 400-token passage runs about 95% when the text is untouched, drops to about 66% once 10% of the words are swapped for synonyms, and falls to about 17% at a 25% swap rate. Translation and heavy paraphrasing can interfere with detection as well.
When does the EU AI Act require AI-generated text to be labeled?
The labeling obligation takes effect December 2, 2026, under the Digital Omnibus on AI agreement between the European Parliament and the Council of the EU. That date is 2 months earlier than the European Commission's original proposal of February 2, 2027, even as several other AI Act obligations were pushed later in the same deal.
Here is what makes Alex a credible voice on this topic: As an AI governance advisor to the California State University system, and after 20 years at Cisco, where he shaped a $1.1B innovation portfolio, Alex helps leadership teams turn new AI rules into compliance plans that hold up against real deadlines.
Bring Alex in to pressure-test your AI compliance plan before December 2 →
.svg)
