The ChatGPT watermark is coming to the EU: what textGrain can prove, and what it cannot

What the ChatGPT watermark is, and what OpenAI announced on 5 October#

On 5 October OpenAI set out how it will mark text under the EU AI Act. In its announcement, the company said that from that day API customers worldwide can opt in to text watermarking for select models, that it stays off by default in the API, and that over the coming weeks it will add an invisible watermark to eligible ChatGPT and Codex text in the European Union. It also opened applications for its detector, with access limited at first to approved researchers and expert organisations. The Next Web reports that the change covers eligible users on every plan, in the EU only, and that OpenAI is not making text watermarking a global default at launch.

That is the whole shape of the ChatGPT watermark today: a switch for developers, a default for one jurisdiction, and a detector most people cannot touch. The rest of this piece is about what that shape means for anyone who publishes words for a living, and it is a smaller thing than the headlines suggest, and a more useful one.

Does ChatGPT watermark text today?#

Not by default, and not outside the EU. For eligible EU users the answer becomes yes within weeks. For API customers it is yes if they have switched it on. ActuIA reports that API customers do this in a Text provenance setting under organisation or project data controls, and that it is not a request parameter. Because the date is "the coming weeks" rather than a day, treat any claim about what ChatGPT output carries on a given morning as unproven until you can read the setting yourself.

What textGrain does to a sentence#

A language model picks each word from a ranked list of likely next words. textGrain tilts that pick using pseudorandom values derived from a secret key and the words that came before. Nothing visible changes. OpenAI's technical report, written with researchers from the University of Pennsylvania and Yale, calls the randomness it gives up an entropy budget, and the detector later tests whether the word choices lean toward the key more than chance would allow.

This is why the ChatGPT invisible watermark is invisible in a literal way. ActuIA reads the report as adding no hidden characters or odd spacing, because the signal lives in which words were chosen. In our reading, that means a character cleaner has nothing to clean. It also means the mark is a statistical scent, not a stamp: it gets fainter as the wording is changed. The idea is not new either; the field's reference point is a 2023 paper that promoted a randomised set of "green" words during sampling, and OpenAI cites that line of work in its own report.

Mind the vendor's own framing. OpenAI says textGrain matched or exceeded other approaches, including Google DeepMind's SynthID for text, and it also says textGrain "does not guarantee reliable detection". Both are the company talking about its own system. We carry them as claims, not as rankings.

How well the ChatGPT watermark detector works, on OpenAI's own numbers#

Every figure in this section was published by OpenAI about OpenAI's detector. No outside party can reproduce them yet. A 2026 governance paper on AI text watermarking after the EU AI Act argues that "neither the objections nor the assurances can currently be verified", because no public tool can test the deployed systems. Read the numbers as a fox reads tracks in snow: a good clue, not a confession.

• 95%

  detected in 400-token English passages, unedited

• 80%

  detected in 200-token passages

• 66%

  after 10% of words swapped for synonyms

• 17%

  after 25% of words swapped

The ladder is the unedited case. A marketer's draft is rarely that. Search Engine Journal sets out OpenAI's edit test: swapping some words for synonyms in 400-token English passages cut detection from approximately 92% to 66% at ten per cent, and to 17% at twenty-five per cent. Short passages, maths answers and translated text were weaker still.

None of this should shock a researcher. A 2023 study found that paraphrasing evades several detectors, including watermarking, while proposing a retrieval defence that sits with the provider. The new fact is not that edits weaken a watermark. It is that the maker of the world's most used chatbot has printed the size of the weakening, and shipped anyway.

Why a free ChatGPT watermark detector cannot be real#

Here the report matters. It says the detector needs only the generated text and the secret key. OpenAI is keeping that key private, and its developer documentation says text verification is currently available only to approved bodies, including research and academic institutions. So in our reading, a web page offering a free ChatGPT watermark detector cannot be reading OpenAI's mark. At best it guesses at style. Anthropic is similar: its help page says detection is in private preview for eligible organisations. Treat any free verdict, in either direction, as the sound of a fox in the next field: something is moving, but you have not seen it.

Does ChatGPT have a watermark? Three vendors, three different answers#

Does ChatGPT have a watermark? In the EU, eligible text gets one within weeks, and elsewhere only API customers who opt in. But a content team rarely uses one tool. Anthropic says it marks Claude output worldwide by default, giving as its reason that it does not yet have a durable way to scope it by region. Google, according to the governance paper cited above, has deployed SynthID-Text in Gemini. Your drafts may already be marked by a vendor you were not watching.

• OpenAI

  ChatGPT and Codex

  Default

  On in the EU within weeks; API opt-in, off by default

  Reach

  EU for products; worldwide for API opt-in

  Detector

  Approved researchers and expert organisations

  Best for

  • Teams with EU readers
  • API builders who want a record

• Anthropic

  Claude, all supported models

  Default

  On, worldwide; no user opt-out, per the arXiv paper

  Reach

  Every Claude surface, per its help page

  Detector

  Private preview, eligible organisations

  Best for

  • Teams who assume global scope
  • Anyone drafting in Claude

• Google

  Gemini, SynthID-Text

  Default

  Deployed in Gemini, per the arXiv paper

  Reach

  Not examined here

  Detector

  Not examined here

  Best for

  • Teams on Gemini and Workspace
  • Anyone comparing the three

The picture is uneven because the law is. The EU AI Act sets one duty for the people who build the model and another for the people who publish its output. That difference is the heart of the next section, but first, a reaction worth hearing from the practitioner end of the timeline.

Weigh that poster fairly. The post discloses that its author sells a hosted watermark-removal service, ten test reports is a small sample, and it tested the open Google scheme, not OpenAI's. We quote it for its last sentence, which is the most honest line in the thread: nobody outside the labs can currently check either claim. Whether a rewrite scrubs a mark is a question for the detector's owner, and the owner has restricted the tool.

So the fox's answer to does ChatGPT watermark text is a staged one: yes for some, soon for others, optional for developers, and unverifiable for everyone else. That is awkward for a marketer, and it is exactly the kind of awkward folkfox is built to work in. If you want the wider frame for governing AI in a content operation, our AI consultancy work starts from this sort of dated, sourced uncertainty rather than from a vendor's slide.

What the EU AI Act asks of a brand that publishes, in plain terms#

Start with who owes what. Article 50(2) of the AI Act says providers of generative systems shall ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. That duty falls on OpenAI, Anthropic and Google. The watermark is their homework, not yours.

Your duty sits in Article 50(4), and it is narrower than people fear. For text, it covers deployers who publish AI-generated text with the purpose of informing the public on matters of public interest. Even there, the article carves out content that has undergone human review or editorial control, where a person or organisation holds editorial responsibility for the publication.

The Commission's Code of Practice, published on 10 June 2026, is voluntary, but it states the practical line plainly: AI-generated or AI-manipulated text published on matters of public interest must be clearly labelled. The Commission adds that the transparency rules apply from 2 August 2026.

Look at the exemption's hinge, because it is the whole game. Someone holds editorial responsibility. That is a job and a name, not a tool, and a watermark does nothing to give you one. A brand with a named editor, a review step and a record of both is on firmer ground than a brand with a clever detector. We have walked through the labelling side before, in a trade body's disclosure rule and in an Italian regulator's ruling on a label that was on the screen, and the same logic holds for text.

Most marketing copy is not a matter of public interest in the Act's sense, and we are not lawyers. But copy about health, money or safety can drift that way, and a regulator, a client or a platform may ask how a page was made. The thicket is thinner for teams who can answer in one line.

Will the ChatGPT watermark matter for SEO? What Google's own page says#

Content leads ask this first, so take it head on. Google's page on using generative AI content in Search, updated on 1 October 2026, says that using generative AI to produce many pages without adding value for users may violate its spam policy on scaled content abuse. The page is about value, not production method. As we read it, the page says nothing about text watermarks.

The strongest outside reading is from Seer Interactive, which argues that Google has used SynthID-Text on Gemini for two years without making it a ranking factor, and that the watermark is a compliance matter, not a quality signal. That is an analyst's argument, not a Google statement, and silence is not a guarantee. Our line is the careful one: nothing we read says a watermark changes ranking, and nothing says it never could.

There is a quieter point, and it is our inference. A watermark survives best in text that nobody has edited. Text that nobody has edited is also the text that Google's spam policy and the Search Quality Raters' guidance on low-effort content are aimed at. So the pipeline most likely to keep a readable mark is the same lazy pipeline most likely to lose traffic anyway. Edit like a human with something to say, and the mark fades while the page improves. Our notes on reviewing AI content before Google does and on AI content in a spam update lay out that review step.

Do not read this as advice to strip a mark. Editing for voice, evidence and accuracy is the craft. Editing in order to launder a draft is a different act, and it leaves the brand with a record problem rather than a detection problem. Our SEO and GEO work and content marketing work are built on the first kind of editing.

What a detector verdict should and should not change#

If a supplier's article is flagged by a free tool, you have learned almost nothing. If it is not flagged, you have learned less. A better audit asks the supplier for the provenance record: which tool, which review, which editor, which date. Anthropic's own help page says a missing mark does not mean the content was not AI-generated, and OpenAI says the same of its own system. When the vendors tell you the instrument has limits, believe them.

A Monday plan for a content team facing the ChatGPT watermark#

Turn all of this into a routine that a content lead can own. It asks for an afternoon, a shared document and a refusal to treat a detector as a judge.

Page: [URL or title]
AI tools used: [none / name each, and for which step]
Human review: [what was checked, by whom]
Editor with editorial responsibility: [name]
Published: [date]
Notes: [sources checked, claims removed]

The record asks for the very thing a watermark cannot supply: a human being who answers for the page. It will not satisfy every regulator, and we are not claiming it does. It will let you answer one plain question the same way every time.

Do that and the fox is no longer waiting on a vendor's scent trail. The quarry becomes a pipeline that can show its working, whatever any detector says on any Tuesday. The ChatGPT watermark will arrive on its own schedule, and the patient prowl of a good editor is the thing no key can read.

Frequently asked questions#