ChatGPT Will Watermark Its Text in the EU, Not Elsewhere Yet

Yes in the European Union, and not elsewhere for now: OpenAI says it will add an invisible watermark to eligible ChatGPT and Codex text in the EU over the coming weeks, on all plans, and is “not making text watermarking a global default at launch”. OpenAI’s announcement and Help Center describe no setting that lets a ChatGPT user turn it off.

The one switch OpenAI does document is for developers. In its API, watermarking is off unless a customer turns it on.

Where the watermark is on

Where you use it	Text watermark	Can you switch it off?
ChatGPT and Codex in the EU, all plans	On for eligible output, rolling out over the coming weeks from 5 October 2026	OpenAI describes no setting
ChatGPT and Codex outside the EU	Off; not a global default at launch	Nothing to switch off; OpenAI describes no opt-in either
OpenAI API, any country	Off by default	Yes: it is opt-in, per project or for a whole organisation, model by model
OpenAI models through cloud partners	OpenAI says it is working with partners “in the coming weeks”	OpenAI does not say

Outside the EU, ChatGPT text carries no OpenAI watermark at launch, and inside it there is no documented way to opt out.

API customers find the switch under Organization settings › Data controls › Text provenance, or Project Settings › Text provenance for one project: turn on Allow text watermarking, choose the models, then Save.

Images and audio already carry marks. OpenAI’s Help Center says images from ChatGPT, Codex and the API carry Content Credentials (the C2PA standard) and a SynthID watermark, and audio carries SynthID. Those can be checked by anyone at openai.com/verify. Text cannot.

What the mark actually is

OpenAI calls it textGrain. As the model writes, it chooses each next word piece at random from a set of likely candidates; textGrain tilts those choices according to a secret key, in a way that averages out to the model’s normal odds. A detector holding the same key checks whether a passage follows the pattern more often than chance would allow.

That has two practical consequences. There is nothing to find by looking: OpenAI says it “does not add hidden characters, invisible spaces, or unusual punctuation”, so there are no stray characters to strip out. And it needs room to work: the Help Center says answers that must be precise, like a maths result, and text the model is asked to reproduce unchanged leave little room for the pattern.

OpenAI reports no meaningful quality cost. On the eight benchmarks it published for its Astra model, the largest gap between watermarked and unwatermarked scores is about three points, and on several the watermarked version scored higher; OpenAI says the differences fall within the noise it normally sees between runs.

Can anyone tell your text came from ChatGPT?

Not through the watermark, at launch. OpenAI is opening its text detector only to approved researchers and expert organisations, case by case, and says that “given the risk of missed watermarks and false positives, we are not making it publicly available”. A teacher or an employer has no OpenAI tool to run your essay through.

Even with the detector, OpenAI’s own figures show where it struggles, all measured at a 1% target false-positive rate:

Condition, per OpenAI	Watermark found
400-token passage, psychology question	about 95%
200-token passage, same kind of content	about 80%
Maths content	“substantially lower”
Across the 24 official EU languages	from 69.0% (Spanish) down to 42.2% (Romanian), before OpenAI strengthened the weaker ones

Short text is the gap. The EU Code of Practice does not require a watermark on outputs under 200 tokens, which OpenAI puts at about 150 words of English, or on code snippets. A token is the unit a chatbot counts text in, roughly three-quarters of an English word.

The tools people already worry about are a different thing. Classifier-style detectors guess from writing style after the fact, and the evidence on how well those work is poor enough that OpenAI withdrew its own in 2023. textGrain instead checks for a signal OpenAI planted with a secret key, and only works on text from OpenAI models that carried it.

What it survives

Copying and pasting keeps it, as long as the wording is unchanged. Editing wears it down, fast:

• Replacing 10% of words with synonyms cut detection on 400-token passages from about 92% to 66%.
• Replacing 25% of words cut it to 17%.

The Help Center adds that “substantial rewriting, paraphrasing, or translation can make detection less reliable”, and that heavier changes “can make a watermark undetectable”. A quarter of the words swapped, and OpenAI’s detector found the mark 17% of the time in its own test.

What a detection would and would not mean

OpenAI is unusually direct about the limits. A watermark hit says an OpenAI model “likely generated or processed” some of the text. By OpenAI’s account, it does not:

• show how much a person wrote or edited;
• identify the user: it carries no account, prompt or conversation;
• say who owns the text or whether its use was lawful;
• say whether the text is accurate.

And a miss proves nothing. OpenAI lists the reasons its own text can come back clean: too short, edited, translated, made before watermarking started, made by an unsupported model, or made with another company’s tool. No watermark is not evidence of human authorship. What OpenAI keeps of the conversation behind the text is a separate question, covered in whether ChatGPT trains on your chats.

Why only the EU

The EU AI Act requires it there. The European Commission summarises Article 50 as requiring providers of generative AI to ensure output, text included, is “marked in a machine-readable format and detectable as artificially generated or manipulated”, and says those obligations apply from 2 August 2026. OpenAI says it signed the Commission’s Code of Practice on transparency of AI-generated content, which the Commission accepts as a way to show compliance.

OpenAI cites no rule outside the EU that requires it, and says the regional rollout “gives us room to learn from real-world use and feedback”. It says it will revisit each part of the approach, but has given no date for a wider default. Ars Technica, reporting the change on 6 October, framed it as on by default only where regulators require it.

What OpenAI has not said is which output counts as “eligible”, or whether “in the EU” follows your location or your account. Until it does, an EU user should assume longer ChatGPT replies carry the mark.

Questions people ask

Does ChatGPT watermark text?

In the EU, yes, from a rollout OpenAI says takes place over the coming weeks after 5 October 2026, on all plans. Elsewhere it is off: OpenAI says it is not making text watermarking a global default at launch.

Can people tell my text came from ChatGPT?

Not from the watermark, for now. It is invisible to readers, and OpenAI’s detector is limited to approved researchers and expert organisations. Schools, employers and the public cannot use it at launch.

Can I turn off the ChatGPT watermark?

OpenAI’s announcement and Help Center describe no setting for ChatGPT users to turn it off in the EU. The only switch they document is for API customers, where watermarking is off unless the customer turns it on.

Does the ChatGPT watermark add hidden characters?

No. OpenAI says it adds no hidden characters, invisible spaces or unusual punctuation. The mark is a pattern in which words the model picks, so copying and pasting the text carries no extra material.

Does paraphrasing remove the ChatGPT watermark?

It weakens it. In OpenAI’s test on 400-token passages, swapping 10% of words for synonyms cut detection from about 92% to 66%, and swapping 25% cut it to 17%. OpenAI says substantial paraphrasing or translation can make it undetectable.

Does the watermark identify who wrote the text?

No. OpenAI says the watermark does not associate a person, account, prompt or conversation with the text, and that a detection does not show how much a human contributed.