OpenAI is changing how it plans to make AI-generated text traceable. On 5 October 2026, the company introduced textGrain, an invisible statistical watermark for text. The technology is due to roll out gradually over the coming weeks for eligible ChatGPT and Codex outputs in the European Union, across all subscription tiers. For API customers, textGrain is already available worldwide as an optional feature for selected models, although it remains switched off by default.
The move brings a long-running debate about whether ChatGPT text can be technically marked into a new phase. Until now, ordinary text responses did not contain a reliable hidden provenance signal. Rumours about invisible Unicode characters or unusual spaces persisted for years, but these were not genuine OpenAI watermarks. textGrain works in a fundamentally different way: the signal is embedded in the model’s choice of words rather than in special characters.
The main driver is the European AI Act. Since 2 August 2026, new transparency requirements have applied to certain AI systems and AI-generated content. Providers of generative AI are required, where technically feasible, to design their systems so that artificially generated or manipulated content can be detected in machine-readable form. OpenAI explicitly links textGrain to those requirements and to the European code of practice on AI-generated content.
The Watermark Is Embedded in Word Choice
Technically, textGrain is very different from a visible watermark placed on an image.
A language model does not generate a finished paragraph in one go. It predicts the next word, or more precisely the next token, step by step. In many situations, several different words or formulations may all be equally suitable.
This is where textGrain intervenes.
OpenAI slightly alters the probabilistic process by which the model chooses between suitable words or token sequences. Across a sufficiently long passage, those small choices form a statistical pattern. The text is intended to look completely normal to a reader, but a detector with the relevant key can search for that pattern.
Importantly, textGrain does not insert invisible letters, unusual spaces or special punctuation. Copying ChatGPT output into a plain-text editor, stripping formatting or removing hidden Unicode characters does not automatically remove the watermark. The signal is carried by the wording itself.
A Single Word Reveals Almost Nothing
The system works statistically, which means a single word cannot tell you whether textGrain was used.
Only after a sufficiently long passage has been generated are there enough model decisions for the detector to identify a pattern. That makes text length one of the technology’s most important limitations.
OpenAI reports that, at a target false-positive rate of one per cent, detection reaches around 80 per cent for passages of 200 tokens. At 400 tokens, detection rises to roughly 95 per cent for relatively unconstrained writing tasks such as psychology-related prose. In more constrained areas such as mathematics, the detection rate is considerably lower because the model has less freedom in its word choices.
Depending on language and writing style, 200 tokens correspond to roughly a few paragraphs. This alone shows why textGrain cannot become a universal AI detector. Short emails, social posts, headlines or brief responses may simply not contain enough statistical information for reliable detection.
Short Texts and Code Are Special Cases
The European transparency code explicitly acknowledges this limitation. Text outputs shorter than 200 tokens, as well as code snippets, are not subject to the same watermarking expectation.
This matters particularly for Codex. Source code often gives the model much less freedom than ordinary prose. Variable names, syntax, API calls and mathematical expressions cannot simply be changed at will without affecting functionality. That reduced flexibility makes statistical watermarking much harder.
OpenAI therefore refers specifically to eligible ChatGPT and Codex outputs rather than suggesting that every single piece of generated text will be marked.
The Rollout Is Under Way, but Not Every EU Text Is Already Watermarked
An important distinction concerns the current stage of deployment.
OpenAI announced textGrain on 5 October and made the API option available immediately. In ChatGPT and Codex within the European Union, however, the rollout is taking place over several weeks.
That means that, as of mid-October, it would be wrong to assume that every ChatGPT response generated by every EU account already contains the watermark.
OpenAI describes the deployment as gradual and says it will apply across all tiers for eligible users. Outside the EU, textGrain is not initially being introduced as the default for ChatGPT and Codex.
For any individual piece of text, the date on which it was generated is therefore not enough on its own to determine whether the watermark was present.
API Customers Face a Different Set of Rules
OpenAI is taking a more flexible approach with the API.
API customers worldwide can activate textGrain for selected models, but the feature remains off by default. Organisations can therefore decide for themselves whether automatically generated text should carry the provenance signal.
For businesses, this distinction matters. Anyone generating content through the OpenAI API using n8n, custom software, a CMS or a customer platform should not assume that those outputs are automatically watermarked. The setting has to be enabled deliberately and is only available for supported models.
OpenAI is also working with cloud and distribution partners to make watermarking available for OpenAI models accessed through third-party platforms, although exact availability may vary by provider and model.
Even Light Editing Weakens Detection Significantly
The biggest weakness of textGrain lies in something that is especially easy to do with text: rewrite it.
OpenAI has tested what happens when individual words are replaced after generation. In 400-token English passages, the detector initially identified roughly 92 per cent of watermarked texts. When ten per cent of the words were replaced with synonyms, the detection rate fell to around 66 per cent. When 25 per cent of the words were replaced, only about 17 per cent remained reliably detectable.
That is a substantial limitation.
A text does not even need to be rewritten from scratch. Editorial revision, heavy shortening, paraphrasing or translation can significantly alter the statistical pattern. OpenAI explicitly identifies rewriting and translation as situations in which detectability may fall.
That makes textGrain more suitable as a large-scale provenance signal than as forensic proof about an individual document.
A Positive Result Does Not Mean “This Was Written by AI”
Even a detected watermark has to be interpreted carefully.
OpenAI makes clear that a positive result can indicate only that an OpenAI system generated or processed part of the text. It cannot determine how much human work was added afterwards.
Consider an article drafted with ChatGPT but then heavily rewritten by an editor, supplemented with original research and structurally reorganised. Even if some of the textGrain signal remains, that does not mean ChatGPT was the sole author.
Nor does the watermark identify a person. The detector does not reveal the user account, prompt, chat history, organisation or identity of whoever generated the text. It also cannot establish ownership, copyright or legal responsibility.
And it is not a truth indicator. A watermarked text can be factually correct or incorrect. textGrain provides information about possible provenance, not about the quality of the content.
The Absence of a Watermark Proves Even Less
The reverse is even more important.
If the detector finds no textGrain signal, that does not mean a human wrote the text.
The passage may be too short. It may have been heavily edited or translated. It may have been generated using an OpenAI model that does not support the feature. It may have come from the API with watermarking disabled. It may pre-date the rollout or have been produced by a completely different AI provider.
textGrain therefore differs fundamentally from the idea of a universal test for AI authorship.
A positive result is a provenance signal. A negative result is not proof of human authorship.
The Detector Remains Restricted for Now
One of the most controversial questions is who will actually be allowed to check for textGrain.
OpenAI is not making the text detector publicly available at launch. Instead, selected researchers and specialist organisations can apply for access, with approval handled on a case-by-case basis.
OpenAI says the restriction reflects the technical limitations of the system. Because both false positives and missed detections are possible, a fully public detector could easily be misunderstood or misused.
This makes text very different from images and audio. For those media types, OpenAI already offers public verification tools and a Content Provenance API. There is currently no equivalent public interface for textGrain.
OpenAI has said, however, that it plans to make the underlying technology open source in future, with further details to follow.
That Makes textGrain Difficult to Use in Automated Publishing Workflows
For businesses and publishers, restricted detector access has immediate practical consequences.
A publishing workflow cannot currently send every article through a public textGrain API and automatically process the result in n8n, Make or a CMS. Organisations without access to the restricted detector programme cannot reliably verify the OpenAI statistical signal themselves.
Websites that claim today to detect or remove textGrain without access to OpenAI’s key should therefore be treated cautiously. Because the signal is embedded in statistical word choice, simply searching for invisible characters is not enough.
For API-based production workflows, the reverse is already possible: businesses can deliberately decide whether supported OpenAI-generated text should carry the watermark in the first place.
The Watermark Is Designed to Have Little Effect on Model Quality
An obvious concern is that forcing a model to favour certain words might make its writing worse.
OpenAI has therefore compared watermarked and non-watermarked outputs from one of its frontier models across several benchmarks. The results move slightly in both directions depending on the benchmark, and the company says it has found no meaningful overall loss in quality.
That is technically plausible because textGrain does not force arbitrary words into the output. Instead, it nudges the model between several options that were already considered suitable.
Even so, real-world effects will only become clearer after broader deployment. The available room for watermarking may be smaller in highly stylised writing, specialist terminology or languages with fewer interchangeable expressions.
Not All Languages Are Equally Easy to Watermark
OpenAI’s own testing already shows meaningful differences across the 24 official languages of the European Union.
Using one baseline configuration, detection reached about 69 per cent for Spanish, while Romanian achieved only 42.2 per cent. OpenAI says the strength of the watermark can be adjusted and that it has already experimented with stronger settings for languages with weaker detection.
This illustrates a fundamental problem with statistical text watermarking: not all languages offer the same degree of flexibility.
In languages or text types where many formulations are equally natural, a model has more freedom to choose between alternatives and embed a signal. Where wording is more constrained, that freedom decreases.
The Old Unicode Rumours Have Nothing to Do with textGrain
Throughout 2025, claims circulated that ChatGPT secretly inserted special Unicode characters or the narrow no-break space U+202F into text so that its output could later be identified.
textGrain does not validate those claims.
OpenAI now explicitly states that its actual watermark does not rely on hidden characters, invisible spaces or unusual punctuation.
Converting a text to plain text, deleting formatting or cleaning up Unicode therefore does not automatically remove textGrain. Conversely, the appearance of unusual whitespace is not proof that OpenAI deliberately marked a passage.
The EU AI Act Is the Main Trigger
It is no coincidence that OpenAI is introducing textGrain first in Europe.
Article 50 of the EU AI Act requires providers of certain generative AI systems to use technical measures that allow AI-generated or manipulated content to be detected in machine-readable form where technically feasible. The relevant transparency obligations have applied since 2 August 2026.
The accompanying European code of practice sets out more detailed approaches for labelling and detection. Participation in the code itself is voluntary, but the underlying legal transparency obligations are not.
For systems placed on the market before 2 August 2026, European transition rules can provide additional time for certain obligations. That helps explain why technical measures such as textGrain did not necessarily need to be fully deployed on the exact date the new rules took effect.
Machine-Readable Watermarking and Visible Labelling Are Two Different Things
For businesses, another important point is that textGrain does not automatically satisfy every transparency requirement.
The watermark is machine-readable and invisible to people. Depending on the use case, the AI Act may still require visible disclosure. This is particularly relevant for deepfakes and certain AI-generated or manipulated publications on matters of public interest where no appropriate human review and editorial responsibility are in place.
OpenAI itself stresses that provenance signals do not replace visible labels or other disclosures required by law.
For publishers, this means that the presence of textGrain is not a licence to dispense with every other form of transparency.
Text Is Much Harder to Watermark Than Images or Audio
The launch of textGrain also illustrates why text provenance is considerably harder than watermarking other media.
Images and audio contain enormous amounts of numerical information across which an invisible signal can be distributed. Minor transformations can therefore often be tolerated.
Written text contains far less redundant information.
Every editorial change alters the very medium in which the watermark is stored.
Text watermarking therefore involves a constant trade-off. The signal needs to be strong enough to detect, but subtle enough not to damage the quality or naturalness of the writing. At the same time, it should survive ordinary editing even though those edits directly alter the statistical structure that carries the watermark.
OpenAI’s own acknowledgement of relatively high miss rates in shorter or edited passages is therefore an important part of understanding the technology.
textGrain Is Not an Automatic Plagiarism Detector for Universities
The education sector is likely to take a particular interest in textGrain.
At first glance, it may look like a technical solution to the problem of determining whether an essay or assignment was written with ChatGPT. In practice, the system is not reliable enough for that purpose on its own.
A watermark cannot determine the proportion of human involvement or prove who created a text. Heavily edited or translated material can lose the signal, while a positive result may reflect perfectly legitimate AI assistance with editing or translation.
A single detector result would therefore not amount to reliable proof of academic misconduct.
OpenAI’s decision to keep the detector restricted initially may also reflect the risk that people would overinterpret individual results.
For Businesses, It Becomes a Governance Decision
In corporate use, the question is different.
Organisations using the API to generate marketing copy, product descriptions, customer communications, documentation or editorial drafts can now decide deliberately whether those outputs should carry a machine-readable OpenAI provenance signal.
For regulated industries or companies with their own transparency policies, that may be useful.
At the same time, businesses should not assume that textGrain replaces a complete provenance record. If a text is later heavily edited, translated or passed through another model, the signal may weaken or disappear.
A robust governance approach should therefore also document which models were used, where human approval took place and how content was changed before publication.
OpenAI Is Increasingly Combining Multiple Provenance Signals
textGrain is only one part of a much broader provenance strategy.
For supported images, OpenAI now uses both Content Credentials based on the C2PA standard and Google’s SynthID watermark. Audio outputs can also carry SynthID. For these media types, OpenAI also offers public verification options.
The logic is to combine several layers.
Metadata such as Content Credentials can carry detailed information about provenance and editing history, but it can be stripped during file transfers. Watermarks carry less information, but may survive certain transformations.
With textGrain, text is now gaining a comparable additional layer.
textGrain Does Not Solve the AI Detector Problem
The introduction is significant, but it should not be mistaken for a universal solution.
Under suitable conditions, textGrain can provide a statistical signal that an OpenAI system was involved in producing a passage. It cannot reliably determine whether a human or AI was the true author, nor can it identify every rewritten AI-generated text.
OpenAI’s own figures make the limitation clear: detection of around 95 per cent for suitable 400-token passages under favourable conditions falls to roughly 17 per cent after a quarter of the words have been replaced.
That is not a conventional software bug. It is a fundamental property of text watermarking. Once the text itself changes, the medium carrying the signal changes with it.
Even So, textGrain Changes How AI-Generated Text Can Be Traced
The wider significance of textGrain is therefore not that it will allow schools, publishers or companies to conclusively expose individual AI users.
Instead, it creates a systematically embedded provenance signal for ChatGPT-generated text in Europe. Platforms, researchers and specialist organisations may be able to use it to study more reliably how much AI-generated text is circulating across the wider information ecosystem.
For OpenAI, it is also a regulatory step. Europe is becoming the first major market in which eligible ChatGPT text outputs are expected to carry a machine-readable provenance signal by default.
The difference from earlier rumours is fundamental. A ChatGPT text will not contain some secret character that can simply be deleted. The watermark emerges from many small decisions about how the text itself is phrased.
That is both the strength and the weakness of textGrain.
As long as the text remains largely unchanged, the statistical pattern can be relatively clear. The more heavily people or other AI systems rewrite it, the more that pattern fades.
textGrain is therefore unlikely to become the definitive “AI detector”. It is better understood as another provenance signal within an increasingly complex system of watermarks, metadata, visible labels and organisational documentation.
And that distinction is likely to become increasingly important: a watermark may indicate that AI was involved. It cannot tell us who ultimately wrote, shaped or took responsibility for the text.
