A ChatGPT text watermark could soon leave a trace that readers cannot see. OpenAI has described a plan to place an invisible statistical signal into eligible text produced by ChatGPT and Codex in the European Union, as part of its response to new transparency requirements. It is an attempt to address a difficult question: how can anyone assess whether a passage involved AI when the words look entirely ordinary?
OpenAI set out the approach on 5 October 2026 in its explanation of EU text provenance. It said the European rollout would take place over the following weeks, while selected API customers worldwide could opt in for supported models. The announcement did not make watermarking an immediate global default. It also did not provide every member of the public with a tool capable of testing arbitrary text.
How an invisible ChatGPT text watermark works
The approach is called textGrain. Instead of adding a visible label at the bottom of a paragraph, it subtly affects patterns in the words a model chooses. A detector can examine a passage for the expected statistical signal and report whether it appears to be present. For ordinary readers, the words can look the same as any other piece of writing.
This is different from placing a creator’s name in the document metadata or attaching a warning to a webpage. Metadata may disappear when material is copied. Visible labels can be removed from a screenshot or quotation. A statistical watermark is intended to survive at least some ordinary changes to the way a passage is shared, though its survival is never guaranteed.
The system also has an important limitation: it relies on the text having been generated in a way that includes the signal. A detector cannot reliably determine the origin of every sentence on the internet. A passage produced by an older system, another provider, an excluded output category or a person would not necessarily carry this particular pattern. Absence of a signal is not evidence of human authorship.
Why detection can go wrong
OpenAI’s own tests show that detection is easier for some types and lengths of writing than for others. In one reported evaluation targeting a false-positive rate of 1%, the detector identified signals in about 80% of certain 200-token passages and roughly 95% of 400-token passages. Mathematical writing was harder. These are specific internal test results, not a universal accuracy rating for every language and subject.
Editing can weaken the trace too. OpenAI reported that, in one evaluation using 400-token passages, replacing a quarter of the words with synonyms reduced detection substantially. That makes intuitive sense: the watermark relies on patterns among word choices, so replacing many choices may erase the pattern. Translation, shortening and rewriting can create similar practical difficulties.
A false positive is especially serious. It means a detector might report a signal when the passage does not contain the intended watermark. If that result were treated as proof that a pupil cheated, an employee lied or an author misrepresented their work, the consequences could be unfair. LiveAIWire has examined the problems caused by false positives in AI essay detectors, where a confident label can be mistaken for reliable evidence. The technical tool must therefore be evaluated alongside the purpose for which someone wants to use it.
OpenAI is initially limiting detector access to approved researchers and expert organisations. The company says this is partly to support responsible interpretation and further evaluation. That restriction does not establish that the tool is free of errors when used by experts; it simply narrows the first group able to examine the results and provide feedback.
What the EU is asking AI providers to do
The legal background is broader than one company’s technology. The European Commission’s guidance on Article 50 of the AI Act describes duties for providers of generative systems to make certain outputs machine-readable and detectable as artificially generated or manipulated, subject to defined conditions and exceptions. Those provider duties are not the same as every person being required to publish a prominent warning on every piece of AI-assisted writing.
The Commission also distinguishes the role of a provider from that of a person or organisation deploying an AI system. Particular public-interest text may trigger disclosure questions, but human review and editorial responsibility are relevant to the rules. A publisher cannot understand its obligations merely by asking whether a watermark is detectable. How the material was prepared, reviewed and published also matters.
LiveAIWire has previously explained the wider EU labelling framework for chatbots and synthetic media. The OpenAI announcement is a more specific development about the technical marking of eligible text. It does not mean that all synthetic images, voices and written articles are subject to one identical labelling process.
A watermark cannot tell you who wrote the piece
A passage can be drafted by a system, heavily rewritten by a person, quoted from elsewhere or used as one small element in a larger piece of original work. Finding a signal does not explain how much human judgement went into the finished result. Nor does it establish ownership, responsibility or whether a publication was lawful. Those are separate questions involving context and, sometimes, legal analysis.
Likewise, a detected watermark cannot prove that the content is false. An AI system can state a true fact, and a person can write something misleading without AI assistance. The reliability of a claim still depends on evidence, sourcing and editorial checks. Provenance can help explain where a passage came from without certifying whether it deserves belief.
This distinction matters in education and employment. A teacher may reasonably want to know whether a pupil has done the assigned thinking. A manager may want to know whether confidential information has been placed into an outside system. A watermark result might offer a limited clue about one component of a workflow, but it cannot answer all those questions or fairly replace an investigation.
Different tools leave different kinds of traces
OpenAI says it already uses other provenance techniques for supported image and audio outputs, including content credentials and invisible marks. Those mechanisms work differently from statistical marking of prose. A picture can carry structured metadata or an embedded signal that is not the same as a sequence of word choices. Conflating these methods creates unrealistic expectations about what each detector can reveal.
Other companies have investigated related techniques. LiveAIWire reported on a separate debate about watermarking Claude-generated text. The broader issue is shared, but one company’s specific approach and rollout should not be presented as a universal industry standard. Detection across different models, languages and publication methods remains a significant challenge.
There is also a tension between durability and usefulness. A watermark strong enough to survive extensive editing might affect the freedom of expression or quality of output, while a watermark that preserves natural writing may become easier to remove. OpenAI says its internal evaluations did not find meaningful overall performance differences in the benchmarks it tested, but everyday creative and editorial effects still require observation.
What readers should do with a detection result
For a journalist or editor, a text-origin signal is only one part of accountability. Reliable publication still requires checking names, dates, original documents, context and whether the final story reflects an informed editorial decision. A machine-readable watermark cannot record that judgement. Conversely, a newsroom can be responsible for its published work even when software assisted with a limited step. The distinction is why provenance and editorial verification should be treated as complementary activities rather than competing definitions of authorship.
For readers, the practical principle is to treat provenance as evidence with limits. A positive result can suggest that a particular kind of AI generation or processing occurred. It does not name the user, identify a prompt or reveal a conversation. A negative result cannot certify that a human wrote the passage. And neither outcome validates the factual claims inside the text.
Organisations considering use of such a detector should define what question they are trying to answer and what decisions the result may influence. A positive result might warrant a conversation about drafting methods, but it should never substitute for a fair review of the work, the underlying evidence and the rules that were actually in force when the text was produced. They should allow for uncertain cases, protect legitimate privacy interests and avoid automated penalties based on a statistical guess. The standard of evidence must rise when a false conclusion could harm someone.
The European rules create an incentive to improve the traceability of synthetic content, and OpenAI’s October announcement offers one specific response. Whether invisible text watermarking proves useful depends on accuracy across real material, resistance to routine editing and responsible interpretation by those who use it. The most credible promise is a limited additional clue about origin, not a machine that can definitively tell human writing from AI.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.