MENU
CLOSE
The Law Platform
© Copyright 2021 - 2026
The debate over how to identify artificial intelligence-generated content has largely focused on images, videos and deepfakes. But as generative AI becomes increasingly capable of producing essays, reports, legal documents and other forms of written communication, how can society now determine whether a piece of text was generated by AI?
On October 5, 2026, OpenAI announced its approach to text provenance, introducing an invisible watermarking system for text generated by eligible versions of ChatGPT and Codex in the European Union. The move is part of the company's response to transparency requirements under the EU AI Act (Regulation (EU) 2024/1689). The technology, called textGrain, is designed to place a machine-readable statistical signal within AI-generated text without visibly changing the words presented to the user.
What Is Text Provenance?
Text provenance is about the ability to establish where written content came from and, to some extent, how it was produced or modified. This has become very significant as generative AI tools have moved from experimental technology into education, journalism, business, government and professional services.
A person reading a polished article may not know whether it was written entirely by a human, generated by an AI system, or produced through a combination of human and AI input. OpenAI's new system seeks to provide an additional signal.
According to OpenAI, textGrain works by subtly influencing the statistical pattern of word choices produced by its models. A detector can subsequently analyse a passage and assess whether it contains the watermark. Unlike a visible label saying "This text was generated by AI," the watermark is designed to be invisible to ordinary readers.
The EU AI Act Is Driving the Change
The announcement is closely connected to the European Union's comprehensive effort to regulate artificial intelligence. OpenAI says the EU AI Act requires providers of generative AI systems to make generated text identifiable in a machine-readable way. Its response is therefore not simply a technological experiment; it is also part of the company's effort to comply with emerging AI regulation.
Under the new approach, eligible ChatGPT and Codex outputs in the EU will receive invisible watermarks as the system rolls out over the coming weeks.
The company is initially limiting the rollout geographically instead of making watermarking a global default. For API customers, however, OpenAI is taking a different approach. Customers around the world will be able to opt in to watermarking for selected models, while the feature will remain off by default in the API.
24th Sep, 2026
15th Sep, 2026
22nd Aug, 2026
29th Jul, 2026
13th Jul, 2026
But Can AI Watermarks Actually Detect AI Text?
OpenAI says its testing showed textGrain performing as well as or better than other approaches it evaluated, including Google's text version of SynthID. But the company acknowledges that strong laboratory performance does not mean the technology will always work reliably in the real world. For example, the detector's effectiveness depends partly on the length and nature of the text.
At a targeted false-positive rate of 1%, OpenAI reported detecting its watermark in approximately 80% of 200-token psychology passages and about 95% of 400-token passages. The results were considerably weaker for mathematics, where there is less flexibility in word choice.Editing creates another major vulnerability. In OpenAI's testing, replacing just 10% of words in a 400-token passage with synonyms reduced watermark detection from approximately 92% to 66%. Replacing 25% of the words reduced detection to just 17%. AI-generated text is rarely left untouched in many real-world situations.
Users may edit, paraphrase, translate, shorten or combine AI-generated material with their own writing.
In other words, the watermark is not necessarily permanent once the text enters the hands of a human editor.
A Watermark Is Not Proof of AI Authorship
Perhaps the most important part of OpenAI's announcement is what it says the technology cannot establish. A detected watermark does not reveal how much of a passage was produced by AI and how much resulted from human work.
It does not establish who owns the text.
It does not determine whether the use of AI was lawful.
It does not identify the person who generated the content.
And it does not establish whether the information contained in the text is accurate.
The reverse is equally important. If a detector fails to identify a watermark, that does not prove that the text was written by a human.
The watermark could have been weakened through editing or translation. The text might also have been generated before watermarking was introduced, produced using an unsupported model, or generated by another company's AI system.
Source: OpenAI
Unlock premium articles, PDF downloads, cases and more
Learn how to subscribe