These days, it's getting hard to tell apart what's AI-written and what isn't; we are in an era of AI slop, but it looks like OpenAI is making a move to identify text which is AI-generated by watermarking text generated by ChatGPT and Codex. Invisible to humans, these watermarks, OpenAI says, will be detectable by their detection tools. Currently, the company plans to release this feature in the EU only and further global expansion is not on the table for now. But is it just about limiting AI slop, or is there something else behind this move?
What Is This System Called and How Will It Work?
The company calls this system textGrain, and it tweaks the pattern of words used by the model, a pattern which will remain out of human sight but can be detected by AI detectors. This text can also be detected even if it is copied elsewhere, since it works on the very words generated by AI.
ALSO READ: Vivo V80 Marks India Debut: Snapdragon Chipset, 50-MP Zeiss Cameras And 7,200mAh Battery Inside
This is similar to a system employed in Google DeepMind's SynthID for text, and Anthropic's Claude is also likely using a similar system, which it released in August. OpenAI's watermark leaves no personal footprint. It stores zero data regarding the user or the original prompt, acting purely as a simple "yes or no" signal for the detector.
But Why Is OpenAI Adding Watermarks?
OpenAI is bringing its text watermarking system, textGrain, to the European Union as it prepares for new requirements under the EU AI Act. The law requires providers of generative AI systems to make AI-generated text identifiable in a machine-readable format.
The company plans to start the EU rollout in the coming weeks. OpenAI says it is keeping the initial deployment limited so it can see how textGrain works outside controlled testing, while also collecting feedback before making a decision on a wider launch.
There is no need for users outside the EU to wait, either. OpenAI is making the watermarking system available as an opt-in feature to API customers worldwide from today, although it will initially work with only select models. Researchers and expert organisations will also be given access to OpenAI's detection tool as the company continues to develop the technology.
Not A Foolproof Solution
But OpenAI is not presenting textGrain as a foolproof way of finding AI-written content. The company itself points out that detection becomes more difficult when there is less text to analyse or when the writing leaves little room for different word choices.
Its own testing shows the difference. The watermark was picked up in roughly 80 per cent of 200-token psychology passages. When the passages were twice as long, detection climbed to about 95 per cent. The results were considerably weaker for topics such as mathematics, where writers tend to have fewer ways of expressing the same information.
Changing the wording can also make a big difference. In one test using 400-token passages, replacing just 10 per cent of the words with synonyms brought detection down from around 92 per cent to 66 per cent. Once 25 per cent of the words were replaced, the figure fell sharply to about 17 per cent.
Translation is another problem for the system. OpenAI says text that has been translated can also become harder for the detector to recognise.
There is also a common misconception that a watermark could somehow tell whether an AI-generated passage is trustworthy. That is not what textGrain is designed to do. The watermark does not check whether the information in a passage is correct, nor does it indicate how much of the writing came from a human.
The reverse is also important. If textGrain fails to find a watermark, that does not mean the content was necessarily written by a human. Changes made during editing, translation, the use of older AI models or other AI tools can all affect the result.
For now, OpenAI's approach is different depending on where the user is. The EU will get textGrain as the default rollout in response to the region's AI rules, while API customers in other parts of the world can choose to use it on supported models.
The move also highlights a bigger question for the AI industry. As AI-generated writing becomes increasingly common, companies are looking for practical ways to identify it. Text watermarking could become one of those methods, but OpenAI's own results show that making it reliable across every type of writing will be far from straightforward.