OpenAI has announced plans to introduce invisible watermarks on eligible ChatGPT and Codex text generated in the European Union as part of its response to the EU’s artificial intelligence rules.
The company said the watermarking system will be introduced over the coming weeks, while developers using OpenAI’s API globally will be able to opt in to watermarking for selected models from Monday.
OpenAI said the move is aimed at improving transparency around AI-generated content and meeting requirements under the EU AI Act, which calls for AI-generated text to be identifiable in a machine-readable form.
The watermarking technology, called textGrain, works by adding an invisible statistical signal to the way an AI model selects words. OpenAI’s detector then analyses a passage to determine whether the signal is present.
The company said its tests showed textGrain performed as well as or better than other watermarking approaches it evaluated, including Google DeepMind’s SynthID for text. However, OpenAI stressed that the technology is still developing and may not reliably identify every piece of AI-generated writing.
According to OpenAI, longer passages are generally easier to detect than shorter ones, while texts such as mathematics, where there is less flexibility in word choice, present greater challenges.
Editing can also significantly weaken the watermark. OpenAI said replacing 10 per cent of the words in a 400-token passage reduced detection from about 92 per cent to 66 per cent, while replacing 25 per cent reduced detection to 17 per cent.
Because of these limitations, OpenAI said access to its text-watermark detector will initially be restricted to approved researchers and expert organisations rather than being made publicly available.
The company said researchers and organisations can apply for access, with approvals initially being handled on a case-by-case basis to help assess and improve the technology.
See also: Morocco Unveils AI Tools That Understand Local Arabic Dialects
OpenAI also stressed that detecting a watermark would provide only a limited indication of the role its systems played in producing a passage.
A positive result would not show how much human input went into the text, determine who owns it or establish who is responsible for it. It would also not verify whether the content is accurate or identify the user, account, prompt or conversation associated with the text.
Similarly, OpenAI said failure to detect a watermark would not prove that a person wrote the material. Text may evade detection because it is too short, heavily edited or translated, produced by an unsupported model or generated before watermarking was introduced.
The company said it would keep the technology off by default for API customers, who will instead have the option to activate it for selected models. OpenAI is also working with cloud partners to extend watermarking to its model outputs accessed through their services.
The move forms part of OpenAI’s broader approach to content provenance, which already includes Content Credentials and invisible watermarks for supported images and audio.
OpenAI said it plans to continue testing how text watermarks perform after editing and translation and to explore ways of distinguishing between AI assistance and full AI authorship as the technology and regulatory standards develop.
About the Author
Cecilia Attah is a tech analyst with a degree from Benue State University. She covers tech news and startups at TechRegard with a focus on how technology is transforming Africa and shaping the global landscape.
