OpenAI is following in Anthropic’s footsteps by introducing an invisible watermark for text generated by ChatGPT and Codex across Europe, as AI companies face new requirements to make AI-generated content identifiable.
The move comes after its competitor Anthropic introduced invisible watermarking for text generated by its Claude models in August.
OpenAI will introduce the technology to eligible ChatGPT and Codex users across all plans in the EU over the coming weeks, while developers using OpenAI’s API anywhere in the world can switch it on for select models from today. It will be off by default for API users.
The company is not making text watermarking a global default at launch, saying the regional rollout will give it room to “learn from real-world use and feedback” in a blog post on Tuesday.
The watermark itself is not a visible symbol or label. Instead, OpenAI’s system subtly influences the model’s word choices as it generates text, creating an imperceptible pattern that can later be detected by a dedicated system.
OpenAI said the watermark does not identify the user and that it saw no meaningful change in model performance when the technology was switched on.
However, the company acknowledged that the watermark can be weakened through editing.
“Editing can substantially weaken the watermark signal,” OpenAI said in its announcement.
In one test of 400-token passages (see chart below) replacing 10 per cent of words with synonyms reduced watermark detection from about 92 per cent to 66 per cent. Meanwhile, replacing 25 per cent of words reduced detection to 17 per cent.
OpenAI also found that shorter or more constrained passages are harder to detect. Detection rates were also substantially lower for content such as mathematics, where there is less flexibility in word choice.
OpenAI has published a technical report alongside the announcement detailing the method, called textGrain, which was developed with researchers from the University of Pennsylvania and Yale.
The system uses a secret key to influence the ranking of possible next words as the model generates a response. While each individual word choice appears normal, hundreds of these small changes can create a statistical pattern that a detector can identify using the text and the key.
Because the watermark is embedded within the words themselves, OpenAI said it travels with text when it is copied and pasted.
However, the system is not designed to guarantee that AI-generated text will remain detectable after significant editing, with OpenAI noting that translation, shorter passages and mathematical responses can all make detection more difficult.
The rollout follows Anthropic’s decision in August to introduce a similar system for Claude.
“When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself,” Anthropic said at the time, adding that users would not see it and that it would not change the meaning, quality or readability of Claude’s response.
OpenAI is the latest major tech company to commit to EU’s AI code, with others including Anthropic, Google, Meta and Microsoft.
