OpenAI is preparing to add invisible watermarks to text generated by ChatGPT in the European Union, giving certain organizations a way to detect whether text was produced by its AI models.
The company announced the change as part of its response to the EU AI Act, which requires generative AI providers to make AI-generated text identifiable in a machine-readable format.
Watermarking will roll out to eligible ChatGPT and Codex text outputs in the EU over the coming weeks. Unlike a visible label attached to an AI response, its system is designed to leave a statistical signal within the text itself that cannot be seen by readers.
How ChatGPT’s invisible text watermark works
OpenAI calls its watermarking technology “textGrain.” Rather than inserting hidden characters, unusual punctuation, invisible spaces, or additional tokens into a response, textGrain subtly changes how the model chooses between possible words and pieces of words while generating text.
Those choices create a pattern across a passage that the company’s detector can look for later. Because the watermark is built into the wording, simply copying and pasting an unchanged response is not expected to remove it.
The watermark also does not contain information about the person who generated the text, their organization, or the prompt they entered.
Unsplash: glenncarstenspetersOpenAI says API customers around the world can also opt into text watermarking for supported models, although the feature will remain disabled by default. The company is also working with cloud and distribution partners to bring provenance signals to eligible outputs delivered through their services.
The company does not currently plan to give everyone access to the detector used to identify watermarked text. Instead, applications are being opened to approved researchers and expert organizations as the company evaluates the technology’s reliability and potential uses.
The company says this is partly because detecting watermarked text is not foolproof. A detector can incorrectly identify text as containing a watermark or fail to recognize text that actually contains one.
Short responses are particularly difficult to detect because they may not contain enough text to establish the statistical pattern. Code, math answers, and highly factual responses can also be harder to watermark because the model has fewer reasonable ways to phrase its output.
The EU’s transparency rules account for some of those limitations. Under its Code of Practice for AI-generated content, outputs shorter than 200 tokens, roughly 150 English words, and code snippets are not required to carry a watermark.
Editing can also weaken the signal. They said that substantial rewriting, paraphrasing, or translating a response can make its watermark more difficult or impossible to detect.
Freepik/DexertoThat means finding a watermark can indicate that a system generated or processed text, but it cannot determine how much of the finished work was written or edited by a person. Likewise, failing to find a watermark does not establish that something was written by a human.
OpenAI says its testing found no meaningful reduction in model quality from enabling textGrain, while the effect on generation speed was negligible.
The company already uses other provenance systems for AI-generated media.
Images created using supported tools include Content Credentials and SynthID watermarks, while supported AI-generated audio uses SynthID. Unlike the new text system, the company’s verification tools for supported images and audio remain publicly accessible.
This also marks a shift from OpenAI’s previous attempts to identify AI-written text. In 2023, the company shut down its own AI classifier after admitting it suffered from a “low rate of accuracy,” with the tool correctly identifying AI-written text just 26% of the time at launch.
At the time, OpenAI said it was researching “more effective provenance techniques for text.” More than three years later, textGrain represents the company’s latest attempt to make AI-generated writing identifiable without relying solely on traditional AI detection tools.
I’m Abhishek Sharma, an author at TigerJek.com. I enjoy exploring games, testing different strategies, and turning what I learn into clear, useful guides. My goal is to help players understand the game better and improve without the usual confusion.
