OpenAI is planning a significant change for users in the European Union. Texts generated through ChatGPT and Codex will, in the coming weeks, contain an invisible digital watermark that aims to allow machines to recognize that artificial intelligence was involved in their creation.
The company OpenAI announced this new system on Monday, October 5th. The technology, called textGrain, is intended to be gradually implemented into selected text outputs from ChatGPT and Codex across the entire European Union, and will apply to users of all subscription tiers. One reason for this is the new requirements of the European AI Act.
However, this is not a classic watermark that a user would see on their screen. The text will not contain any warnings or special symbols. textGrain works differently – it subtly influences the selection of individual words when generating a response and creates a statistical pattern in the text that can then be recognized by a special detector.
The watermark remains part of the text even after it has been copied, for example, into a document, email, or website.
The Watermark Has Its Limitations
OpenAI also points out that the technology is definitely not 100% reliable. The shorter or more modified the text is, the more difficult it is to reliably determine its origin.
In one test, the system was able to recognize the watermark in approximately 92 percent of cases for an original text consisting of several hundred tokens. However, after replacing 10 percent of the words with synonyms, the success rate dropped to approximately 66 percent. If 25 percent of the words were changed, the detector only recognized the mark in approximately 17 percent of cases.
Very short responses, mathematical texts, or translations are also problematic. Therefore, OpenAI emphasizes that the absence of a watermark will definitely not be proof that the text was written by a human.
Conversely, the same applies to the opposite situation. The presence of a watermark only indicates that the text may have been created or processed by an OpenAI system. It says nothing about how much work a person put into it, whether it was subsequently edited, or who specifically created it.
Users Will Not Be Tracked
An important issue is privacy. OpenAI states that the watermark will not contain information about a specific user, account, prompt used, or the conversation itself.
Furthermore, the detector will initially not be available to the public. Approved researchers and organizations will first gain access to it, which will help evaluate its reliability and potential uses.
For customers using the API interface, the situation will be different. They can voluntarily enable the text watermark for supported models worldwide. However, by default, it will remain disabled.
This change will initially affect primarily regular users of ChatGPT and Codex in the European Union. OpenAI plans to gradually implement it over the coming weeks and also acknowledges that the method of marking texts will continue to evolve based on experience from real-world use.
Jaroslav Holý
