ChatGPT users in the European Union will soon get an invisible watermark in eligible text responses. OpenAI plans to introduce the technology across all ChatGPT and Codex plans in the EU over the coming weeks, raising questions about how people will interpret the results.
OpenAI announced the rollout in response to the EU AI Act. API customers worldwide can also opt in for select models starting today, but watermarking remains off by default in the API. The company is not introducing it as a global default at launch.
The technology, called textGrain, creates a statistical pattern through the model’s word choices. According to OpenAI’s technical explanation, it adds no hidden characters, invisible spaces, or unusual punctuation. Because the signal exists in the wording, copying and pasting does not automatically remove it.
Detection has substantial limitations, however. In one evaluation of 400-token passages, replacing 10 percent of words with synonyms lowered detection from about 92 percent to 66 percent. Replacing 25 percent dropped it to 17 percent. OpenAI will initially limit detector access to approved researchers and expert organizations.
A watermark also cannot measure human contribution, establish ownership, identify the user, or verify accuracy. Its absence does not prove that a human wrote the text. That leaves considerable room for people to draw conclusions the technology cannot support.
Consider someone who writes an original essay and asks ChatGPT to improve its grammar. Compare that with someone who asks the chatbot to produce the entire assignment. A positive detection result cannot explain the difference, even though the amount of human effort could be vastly different.
My concern is that a teacher, employer, or reader could treat that result as a verdict. Someone who developed the ideas, checked the facts, and shaped the argument could find their work dismissed because they used an editing tool. That is a possible consequence, not something this announcement establishes has happened.
I appreciate the EU’s apparent focus on protecting its citizens, but in my view, its regulatory overreach usually leads to negative consequences. Requiring a technical signal does not ensure that people understand it. A transparency measure could become a source of unfair suspicion if users overlook its limitations.
The weakness under editing creates another problem. Someone who makes modest use of AI could retain a detectable signal, while extensive rewriting could make it harder to find. I question how useful that will be for anyone hoping to judge the extent of AI involvement.
I am glad to be an American, and I am relieved this default rollout does not apply to my U.S. ChatGPT usage. I would prefer that decisions about writing tools leave room for individual judgment, particularly when a detector cannot account for the work a person contributed.
OpenAI is right to explain these limitations upfront. The challenge is ensuring that anyone using its detector takes them seriously, because an invisible pattern in someone’s words should not outweigh a fair assessment of their work.
Support independent tech journalism
NERDS.xyz is independently owned and operated. If you enjoy my coverage of Linux, AI, hardware, cybersecurity, and tech culture, consider supporting the site on Ko-fi.
Support NERDS.xyz
