Will People in Europe No Longer Be Able to Write with ChatGPT?
OpenAI is preparing to add a machine-readable, invisible watermark to text generated by ChatGPT and Codex. Known as textGrain, this technology is scheduled to roll out first for European Union (EU) users in the coming weeks as part of OpenAI's steps to comply with the AI-generated content transparency requirements of the EU AI Act.
Here, a "watermark" does not refer to a visible logo in the corner of a document or a hidden control character. Instead, textGrain slightly alters the statistical patterns behind the model's word choices and token selections as it generates sentences. While this pattern remains completely invisible to a human reader, a specialized detector can identify it. Because it is embedded directly within the wording pattern itself rather than relying on hidden spaces or special characters, the watermark stays intact even after copying and pasting.
Can All AI-Written Text Be Detected 100% of the Time?
No. OpenAI itself explicitly states that text watermark detection technology is far from perfect.
Detection reliability decreases significantly with short text, constrained answers, heavily edited content, and certain languages. Detectors carry both a risk of false negatives (missing an AI watermark that is present) and false positives (incorrectly flagging human-written text as having a watermark).
For instance, TechCrunch reported on OpenAI's internal testing: while detection rates were high for original watermarked text, replacing just 10% of the words with synonyms led to a significant drop in detection accuracy. Therefore, one cannot draw the definitive conclusion that "No watermark detected = Human-written."
What Information Does This Watermark Reveal?
The detector can assess whether an OpenAI watermark signal exists within a piece of text. However, it cannot:
- Identify the individual user who generated the text
- Reveal the original user prompt
- Prove whether the text is factually accurate
- Verify copyright ownership
- Determine the exact extent of human editing applied to the text
OpenAI has explicitly highlighted these key limitations.
Can Anyone Check for This Watermark?
Not currently. OpenAI plans to restrict initial access to its text watermark detector to approved researchers and expert organizations. The detection tool has not been launched for public access, as OpenAI emphasizes the need for responsible evaluation due to the risks of false positives and false negatives.
