OpenAI is implementing invisible watermarks in text generated by ChatGPT and Codex for European Union users. The technology, called textGrain, subtly modifies word choices to create a detectable statistical pattern that remains hidden when text is read or copied. Starting today, API developers worldwide can voluntarily enable watermarking for compatible models, though it is currently disabled by default. OpenAI is accepting applications for access to its watermark detector, with initial access limited to approved researchers and expert organizations.
However, OpenAI acknowledges significant limitations with its watermarking approach. Testing reveals that ordinary text editing substantially reduces detection effectiveness—replacing just 10 percent of words with synonyms lowered detection rates from 92 percent to 66 percent. Detection performance also varies by subject matter, performing worse in mathematics where models have fewer word choice alternatives. OpenAI cautions that the absence of a detected watermark does not confirm human authorship, and that text may be too brief, edited, or translated for reliable detection. The company states that detected watermarks reveal neither the creator’s identity nor account information.
Source: Read the full story at BleepingComputer
This summary was written with AI assistance from the linked source.
Send this to yourself, or to someone who should see it
Enter your email and we will send you a link to this article.
