People worry that watermarks, hidden signals that text came from artificial intelligence, let companies track what they write. Users often paraphrase to remove the watermark, making it useless.
A new approach protects privacy by only measuring how many AI-written texts exist overall, not which ones. It adds random noise to each text's signal so a single text reveals very little, called local .
Tests on an essay task with only a handful of samples showed the method is accurate. Its estimates of the fraction of watermarked texts were off by only 0.0113 on average and had a bias of only 0.004. This means AI use can be monitored at scale without invading privacy.