AI-generated text is everywhere, raising concerns about authenticity and misuse. One promising solution is watermarking, subtly embedding invisible signals into the text during its creation. This allows for later verification, confirming whether a piece of text originated from a specific AI. However, a major challenge arises when humans edit AI-generated text. Does this editing process destroy the watermarks, rendering them useless? New research tackles this critical question, exploring how robust current watermarking techniques are against human meddling. Researchers analyzed the popular Gumbel-max watermarking method, simulating various scenarios where humans edit AI text. Surprisingly, even simple edits like replacing a few words can significantly weaken watermark signals. This vulnerability arises because the watermark relies on the statistical relationships between words and hidden pseudorandom numbers. When words are changed, these relationships are disrupted, making the watermark harder to detect. To address this, the researchers developed a new detection method called Tr-GoF. This approach uses a statistical technique that compares the distribution of word probabilities in the edited text to what would be expected in human-written text. Tr-GoF cleverly truncates or removes extreme values in its calculations, making it more stable and robust against the noise introduced by human edits. Results show that Tr-GoF outperforms existing methods, especially when dealing with targeted “adversarial” edits designed to erase the watermark. It’s particularly effective when the AI’s output is less random, like when it’s reciting a poem rather than generating a new one. This research is a vital step towards ensuring AI watermarks can hold up in the real world. While Tr-GoF shows promise, the ongoing challenge is to make watermarks even more resilient against a wide range of human editing behaviors, while also preserving the natural flow and quality of AI-generated content.