Text AI watermarks will always be trivial to remove

An article argues that text-based AI watermarks, such as those proposed by OpenAI and other developers to detect AI-generated content, are fundamentally ineffective because they can be easily removed or bypassed. The piece explains that watermarking relies on subtle statistical patterns in token selection, but these patterns can be disrupted by simple post-processing techniques like paraphrasing, retranslation, or adding noise. The author contends that as long as AI-generated text is indistinguishable from human-written text in quality, any watermarking scheme will be trivial to circumvent, making it an unreliable tool for content authentication. The article cites technical limitations and the adversarial nature of the problem, noting that even sophisticated watermarks can be stripped without significant effort. This critique comes amid growing regulatory and industry interest in watermarking as a solution to misinformation and academic dishonesty.

Global Impact

The article's argument has significant implications for the ongoing debate around AI regulation and content authenticity. If watermarks are indeed trivial to remove, then proposed policies requiring AI-generated text to be watermarked (e.g., the EU AI Act or US executive orders) may be ineffective, potentially eroding public trust in digital content.