Researchers say they've built a watermarking system for AI-generated text that survives one of the most common evasion moves: rewording entire paragraphs at once.
Previous semantic watermarking schemes tagged text at the sentence level. That worked against minor edits, but fell apart once an attacker scrambled sentence order or rewrote whole paragraphs. The new framework, SAMark, sidesteps that weakness by anchoring watermarks to a position-independent region in semantic space rather than to sentence sequence. It adds a multi-channel scoring mechanism that amplifies the watermark signal and a filtering layer that catches semantic redundancy beyond the usual n-gram tricks. In tests, SAMark hit 90.2% true-positive rate at a 1% false-positive threshold under paragraph-level paraphrasing, outperforming the prior best method by more than 30 percentage points on average.
Watermarking AI output is the simplest proposed answer to a real disclosure problem: once text leaves a model, how does anyone confirm it came from one? Most schemes so far have been defeatable with a single paraphrase pass, so holding 90% detection under that attack is a meaningful bar. The researchers also claim to break the usual robustness-quality trade-off, where stronger watermarks degrade the readability of generated text, though that claim deserves scrutiny at scale.
The work is a preprint on arXiv, not yet formally peer-reviewed, and how it holds up against commercial paraphrasers in the wild is still an open question.