Saifur Rahman Tamim, Amir Labib Khan
Several governments have decided that text written by a language model should carry a watermark. The EU AI Act asks for markings that are sufficiently reliable and robust. California's SB 942 wants disclosure that is permanent or extraordinarily difficult to remove. Both are written as if the technical question were settled.
Underneath sits an assumption nobody had tested: that when a watermark detector fires, the result is solid enough to be treated as evidence. Those are very different standards. A detector that is usually right is a useful research tool and a dangerous basis for a legal finding, because the cases that matter are the contested ones.
This paper tests the assumption directly, evaluating three representative watermarking methods including KGW, Unigram, and the MarkLLM implementation of SynthID-Text. The framing is what makes it worth reading: it treats forensic readiness as a property to be demonstrated rather than assumed, at a moment when law is already being written on the assumption.
Governments are increasingly mandating that LLM-generated content carry watermarks. The EU AI Act calls for markings that are "sufficiently reliable and robust." California's SB 942 requires disclosure that is "permanent or extraordinarily difficult to remove." Both mandates rest on an untested assumption: that watermark detection yields evidence reliable enough for courts. This paper tests that assumption directly. We evaluate three representative LLM watermarking methods…
More with Less: a Large Scale Remote Sensing VLM with a Simple Recipe
arXiv (cs.CV) · July 17, 2026Handwritten and Printed Text Segmentation via Region-Aware Human-Writing Descriptor Engineering
arXiv (cs.LG) · July 17, 2026Learning Reach-Avoid Task with Reinforcement Learning: Vectorized Simulation and Benchmark
arXiv (cs.CV) · July 16, 2026Hierarchical Denoising For Multi-Step Visual Reasoning