Watermarking AI-generated text is a promising tool for tracing provenance and detecting edits. A broad class of these methods, known as logit-based watermarks, operate by increasing the logits of a chosen token set at each step of generation. But not all language models may be equally receptive to this kind of manipulation.
The new arXiv paper proposes a concept called 'watermarkability' to measure how compatible a given model is with logit-based watermarks. Rather than treating watermarking as a one-size-fits-all procedure, the authors frame it as a property of the model itself. This allows for systematic comparison across models and watermarking strategies.
Beyond measurement, the paper also addresses how to improve watermarkability. By identifying factors that make models more or less amenable to logit adjustments, it points toward ways to design or fine-tune models that are easier to watermark effectively. The work is based on a single source, so the details of the proposed metric and improvement techniques are not yet fully specified here, but the direction is clear: making watermarking more predictable and robust.