Ideal Attribution and Faithful Watermarks for Language Models

We introduce ideal attribution mechanisms, a formal abstraction for reasoning about attribution decisions over strings. At the core of this abstraction lies the ledger, an append-only log of the prompt-response interaction history between a model and its user. Each mechanism produces deterministic decisions based on the ledger and an explicit selection criterion, making it well-suited to serve as a ground truth for attribution. We frame the design goal of watermarking schemes as faithful representation of ideal attribution mechanisms. This novel perspective brings conceptual clarity, replacing piecemeal probabilistic statements with a unified language for stating the guarantees of each scheme. It also enables precise reasoning about desiderata for future watermarking schemes, even when no current construction achieves them, as the ideal functionalities are specified first. In this way, the framework provides a roadmap that clarifies which guarantees are attainable in an idealized setting and worth pursuing in practice.

Large language models now routinely produce human-quality text; for the first time in history, vast amounts of writing are coming from non-human sources.Large language models now routinely produce human-quality text; for the first time in history, vast amounts of writing are coming from non-human sources. This makes it important to be able to trace text back to the model that produced it, and watermarking, which embeds a hidden statistical signal in AI-generated text so it can be recognized later, is a leading tool for the job. But pinning down what it actually means for a piece of text to be attributable to a given model turns out to be conceptually difficult, and different methods provide different, hard-to-compare guarantees. We define an ideal notion of when text should be attributed to a particular model. A watermarking method is then judged by how closely its detector matches this benchmark. This separates the goal from the implementation, making it easier to state, compare, and understand different watermarking guarantees.