Writing assistants can watermark the words they suggest. In an audit, when a school or publisher asks the provider of such an assistant about a submitted report, the provider checks how many positions of a passage match tokens its service returned, without having retained request logs. A watermark test shows that such matches exist, not how many. We introduce CountScan, a scan that turns per-token watermark evidence into lower bounds on the number of matches, valid for every passage at once and after key-blind editing by the writer, at a cost logarithmic in the text length. Since different hidden histories can release the same text with different numbers of matches, no valid count can certify more than they all share. In explicit models of the writing interface, in which the writer may request, retry and discard suggestions, retain their own draft and edit under a request cap, we compute this limit exactly and show that a variant of CountScan attains it. In simulations it stays valid and, under a hard history, certifies at least 2.5% of the positions in 100% of documents with 10^4 positions.
翻译:暂无翻译