Expand the digital commons: human-curated and AI-generated defensive publication

This is the authors' abstract. We don't add key points for this paper.

Language is the substrate of the patent system, and that substrate is arithmetizable. The claim is the invention's true text, and what can be written can be enumerated, varied, and generated. First, we argue that the patent system's foundations are empirically failing on their own evidence: search does not find, the inventive step does not discriminate, and classification silos conceal from the examiner the analogical transfers that drive real invention. Second, we argue that doctrine itself supplies the remedy's mechanism. Prior art turns on accessibility and enablement, not on authorship or readership: hidden but published is doctrine. This creates a decisive asymmetry: machines cannot invent, but they can anticipate. Third, we argue that generative models, today devoted almost entirely to the production of exclusivity, can be turned toward the opposite end, the digital commons: distributed, timestamped, enabling defensive publication at scale, with the human contribution quantified and placed on record. The service txtcraft proposes a concrete implementation. Copyleft was built on top of copyright. Generative patentleft can now be instrumented. Where copyleft used copyright's own lever to guarantee freedom, patentleft-by-publication needs only patent law's own definition of novelty. Four ideas carry the argument: the arithmetization of the claim, the asymmetry theorem, patentleft-by-publication, and the spread between accessibility, which the law pays out on, and attention, which the economy pays out on. One major research question is whether generative models augmented with scarce human curation can multiply the defensive value and the retrievability of contents, without multiplying noise and verification costs at the same rate. Machines now make claims and disclosures cheap to write on both sides. They can also check them. World models and AI systems can test enablement in silico, automated laboratories may test it in reality, and humans and machines may audit the results afterwards. A commons that publishes prior art together with the means to check claims against it could become a public verification layer for the patent system. This paper is also a call for help: its theses are provisional, and researchers, practitioners and institutions are invited to test and correct them, and to measure what the approach achieves. Preprint, not peer-reviewed. English is the reference text. Version 1.0. The text is released under CC0 1.0