The Missing Variable in AI-Assisted Litigation: Architecture and the Quality of Pro Se Access to Justice

Abstract

Version of record: SSRN 7120940, DOI 10.2139/ssrn.7120940. Cite that version. The same text is also at https://vadymchernets.netlify.app/missing-variable.html and https://github.com/vadimchernets/papers/tree/main/missing-variable. This deposit is an archival copy made so that the full text stays retrievable; it makes no separate claim. Nearly every account of AI in litigation rests on an unstated assumption: that AI assistance is one thing. The debate that assumption produces is about volume and fabrication: a federal pro se plaintiff rate up from 11% to nearly 17%, courts describing an "existential threat," a growing docket of sanctions decisions over fabricated citations. What the aggregate studies cannot yet show is whether any configuration of AI assistance changes the quality of what self-represented litigants do. The record used here is the public docket of one federal case, and it holds a pattern the volume-and-hallucination account does not predict. A self-represented litigant with no attorney of any kind, using, by the author's account, a multi-model orchestration system (independent models queried in parallel, disagreement surfaced rather than averaged away, citations verified against external sources), answered six coordinated pre-motion letters within four days, filed a substantive opposition six weeks ahead of deadline, and identified a verifiable statutory-citation error in a brief filed for seven defendants represented by counsel from leading firms, in a case naming thirteen technology defendants. The last of those three is the inverse of the failure mode that dominates the literature. The Article proposes the architecture-dependence hypothesis: the architecture of AI assistance, single-model prompting versus multi-model orchestration, is an empirical variable in the quality of self-represented litigation in its own right, not an implementation detail and not the same thing as the availability of AI. The doctrinal consequences taken up here are for the work-product doctrine, where federal courts split in early 2026, and for the unauthorized-practice framework. The case is offered as a falsifiable existence proof and a research agenda; the author is the plaintiff, the system developer, and the analyst.