Dual-Audience Advocacy: Adversarial Inputs in AI-Mediated Litigation

Adjudication relies on a shared record through which parties present their proofs and arguments and against which one can assess the decision-maker’s responsiveness and diligence. Delegating the reception of that record to artificial intelligence introduces a differential-legibility gap: a divergence between what a human understands a submission to contain and what a machine intermediary may instead do with it. This article examines how adversarial legal actors may exploit that gap. It develops a taxonomy of three modes of exploitation: Corrupt, which relies on the fact that humans and computers do not parse input the same way; Command, which induces AI-based systems to follow adversarial instructions; and Seduce, which exploits AI’s processing and attentional tendencies through the medium of otherwise ordinary legal prose. It then traces how their effects propagate through the various tasks increasingly delegated to AI models, potentially acquiring professional or judicial authority as their adversarial origin recedes. This analysis informs both the boundary between permissible advocacy and interference and the design of institutional and technical remedies. As a remedy, this article proposes a defence-in-depth approach across the processing pipeline, combining controls on incoming material, constraints on delegated authority, targeted human review and preservation of provenance. Its central claim is that, once legal advocacy is mediated through AI, the temptation will grow – and the capabilities will increasingly allow – to exploit that mediation, triggering a new relationship of offence and defence in advocacy based on how text may be read (or misread) by AI.