Legal Reasoning Is Not Lawyering: Rethinking Legal Benchmarks for Pro Se Access to Justice

Shallow read  ·  2026  ·  source  ·  all reading

Legal Reasoning Is Not Lawyering: Rethinking Legal Benchmarks for Pro Se Access to Justice

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2606.23716 Date read: 2026-06-24 Connected to: none Escalation: escalate-to-deep Escalation rationale: Identifies a fundamental measurement gap in how AI systems interact with human institutions—benchmarks measure expert-preprocessed inputs rather than real-world messy inputs—which challenges the validity of access-to-justice claims and suggests a generalizable pattern about protocol-system validity under distribution shift.

What this is

This is a critical empirical/methodological paper arguing that legal AI benchmarks systematically misrepresent model capability for real-world access-to-justice applications. The core claim: existing benchmarks test performance on cleaned, expert-preprocessed legal inputs, not the unstructured, ambiguous, partially-informed inputs that pro se (self-represented) litigants actually produce. This creates a measurement illusion where models appear competent for justice access when they may not be.

What I took from it

This work identifies a validity collapse in protocolized systems: when an AI system is evaluated in an idealized protocol environment (curated benchmarks) but deployed in an adversarial or uncontrolled environment (actual human legal needs), the measured capability becomes non-predictive. The paper suggests that institutional integration cannot be assumed from task performance—there is a hidden lower bound problem.

More broadly, this signals a pattern relevant to all protocol-system research: benchmarks that measure only expert-mediated or pre-processed versions of a task are systematically misleading about real-world utility. The gap between upper-bound and lower-bound performance may be a structural feature of how AI systems encounter human institutions, particularly where humans have heterogeneous inputs, knowledge states, and incentives.

This also touches on legitimacy and capture: benchmarks drive funding and deployment claims, but if benchmarks don't measure what matters for the stated purpose, the system creates false confidence in institutional adoption.

Research connections

  • Protocol-system validity under distribution shift: Real-world inputs from pro se litigants are a severe distribution shift from expert-curated legal text; benchmarks do not measure robustness to this shift.
  • Institutional integration assumptions: Claims about AI improving access to justice assume seamless human-protocol interaction, but this paper suggests the assumption is empirically ungrounded.
  • Measurement bias in capability claims: Benchmarks become decoupled from real-world utility when they test only the idealized protocol path, not the messy deployment path.

Candidate laws or signals

  • CL-2606-1: Benchmark-Reality Gap in Institutional AI: When an AI system is evaluated on expert-preprocessed or curated inputs and deployed against unmediated human inputs, measured capability becomes a misleading upper bound; real-world utility requires lower-bound testing on uncontrolled inputs from the actual user population.

  • CL-2606-2: Institutional Legitimacy Through Measurement Decoupling: Benchmark design in high-stakes institutional domains (law, medicine, justice) can systematically inflate capability claims by testing only the path through expert gatekeepers, obscuring failure modes when users bypass or misalign with that path.