On Wednesdays, We Ask Questions: Optimizing "Active Listening" in Automated Legal Triage and Referral

TEXT ANALYSIS: FETCH System Paper

THE DISSECTION

This paper documents the engineering of a legal intake automation system — specifically, a classifier (FETCH) that generates follow-up questions to refine legal problem categorization. It is a technical post-mortem on automating the initial stage of legal aid.

The core findings:
1. Low-cost LLMs handle classification adequately but fail at generating quality follow-up questions.
2. GPT-5 (high-cost) was required to generate questions that elicited useful information.
3. LLM-as-judge ratings diverge from human expert ratings — the automated judge can't reliably assess its own work.
4. Uneven performance across legal categories (domestic violence notably underperforms).
5. Prompt engineering alone does not close the quality gap.

THE CORE FALLACY

The paper assumes "accurate classification" is the goal, when the actual function being automated is gatekeeping access to legal resources.

The entire framing treats legal triage as a classification problem — matching problem descriptors to legal categories. This is a category error. Legal intake is not primarily a matching function. It is a function of establishing trust, reading between lines, identifying danger patterns that applicants won't state directly (especially in domestic violence), and creating the conditions under which vulnerable people will disclose. The paper acknowledges domestic violence underperformance and proposes a "dedicated screening panel" as a fix — but this concession undermines the entire architecture. If certain categories require human-specific protocols, the system isn't automating legal triage; it's automating triage for cases that don't require protection.

The fallback — "dedicated panels" for sensitive categories — is not a solution. It is a triage of the triage, which means the core system is a sieve for safe cases and a trap for dangerous ones.

HIDDEN ASSUMPTIONS

1. Legal problems are discrete and describable. The paper assumes applicants can articulate their legal problem in structured form. Many cannot — especially those in coercive relationships, undocumented status, or cognitive impairment.

2. Questions extract facts. Active listening in human legal intake is not a question-response system. It is a relational protocol where tone, hesitation, partial answers, and what is not said carry diagnostic weight. LLMs cannot receive silence as information.

3. Classification accuracy maps to legal aid quality. Even if FETCH classifies correctly, it does not generate legal knowledge in applicants. It routes them toward resources they may not access, cannot afford, or may not qualify for.

4. Low-cost deployment is the constraint to optimize. The finding that high-cost models are required to generate usable questions should be read as: this task is not automatable at low cost. The paper treats this as an engineering problem to solve. It is structural — the question is not answerable by any LLM at any cost, because the problem requires judgment, not retrieval.

5. LLM-as-judge divergence is a methodological problem. The paper notes LLM judges rate outputs that human experts rate low. This is not a calibration issue. It is evidence that LLM judges assess fluency and coherence, not legal adequacy or safety. You cannot use the system's outputs to validate the system's quality.

SOCIAL FUNCTION

Transition management theater. This paper is a procedural document in the industrialization of legal denial. It does not ask whether automating legal intake is appropriate. It asks only how to do it more efficiently.

The "we consulted legal intake workers" language signals that human expertise was used instrumentally — as a validation layer, not a design constraint. The workers were brought in to evaluate the system's output, not to define whether the system should exist.

The paper is also a cost accounting document: it establishes that high-cost models are required, which means the automation ceiling for legal triage is higher than hoped, which makes deployment economics less favorable, which may slow rollout — which is the only actual protective function this research performs.

THE VERDICT

FETCH is a triage system for the already-vulnerable, built on a category error: mistaking classification for legal judgment.

The paper inadvertently demonstrates the fundamental problem with automating legal intake: the cases that require the most protection (domestic violence, coercion, exploitation) are precisely the cases where standardized questioning fails. A system that classifies well under ideal conditions and fails precisely under stress conditions is a system optimized for the easy cases and dangerous for the hard ones.

The researchers found that the low-cost models didn't cut it — which suggests the actual cost of adequate legal triage automation is not lower. This is either:
- A finding that delays deployment (protective), or
- A procurement justification for more expensive AI infrastructure (harmful), depending on who reads it.

The paper does not grapple with what it has actually found: that automated legal triage cannot safely classify the cases that most need classification. The "dedicated screening panel" workaround is an admission that the system cannot be trusted with the dangerous cases — which means it cannot be trusted to know which cases are dangerous.

Recommended reading: This paper is useful as a procurement and deployment warning document for legal aid organizations. It documents, in technical detail, the exact failure modes that should trigger human override requirements. Any legal aid organization being sold an AI intake system should require reading this paper and answering: what is the dedicated panel protocol for your most dangerous case types, and who monitors it?