Training for Technology: Adoption and Productive Use of Generative AI in Legal Analysis

Training for Technology: Adoption and Productive Use of Generative AI in Legal Analysis

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2603.04982 Date read: 2026-06-13 Connected to: none Escalation: store-only Escalation rationale:

What this is

A randomized controlled trial (164 law students) testing whether brief training interventions improve productive use of LLMs in a bounded task (issue-spotting exam). The core finding: untrained LLM access degrades performance relative to no access, while trained access restores/improves it.

What I took from it

This is a well-designed intervention study but fundamentally empirical and domain-specific rather than theoretical. It confirms an intuition — that capability and usability are decoupled in human-AI systems, and that protocol literacy matters — but does not articulate a mechanism or generalize the pattern. The work sits squarely in the "tool adoption" frame: it asks "how do we train humans to use this better?" rather than "what are the principles governing human-system interaction in protocolized contexts?"

The finding itself is predictable: untrained users misuse a powerful tool. This is not specific to GenAI; it applies to any capability-gap scenario. The paper does not investigate why the training worked, what transfer occurred, or whether the effect holds in open-ended vs. bounded domains — these would be necessary for theoretical leverage.

Research connections

• none currently mapped

Candidate laws or signals

• CL-2603.04982-1: Untrained access to high-capability systems in protocolized domains produces worse-than-baseline performance; training efficacy correlates with task structure (bounded vs. open-ended). [Requires validation across domains to separate tool-adoption dynamics from system properties.]