Legal AI models exploit 'shortcut learning' on outcome-predicting cues

A new paper published on arXiv investigates shortcut learning in legal judgment prediction (LJP) models, specifically within the context of the UK Employment Tribunal. Researchers found that current LJP models, trained on post-hoc judicial materials, often achieve high predictive performance not through true forecasting but by exploiting linguistic cues that reveal the outcome. The study demonstrated that a model trained on only 4% of identified leakage features could outperform human experts, highlighting concerns about inflated performance metrics in LJP systems. However, the paper suggests this vulnerability is not insurmountable, as models still retain predictive capability even after these outcome-revealing artefacts are removed. AI

IMPACT Highlights potential overestimation of AI performance in legal contexts due to linguistic shortcuts, urging for more robust auditing methods.

RANK_REASON Academic paper detailing empirical findings on AI model behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• LLMs
• Macro F1
• Shortcut Learning
• TF-IDF
• UK employment tribunal

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →