Legal AI study: Trust matters more than accuracy in court predictions

A new study from researcher Surya Saka challenges the assumption that fusing uncertainty tools with large language models (LLMs) improves legal prediction accuracy. Testing on 1,000 European Court of Human Rights cases using frontier models GPT-5.5 and Claude Opus 4.8, the research found that raw LLMs alone outperformed any fused pipeline in discrimination tasks (AUROC ~0.83).

The study evaluated three families of evidence estimators: raw LLMs, LLMs routed through uncertainty fusion pipelines (evidence graphs with belief propagation, Bayesian odds updating, Dempster-Shafer combination, and conformal prediction), and a term-frequency baseline. Results showed that naive fusion of uncertainty tools with LLMs doubled calibration errors (ECE from 0.16 to 0.46) due to prior-mismatch mechanisms. Dempster-Shafer fusion was particularly unsafe, committing confidently to wrong labels at below-chance accuracy. However, when Dempster-Shafer was removed, the pipeline achieved 96.8% accuracy with only 0.5% errors escaping review, compared to 85.9% accuracy for an untuned baseline. The key takeaway: calibrated trust, not sharper prediction, is the real value of these pipelines.