AI legal research startup Descrybe launched DescrybeLM, a purpose-built legal reasoning engine that...
The AMW Read
The article provides empirical evidence for the 'Vertical AI bull' frame by demonstrating that domain-specific reasoning engines can outperform frontier general-purpose models in high-stakes legal accuracy and hallucination mitigation.
AI legal research startup Descrybe launched DescrybeLM, a purpose-built legal reasoning engine that achieved 100% accuracy on 200 NCBE bar exam questions, while ChatGPT 5.2, Claude Opus 4.5, and Gemini 3 Pro scored 88.5% to 93.5%. The critical finding: 94% of general-purpose model errors (49 out of 52) were "confidently wrong" - fluent, assertive responses with no uncertainty signal, imposing massive verification burdens on practitioners. Even cross-checking between models failed as a safeguard, with non-overlapping errors across 40 questions. This proves domain-specific AI trained on structured legal data fundamentally outperforms general models for legal reasoning. The legal industry now faces a clear choice: continue with hallucination-prone general AI or adopt purpose-built systems designed for legal accuracy. #LegalAI #AI #LegalTech #LawAndTechnology #ProfessionalServices