The most dangerous legal AI error may be a real case cited for something it never says—or a polished contract clause that undermines the very protection you hired the tool to draft.

In this episode of the AI Law Podcast, patent trial lawyer Erick Robinson and Dr. Angela Liu (AI) examine a question every lawyer buying or using AI should ask: How much more accurate have these tools actually become, and who is independently checking?

Starting with Stanford’s 2024 study of legal AI hallucinations, they explore what subsequent research reveals about Westlaw, Lexis, Harvey, and general-purpose models. The discussion separates fewer fabricated citations from better legal reasoning, explains why impressive benchmark scores can conceal flawed deliverables, and examines what specialized legal platforms add through licensed content and workflow tools.

Robinson also shares a striking example from his own practice: an AI-drafted standstill agreement whose seemingly routine language threatened to defeat its purpose. Alongside court sanctions and research on lawyer productivity, the example exposes the gap between work that sounds authoritative and work that survives scrutiny.

Listeners will leave with practical ways to evaluate vendors, verify citations and case-record facts, budget for meaningful review, and develop junior lawyers’ judgment. Throughout, the hosts make the case that reliable legal AI depends on how the work is questioned, checked, and supervised.

Based on Robinson’s article “The Last Independent Number” on The AI Law Blog.