Audit: AI Hallucinations Pose Ongoing Risks in Legal Research Platforms
Independent audits confirm AI hallucinations are systemic in major legal research tools.
Why it matters: As legal professionals deploy AI from Westlaw and LexisNexis in workflows, built-in hallucinations can trigger erroneous filings, sanctions, and reputational harm. Understanding and vetting AI output is critical for risk management and client service.
- Auditors and analysts find hallucinations reliably arise from how LLMs generate language, not from random bugs.
- Stanford research shows Westlaw AI made up citations 33% of the time; Lexis+ AI did so at least 17%.
- False AI outputs have factored in hundreds of documented legal proceedings and attorney sanctions since 2023.
- Firms must implement validation processes when using AI for legal research, drafting, and client deliverables.
Independent auditors and academic researchers are urging caution as large language models (LLMs) cement their status in U.S. legal workflows. Hallucinations—where AI generates plausible but false information—are an intrinsic feature of these models, stemming directly from how they predict language rather than from technical malfunctions.
- An industry report highlights: Hallucinations occur predictably in LLMs due to their design, making them a consistent risk for legal users.
- Physics-based analysis explains that the models' method of predicting likely words can suddenly lead to fabrications, especially when prompted for authoritative-sounding answers. This deterministic process is fundamental to LLMs, not a remediable software bug (see technical review).
Most striking: Stanford Law School analysis tested real-world legal research scenarios with major platforms:
- Westlaw AI-Assisted Research hallucinated citations in 33% of queries.
- Lexis+ AI produced at least 17% fabricated outputs in similar tests.
The real-world impact is growing. Documentation after 2023 reveals hundreds of legal proceedings where lawyers filed briefs containing fabricated citations generated by AI. These cases led to court sanctions and lasting reputational issues (see analysis).
Takeaway for legal teams: AI output should always be independently validated before use in client or court documents. Routine due diligence—human review, parallel research, and built-in verification tools—is now essential when using AI in research, drafting, or advisory roles.
By the numbers:
- 33% — share of Westlaw AI queries with hallucinated citations in the Stanford study
- 17% — minimum rate of false outputs from Lexis+ AI in parallel tests
- Hundreds — legal proceedings since 2023 in which AI-generated errors yielded attorney sanctions
Yes, but: Some authorities note that hallucination rates can vary based on prompt structure and model updates, so ongoing monitoring is essential.
What's next: More legal research providers are rolling out prompt guardrails and citation checking features in upcoming software releases, aiming to reduce risk.