Legal AI systems like Westlaw CoCounsel and Lexis Protégé are trained on curated legal data sets and are designed specifically for legal research. General-purpose AI systems like ChatGPT and Claude (whether free or paid tiers) draw from the internet and proprietary data and are not reliable for legal research.

Legal AI systems can help identify relevant authorities, summarize case holdings, and flag potential issues, such as invalid citations or quotations. However, even these tools remain susceptible to errors and require independent verification. The types of errors a legal AI system might make include confusing a dissent with a concurrence, confusing trial and appellate court holdings, or confusing a quote from a party’s brief with a quote from a court’s holding. A 2025 study conducted by Stanford University researchers found that although RAG-based proprietary legal AI tools hallucinate less often than general-purpose AI systems, hallucinations and other errors still appear 17 - 33% of the time. Given the pace at which AI technology evolves, the models used in the Stanford study may have improved since these results were reported. However, no GenAI system is currently hallucination-proof.

General-purpose AI systems, whether free or paid tiers, (e.g., ChatGPT, Claude, or Gemini) are not trained on curated legal datasets and are prone to hallucination. They may fabricate citations or legal authorities that do not exist, misstate holdings, or invent quotations. They are not reliable tools for legal research.

Tools and prompting options integrated into the Westlaw and Lexis legal AI suites can flag potential hallucinations as well as citation and quotation errors in the parties’ filings and in draft decisions and orders. Westlaw and Lexis also have non-GenAI tools that can be used to check for hallucinations. No tool, however, can identify citations to hallucinated authority with perfect accuracy. 

GenAI (whether general-purpose or legal) cannot reason through unsettled law or navigate genuinely novel or nuanced legal issues the way a human can. The use of GenAI for such questions risks delegating core judicial functions as well as importing errors that are difficult to detect because there is no established correct answer to verify against.

Every citation must be independently verified using authoritative sources. Every holding and quotation must be checked against the original. This is true even for legal AI systems like Westlaw CoCounsel and Lexis Protégé.