Large Language Models & Generative AI in Legal

Legal AI platforms use multiple strategies to minimize hallucination. RAG (retrieval-augmented generation) grounds LLM outputs in verified legal databases: when CoCounsel cites a case, it retrieves the actual case from Westlaw rather than generating text from memory. Citation verification systems check every legal reference against authoritative databases before presenting results. Confidence scoring flags outputs where the model is uncertain. And some platforms restrict generative capabilities to domains where they have verified training data, declining to answer questions outside their knowledge boundaries. These approaches don't eliminate hallucination entirely but reduce it to manageable levels with attorney oversight.

Harvey AI is the most prominent purpose-built legal LLM platform, used by major law firms globally. Thomson Reuters' CoCounsel integrates LLM capabilities across Westlaw, Practical Law, and other Thomson Reuters products. Lexis+ AI from LexisNexis provides similar integration with the Lexis research platform. Spellbook focuses on transactional contract drafting and review. EvenUp specializes in personal injury demand letter generation. Microsoft Copilot for Legal uses Microsoft 365 integration.

Most Am Law 200 firms have established AI governance frameworks that include approved tool lists, usage policies, training requirements, and oversight committees. Common rules include: never input client confidential information into consumer-grade AI tools, always verify AI-generated citations before filing, disclose AI assistance where required by court rules, and restrict use to approved enterprise platforms with appropriate data security. Many firms have Chief AI Officers or AI steering committees. Bar associations including the ABA, state bars, and international equivalents have issued guidance on AI competence, confidentiality, and supervision obligations.

LLMs are augmenting rather than replacing legal professionals, but they are changing the mix of work significantly. Tasks that are primarily information retrieval and text production (basic research, first-draft writing, document summarization, contract markup) are increasingly automated. But judgment, strategy, client counseling, courtroom advocacy, negotiation, and creative problem-solving remain firmly human. The net effect is that each legal professional can handle more work, which may reduce headcount growth over time without mass displacement. The professionals most at risk are those whose roles are primarily document-production focused, while those with strong judgment, client relationship, and strategic skills become more valuable.