The future of AI and the law

Episode Summary

Executive Summary: The episode explores how large language models can help legal research and government reform, but also how hallucinations, overbroad use cases, and governance choices limit their reliability. Stanford’s Dan Ho describes AI systems that scan massive legal corpora to identify racist covenants and obsolete reporting rules, while warning that general-purpose legal chatbots still need narrow, high-confidence applications.

Main Topics: Why law is a fertile but difficult domain for AI (Priority: 5/5): Dan Ho explains that lawyers spend much of their time on research and writing, and that law’s adversarial, precedent-based nature makes it a rich testbed for AI—but harder than ordinary question answering. Legal hallucinations and retrieval-augmented generation (Priority: 5/5): Ho summarizes studies showing general-purpose models hallucinate frequently on legal facts, and that retrieval-augmented systems improve performance but still produce non-trivial errors. AI for identifying racist covenants in property records (Priority: 5/5): The team built a system to scan millions of deed records for racially restrictive covenants, preserving historical records while enabling formal redaction under California law. STARA and statutory/regulatory sludge (Priority: 4/5): A broader legal research assistant, STARA, was built to ingest statutes and regulations and identify obsolete reporting mandates, helping governments reduce administrative burden. Examples of obsolete reporting requirements (Priority: 4/5): The conversation highlights vivid cases like reports on defunct programs, newspaper rack zones, and the presidential dollar coin program—illustrating how old rules create bureaucratic waste. Open vs. closed models in legal and broader AI (Priority: 4/5): Ho argues openness should be judged by marginal risk and marginal benefit, noting that closed legal research tools have a long precedent and that risk from open models is not straightforward. Future of legal chatbots: specialized, not universal (Priority: 5/5): Ho concludes that chatbots may democratize access to legal information, but the safest and most useful path is narrow, task-specific systems rather than end-to-end legal advice bots.

Key Arguments: Law is a strong AI application area because much of legal work is text-heavy research and synthesis, not just courtroom performance. General-purpose LLMs can contain legal knowledge, but they hallucinate too often to be trusted for uncertain or novel legal questions. Retrieval-augmented generation improves legal accuracy, but still leaves meaningful hallucination rates that require caution. AI can materially accelerate legal reform by scanning massive corpora that humans cannot review efficiently, as shown in the racist covenant project. A general statutory search platform can help governments find obsolete, duplicative, or politically outdated reporting requirements and reduce regulatory sludge. Open-vs-closed model debates should focus on marginal risk and marginal benefit, not ideology, because openness exists on a spectrum and safety measures can be stripped away. The most promising legal AI systems will be specialized for defined use cases, where evaluation is easier and hallucination risk can be lower.

Data Points: Legal hallucination rate for general-purpose models: 60% to 80% - Study of roughly 800,000 benchmark legal queries asked of models like ChatGPT on basic legal facts Hallucination rate for retrieval-augmented legal systems: 20% to 33% - Follow-up study using retrieval augmented generation to answer legal questions Benchmark queries: roughly 800,000 - Scale of queries used in the legal hallucination study Santa Clara County deed records: 84 million pages - Volume of property deed records that needed scanning for racially restrictive covenants Manual review output: nearly 90,000 pages; 400 covenants found - County team manually reviewed records before AI assistance Automated scan scale: 5 million deed records - AI pipeline used to accelerate covenant identification in Santa Clara County Los Angeles vendor contract: $8 million - Cost quoted for keyword-search-based covenant identification over about seven years U.S. code length: over 30 million words - Referenced as the scale of federal statutory text Ruth Bader Ginsburg’s team searched San Francisco municipal corpus: about 16 million words - Size of the municipal code and resolutions ingested into STARA Reports identified in San Francisco: around 528 - Number of reporting requirements found by the STARA system Resolution size: 351-page resolution - Proposed deletion or modification of more than a third of reporting obligations Federal instances documented: over 140 - Documented filings with legal hallucinations Human effort on SSA report: 95 federal employees for four months - Effort required to compile a report on printing operations

Pivotal Quotes: "law has become a really interesting terrain for AI research" — Dan Ho: Explaining why legal reasoning is a challenging but valuable domain for LLMs "hallucinate 60 to 80% of the time" — Dan Ho: Describing results from studies of general-purpose language models on legal facts "the future of a general purpose legal chatbot ... they will probably be high specialized and very narrow in their application domain" — Russ Altman / Dan Ho: Closing exchange about whether chatbots can replace lawyers

Implications: Legal AI is powerful for search, document review, and reform, but not yet reliable enough for broad legal advice. Expect narrow, audited tools to spread first, while institutions debate transparency, safety, and accountability.