Best of: The future of AI and the law

Episode Summary

Executive Summary: Stanford law professor Dan Ho explains how AI is reshaping legal work: it can rapidly identify racist deed covenants and obsolete government reporting rules, but general-purpose legal chatbots still hallucinate often and need careful human oversight. The episode argues AI’s biggest legal value is in narrow, high-volume tasks that improve justice and reduce bureaucratic waste.

Main Topics: Why law is a fertile domain for AI (Priority: 5/5): Ho explains that legal work is dominated by research, writing, precedent, and contested interpretation—making it both difficult for humans at scale and well-suited to AI assistance, but not simple question-answering. AI hallucinations in legal settings (Priority: 5/5): The conversation details studies showing general-purpose models often hallucinate legal facts, while retrieval-augmented systems improve performance but still make significant mistakes, especially in uncertain or specialized cases. Detecting racially restrictive covenants in property deeds (Priority: 5/5): Ho describes a California project using AI to scan millions of deed records to identify and redact illegal racist covenants, preserving them for history while removing them from active records. Finding obsolete reporting requirements and regulatory sludge (Priority: 4/5): A second project uses AI to search municipal codes for outdated, burdensome reporting mandates that waste staff time, enabling governments to delete or consolidate them. Open vs. closed AI models and marginal risk (Priority: 4/5): Ho discusses the policy debate over model openness, arguing that regulators should assess the marginal risk and benefit of openness versus closed systems rather than assume openness is inherently dangerous. The future of legal chatbots (Priority: 5/5): Ho argues that fully general legal chatbots are unlikely to replace lawyers soon; the better path is specialized tools for defined legal tasks where accuracy and evaluation are easier.

Key Arguments: Law is an ideal but risky area for AI because legal work involves massive text volumes, precedent, and adversarial reasoning rather than straightforward facts. General-purpose LLMs can perform well on widely available legal knowledge, but they fail badly on uncertain, jurisdiction-specific, or novel legal questions. Hallucination remains a major issue: even with retrieval-augmented generation, legal systems can still produce non-trivial error rates. AI can produce concrete public benefits by accelerating legally mandated redaction of racist covenants and exposing obsolete reporting obligations. Multimodal models can outperform and undercut traditional OCR for digitized legal archives, making large-scale document review more efficient. Regulatory reform often gets stuck in “sludge,” where agencies keep filing old reports because statutes have not been updated or sunsetted. Debates about open vs. closed models should focus on marginal risk and marginal benefit, not ideology, because openness exists on a spectrum and safety features can be stripped through fine-tuning. The most promising legal AI products will be narrow, task-specific tools rather than all-purpose legal advisors.

Data Points: Legal hallucination rate: 60% to 80% - General-purpose models like ChatGPT hallucinated on basic legal facts across roughly 800,000 benchmark queries. Hallucination rate with retrieval-augmented generation: 1/5 to 1/3 - Systems using retrieval-augmented generation improved but still produced non-trivial hallucination rates. California counties affected: 58 counties - 2021 California legislation required county recorder offices to search deed records for racially restrictive covenants. Property deed volume in Santa Clara County: 84 million pages - Illustrates the scale of manual review required for deed record redaction. Manual review result: 90,000 pages reviewed; 400 covenants found - A two-person county team manually screened deed pages before AI assistance. Los Angeles vendor contract: $8 million over 7 years - Reported cost of keyword-search-based processing for covenant review. AI processing speed for deed records: 5 million records in a couple of days - AI system dramatically accelerated identification of racial covenants. San Francisco municipal corpus size: 16 million words - Large enough to make comprehensive human review impractical for statutory/reporter analysis. Reports identified in San Francisco code: 528 reports - STARA ingested municipal code and resolutions to surface reporting requirements. Reduction in reporting obligations: Over one-third - City attorney proposed deleting or modifying more than a third of identified reporting requirements. Documents searched in U.S. Code example: Over 30 million words - Used to illustrate the scale of statutory review in the federal context. Federal labor example: 95 federal employees, 4 months - Time spent compiling a Social Security Administration report on printing operations. Documented legal hallucination instances: Over 140 - Referenced as filings where hallucinations appeared in legal proceedings.

Pivotal Quotes: "large language models have gotten a lot of attention in the last couple years. They are amazing at taking large amounts of text, summarizing it, searching through it for the pieces of information that you need" — Russ Altman: Opening framing of why AI matters for legal text-heavy work "hallucinate 60 to 80% of the time on basic legal facts" — Dan Ho: Describing empirical findings from legal AI benchmark studies "what we really need in this ecosystem is much more analysis of both marginal risk and marginal benefit of open versus closed" — Dan Ho: Summarizing his view on the open-vs-closed AI debate

Implications: AI is already useful for high-volume legal cleanup and research, but safe deployment depends on narrow use cases, human review, and policy scrutiny. For governments and legal systems, the biggest near-term gains are justice and efficiency, not lawyer replacement.