AI & The Law: Changing Practice, Claude Constitution, & New Rights, w/ Kevin & Alan of Scaling Laws

What It Covers

Kevin Frazier and Alan Rosenstein examine how AI transforms legal practice and policy. They cover frontier models outperforming median lawyers, the billable hour disincentivizing AI adoption, legal deserts requiring better access, complete contingent contracts, outcome-oriented legislation with AI simulation, the unitary artificial executive concept, new rights like compute access and data sharing, and emerging questions around AI sentience and welfare.

Key Questions Answered

• •Current AI Legal Capability: Frontier models like Claude Opus 3.5 win one in three head-to-head comparisons versus human lawyers and tie or win 70% of matchups. They already exceed median lawyer performance in raw intellectual horsepower, though hallucinations and database access limitations remain. OpenAI appears to have the best legal taste from focused RLHF training. Within a few years, models will likely be vastly superior to most practicing attorneys across standardized legal tasks that constitute the majority of legal work.
• •Adoption Paradox in Law Firms: Despite 70% of top 100 US law firms licensing Harvey AI, actual daily usage remains surprisingly low. Lawyers receive minimal training beyond initial email announcements and face no obligation to use AI tools. The billable hour compensation structure creates perverse incentives where attorneys maximize time spent on tasks rather than efficiency. Secret cyborgs who master AI tools quietly outperform peers without revealing their methods, while firms whisper about hiring fewer junior associates.
• •Legal Desert Solution: One lawyer serves every 1,000 residents in legal desert areas, leaving people unable to access services for leases, business formation, divorces, and disputes. AI can democratize access to quality legal services at dramatically lower costs. Landlord-tenant dispute trials show tenants with even minimal legal counsel achieve significantly better outcomes. The latent demand for affordable, accessible legal services represents a massive untapped market that AI could address through scalable expertise delivery.
• •Complete Contingent Contracts: With infinite time and resources, optimal contracts would address every possible contingency between parties, creating socially optimal agreements. Current contracts remain incomplete because negotiating comprehensive terms is prohibitively expensive. AI agents representing each party could negotiate at 400 tokens per second, exploring the full contingency space to create near-complete contracts. This transforms contract law from default rules that frequently misfire into precisely tailored agreements reflecting actual party preferences across all scenarios.
• •Outcome-Oriented Legislation: Current laws follow centuries-old formats without defining intended outcomes or running simulations before passage. Legislators should specify explicit goals like reducing unemployment to 7% or cutting carbon emissions by specific percentages, then use AI to simulate proposed legislation against these targets. NEPA environmental law created unintended veto points that block affordable housing, problems that simulation could have identified. Future generations will view failure to test laws through AI simulation as incomprehensible negligence.
• •Unitary Artificial Executive: AI enables unprecedented presidential control over the executive branch's millions of employees through perfect enforcement, mass surveillance, and propaganda creation at scale. An AI trained on presidential preferences and injected throughout bureaucracy can read all communications and ensure alignment with executive intent. This dramatically increases executive power beyond current legal authorities through practical management capabilities. The challenge involves encouraging AI adoption for improved government services while preventing authoritarian abuse of centralized control mechanisms.
• •Right to Compute and Data Sharing: Montana enacted right to compute legislation, with Ohio and New Hampshire considering similar bills protecting access to computational tools from government restriction. Individuals need the right to share personal data frictionlessly with AI services for personalized tutoring, health optimization, and other applications. Current privacy laws like FERPA create burdensome barriers requiring annual consent forms. Only wealthy individuals access services like comprehensive health data analysis, while others remain limited to basic checkups and fragmented records.