Learning Machines
Lawyers are learning to work with artificial intelligence. Artificial intelligence is learning to work with law. This blog explores how — through pedagogy, practice, policy, and the ethical questions that connect them.
-
Consumption Pricing and the Return of Cost-Effective Research
Legora moved its most capable product to consumption-based pricing, which means the bill now tracks how the tool is used. Lawyers who remember per-search Westlaw charges already know what that does: it turns efficient use into a billable skill, and it puts the cost of an associate’s fumbling on a dashboard the client can read.
Read more -
The Verification Standard: Training New Lawyers to Use AI
Law firms have rebuilt their workflows around generative AI, and the document review, first drafts, and disclosure-schedule checks that once trained associates are now the tool's work. Prohibiting the tool is unrealistic and counterproductive; handing it over with no change to how juniors are trained lets them skip the work that builds judgment. The standard that fits is verification: let new lawyers use AI from the start, but require corroboration of every substantive output against independent authority, and against more than one source when the training is serious.
Read more -
Three Approaches, One Goal: AI Policy at Berkeley, Texas, and Boston College
Within a few weeks, Berkeley adopted a rule restricting student AI use, the dean of the University of Texas sent his faculty a letter, and Boston College Law named its first director for AI. The three read as a disagreement, but they serve the same goal the ABA already sets for every law school: graduates who can do legal work and exercise judgment, competently and ethically. What divides them is the instrument and the sequence, when and how to build legal judgment and AI fluency into the same lawyer. Berkeley’s ban is the broadest, and because much of it cannot be detected or enforced, it would work better as a statement of values than as a policy.
Read more -
AI Prompts as Work Product: The Civil Courts Answer Heppner
In March I argued that Heppner’s narrow view of AI work product would not survive contact with civil litigation. Three decisions, capped by the Nassau County order in Assini v. Hayward, have now protected a self-represented litigant’s ChatGPT history from discovery. They reach a defensible result through reasoning that borrows the wrong test and skips the analysis the result requires.
Read more -
Answer Quality Is Not Learning Impact: The Stanford AI-Tutoring Study and Hybrid Legal Education
Sixteen law professors judged nearly 3,000 blind comparisons between AI-generated and human-written answers to contracts questions, and preferred the AI 75 percent of the time. The finding is hard to dismiss and easy to overread. Hybrid JD programs that deliver most coursework asynchronously might see the study as a case for adopting AI tutors as the primary support mechanism for their distance components. The case is strong as far as it goes. The study measured whether professors preferred the AI’s answers, not whether those answers produce the learning outcomes that legal education is supposed to serve.
Read more -
Offensive Prompt Injection as Professional Misconduct
A Brazilian labor court has fined two lawyers for hiding white-text commands in a petition to manipulate the court’s AI, and the Pará bar has suspended them. U.S. bar opinions still treat the lawyer as the user of AI tools and the potential victim when those tools fail. The harder scenario has arrived: the lawyer embedding hidden instructions in a document to manipulate an adversary’s AI workflow. The Model Rules already prohibit the conduct. The bar should say so before a U.S. sanctions order has to.
Read more -
Reading the Limitations Section: What the First RCT on AI and Legal Reasoning Actually Shows
A new RCT finds that law students who used AI on an early synthesis task outperformed their peers on a later reasoning task, even without AI. The finding is encouraging, but conditional. The experimental design gave participants curated sources, decomposed tasks, structured prompts, and a selected model: the same institutional scaffolding most firms have not built. The study’s limitations section is a blueprint for the environment that produced the positive result.
Read more -
How Lawyers Should Prompt in 2026
Anthropic, OpenAI, and Google all updated their prompt engineering guidance in late 2025 and early 2026, and a striking amount of the advice from 2023 and 2024 now degrades the newer models. This post lists the changes practicing lawyers should know about, with concrete prompt patterns drawn from the three vendors’ current documentation.
Read more -
Client-Side AI Recording and NYC Bar Formal Opinion 2025-6
The NYC Bar’s Formal Opinion 2025-6 addresses what happens when clients use their own AI tools to record and transcribe conversations with their lawyers. Read alongside Heppner, the opinion establishes one clear duty (warning clients of AI-related privilege risks) and suggests two further responses (providing privilege-preserving alternatives and redesigning communication channels) that the rules do not yet require but that firms should weigh in light of the foreseeability shift Heppner introduces.
Read more -
SB 26-189's Obsolete Exemption: Statutory Categories and Moving Technology
Colorado’s SB 26-189 exempts AI tools used solely to ‘summarize, organize, translate, draft, route, or present information for human review.’ That exemption was drafted for a model of AI use—ask a question, get an answer, review the answer—that the legal technology market has already moved past. The tools law firms are buying don’t summarize information for lawyers; they make the analytical choices that determine what gets summarized, in what order, and with what emphasis.
Read more