Teaching AI to Lawyers: The Four-Layer Framework (Levy, 2026)

Every law firm is already teaching AI to its lawyers. The question is whether anyone is doing it on purpose.

That is the opening premise of Teaching AI to Lawyers, a 2026 guide by Colin S. Levy — legal technology consultant and editor of the Handbook of Legal Tech. The document is not a technical manual or a tool comparison. It is a curriculum design framework for anyone responsible for AI training inside a legal organization: law professors, training partners, CLE coordinators, and legal operations leaders.

📄 Download Teaching AI to Lawyers — Colin S. Levy (2026)

The Training Gap and Its Consequences

Levy's diagnosis is straightforward: two years of generative AI in legal practice have produced a generation of lawyers who use these tools daily without understanding what they do, what they cannot do, and what professional obligations travel with their use.

The consequences are already documented. Researcher Damien Charlotin maintains a public database of judicial decisions addressing lawyer reliance on AI-generated content. As of early 2026, that database contains over 1,300 documented decisions across multiple jurisdictions.

The anchor cases are now well known. In Mata v. Avianca (S.D.N.Y. 2023), a federal court sanctioned counsel $5,000 for filing six fabricated citations generated by ChatGPT and then providing fabricated opinion texts when challenged. In Park v. Kim (2d Cir. 2024), the Second Circuit referred an attorney to its Grievance Panel for citing a nonexistent case with no inquiry into its validity. In Noland v. Land of the Free (Cal. Ct. App. Sept. 2025) — the first published California Court of Appeal opinion on AI-fabricated authority — a $10,000 sanction was imposed and counsel was reported to the State Bar. The same pattern emerged in Zhang v. Chen (2024 BCSC 285), where the Supreme Court of British Columbia ordered counsel to personally compensate opposing counsel for time spent unwinding ChatGPT-generated fabrications.

Levy's reading is consistent across all cases: none of these lawyers acted with malice. They were trying to do their jobs faster with a tool no one had taught them to use properly.

The Standard the Profession Needs

The guide rejects two common failure modes in AI training programs. The first is trying to turn lawyers into junior machine learning engineers. The second is treating the tools as a black box and teaching nothing technical at all.

The standard Levy proposes is functional. A lawyer needs to know enough to ask the right questions of a vendor, recognize when output is suspicious, push back on an associate who cannot explain what tool they used, and read a data processing addendum without losing the thread. It is the same kind of competence a litigator already brings to a damages model from an opposing expert: you do not need to derive the regression, but you do need to know that a regression has assumptions and that the right questions will surface them.

The Four-Layer Curriculum

The guide's central model is a four-level structure that must be taught in order, because each layer rests on the one below it.

Layer One: Conceptual Foundations

What a large language model does. Why it produces confident text that turns out to be wrong. What retrieval-augmented generation (RAG) does, and what it does not fix.

Levy distills the core concept into a single sentence he recommends using to open any AI training: a large language model produces text by predicting the most probable next token, given the text that came before. That sentence — technically simplified but functionally correct — explains why citations get fabricated, why the model cannot tell you what it does not know, and why two nearly identical prompts can produce very different outputs.

The empirical grounding comes from the Magesh et al. study, published in 2025 in the Journal of Empirical Legal Studies. Measured hallucination rates on federal case-law queries: 17% for Lexis+ AI, 33% for Westlaw AI-Assisted Research, 43% for GPT-4. Purpose-built legal tools using RAG hallucinate substantially less than general-purpose models — but they do not reach zero.

Layer Two: Practical Skills

How to write a prompt that produces useful work. How to verify what comes back. How to decide when a task does not belong in AI at all.

Levy introduces a three-tier classification framework: red for tasks where an error would cause material harm and verification is difficult; yellow for tasks that belong in AI only with rigorous verification; green for tasks where the cost of error is low and verification is straightforward. Most of the real, reliable value AI delivers sits in the green tier.

The lab exercise he recommends is simple and replicable: each participant uses an approved tool to produce a first draft, runs a citation-by-citation verification pass against primary sources, and documents the errors found. Hallucinations surface in nearly every session. An error seen once in a low-stakes practice exercise is dramatically easier to catch in a real matter.

Layer Three: Ethics and Professional Responsibility

The guide anchors this section in ABA Formal Opinion 512, issued July 29, 2024 — the first comprehensive ABA opinion on lawyer use of generative AI. The opinion applies six existing duties to the new tool: competence (Rule 1.1), confidentiality (Rule 1.6), communication (Rule 1.4), candor (Rules 3.1 and 3.3), supervision (Rules 5.1 and 5.3), and reasonable fees (Rule 1.5).

The pedagogical point Levy emphasizes: Opinion 512 does not invent new duties. It applies existing duties to a new tool. That framing matters in a classroom or CLE because it tells the audience that the professional framework they already carry is mostly correct — what changes is the mechanism through which those duties show up in daily practice.

The supervisory analogy that runs throughout the guide is the junior associate: the reasons we verify a first draft, supervise research, and never let an associate sign a brief unread are the same reasons we verify, supervise, and review AI-assisted work. Opinion 512 adopts this framing explicitly, treating generative AI tools as nonlawyer assistants for purposes of Rules 5.1 and 5.3.

On fees: a lawyer cannot bill the client for time saved by using AI, cannot bill for time spent learning a generally applicable tool, and may charge for the tool's actual cost only with adequate disclosure and consent.

Layer Four: Governance

How an organization translates individual competence into a defensible system: written policy, an approved-tool list, training records, and incident reporting. This layer is where individual skill becomes institutional practice.

How Lawyers Actually Learn

A chapter worth separate attention addresses pedagogy. Lawyers are adult learners with deeply formed mental models of how knowledge work happens. Training that fails treats those frameworks as obstacles. Training that works treats them as leverage.

Lawyer skepticism is a pedagogical asset, not a problem. It is the same professional reflex they use to cross-examine experts or read contracts against their plain meaning. The training that gets lawyers engaged shows a hallucinated citation and asks the room how they would have caught it. Or hands them a vendor data processing addendum with a training-rights clause embedded in it and asks what the negotiation looks like.

Levy identifies five audience segments that each require different curriculum adjustments: solo and small firm, BigLaw, in-house, government and public sector, and the bench. The solo practitioner needs an opinionated, time-compressed approach: one or two tools, a clear contractual posture, and a pre-filing checklist they can adopt the same week. The BigLaw associate is already operating inside a firm policy and approved-tool list; the curriculum aligns them to that architecture. In-house counsel adds vendor diligence, data processing terms, and the EU AI Act exposure that internal legal teams now face.

The Maturity Model

The guide closes its assessment section with a four-stage maturity model any program can use to locate itself and plan the next step.

Reactive: No formal curriculum. AI use is informal. The firm learns about problems when opposing counsel or a court raises them.

Compliant: An annual training session is in place, typically slides, no hands-on component, attendance as the only evaluation criterion. It meets the floor of the supervisory rules but does not yet produce measurable competence.

Proactive: A multi-module curriculum covers all four layers. Hands-on labs are part of the design. Assessment uses scenario-based evaluation. The program updates when guidance or case law changes.

Strategic: An owned program with dedicated capacity. The curriculum iterates from incident data. A mature case bank supports labs and assessments. The supervisory architecture is documented and would be defensible under examination.

Levy's practical advice: move one stage per year. Two stages in one year are rare. And no stage is permanent — a strategic program that loses its owner reverts to compliant within eighteen months without anyone noticing until something goes wrong.

What Can Be Replicated Now

The guide includes three directly usable artifacts: a complete fourteen-week law school syllabus, a primary-sources reading list organized by module, and assessment rubrics for three evaluation formats — scenario-based written assessment, lab portfolio, and capstone project.

For CLE designers, the three-hour outline Levy proposes is operational: hour one on foundations and prompting; hour two on verification protocols and the triage framework; hour three on the six duties under Opinion 512, the firm's own policy, and three concrete action items each participant will complete that week.

A Note on the EU Context

The guide is written from a U.S. regulatory framework, but its international references include the EU AI Act (Regulation 2024/1689), whose majority of obligations apply from August 2, 2026, the Supreme Court of New South Wales Practice Note SC Gen 23 (effective February 2025), and Singapore Registrar's Circular No. 1 of 2024. Legal teams operating in EU jurisdictions should layer the AI Act's risk classification requirements on top of Levy's four-layer model, particularly for high-risk system categories that include uses affecting access to justice and the administration of law.

Teaching AI to Lawyers is available free of charge. Download the full document below.

📄 Download Teaching AI to Lawyers — Colin S. Levy (2026)