Lesson objective

Diagnose why an information-led AI lecture leaves a lawyer's working practice unchanged, and audit an existing deck against the behaviour it should produce.

Plain-language explanation

Why this matters now

Much of the AI training delivered to lawyers takes one shape. An hour, a slide deck, a speaker, a list of risks, a demonstration of the tool doing something impressive, and ten minutes of questions at the end. Participants leave having heard accurate things. On Monday morning they open the same document, face the same deadline, and do what they did last week.

This lesson makes an argument from reasoning and from the structure of legal work, not from education research. It does not claim that lectures are proven to fail, or cite a percentage of what people remember. This programme holds no such evidence and you should not repeat anyone else's version of it. The argument stands without it, and a room of lawyers can follow it step by step.

A question to sit with

Every lawyer who has attended an AI session in the past two years has been told to check the output. Ask yourself honestly: what did being told add, and what would a lawyer need, beyond being told, to actually do it at 6pm with a filing due?

What goes wrong

1. The deck is the product. The instructor's preparation goes into slides, so the session is organised around the slides. The question "what will anyone do differently" is never written down, and so nothing in the session is built to produce it.

2. The behaviour is a sequence; the lecture delivers a fact. "Verify every citation" is a fact. Doing it is a sequence of actions: noticing that a paragraph contains an authority, stopping, opening a research service, finding the authority, reading the passage, comparing it with the text, and recording that it was done. A lecture can state the fact in one slide. It rehearses none of the sequence.

3. The demonstration only succeeds. A tool that produces a good summary on stage teaches the room that the tool produces good summaries. It teaches nothing about what to do when it does not, which is the moment the working lawyer actually needs to handle.

4. The tips slide. Ten bullet points of good advice. Each is correct. None is attached to a moment in the lawyer's day, so none is retrieved when that moment arrives.

5. The research is taught as a headline. A slide says "AI makes lawyers faster". That is half of what the study it came from found, and the missing half is the part that matters for Monday.

The practical answer

Best practice: apply the Monday test to every element. For each slide, demonstration and discussion in a session, write one sentence completing: "Because of this, on Monday a participant will..." If the sentence ends in "know", "understand" or "be aware", the element produces information and nothing else. If it ends in a verb a lawyer performs at a desk, it may produce behaviour, provided the session also gives them a chance to perform it.

Separate the three things a lecture confuses.

Layer What it is Can a lecture deliver it?
Information The fact: output must be verified, confidential material must not go into an unapproved tool Yes, efficiently
Decision Recognising, in a real document, the moment the fact applies Poorly; recognition needs examples to recognise
Routine The sequence of actions, fast enough to survive a deadline No; it has to be done

A lawyer's Monday failure is almost never at the information layer. They know. The failure is at the decision layer, where a fluent paragraph does not announce that it contains an unverified authority, or at the routine layer, where the check takes longer than the time left. A session aimed only at the first layer is aimed at the one place the problem is not.

Best practice: audit an existing deck in four passes.

  1. Count. Mark every slide as information, decision practice or routine practice. Most AI decks are nearly all information.
  2. Find the moments. List the moments in the participants' actual week where the session's content should change what they do: before pasting text into a tool, before relying on a summary, before a draft leaves the building.
  3. Match. For each moment, find the part of the session that rehearses it. Gaps are where Monday goes unchanged.
  4. Cut. Remove information slides that do not serve any listed moment. The time released pays for practice.

Best practice: attach every piece of advice to a moment. Replace the tips slide with a short table the participant will recognise at their desk.

When you are about to... Stop and...
Paste client material into a tool Confirm the tool is approved for that material under your firm's policy
Rely on an AI summary of a document Open the source and check the passages the summary turns on
Send a draft containing any cited authority Confirm each authority exists and says what the draft says it says
Tell a client how long something took Record the time actually spent, not the time it would have taken

The last row reflects Professional guidance: ABA Formal Opinion 512 (29 July 2024) addresses billing for AI-assisted work in terms of time actually expended. It interprets the Model Rules, which each jurisdiction adopts differently, so the instructor tells participants to check their own jurisdiction's rules and guidance on fees.

Accept what a lecture is for. Short input still has a place: to give the minimum information a practice task needs, to frame why the task matters, and to name a pattern after participants have met it. Module 2 treats that minimum in detail. The fault is not in explaining. It is in stopping there.

Remember this

  • A lecture delivers information; a lawyer's Monday failure is usually at the decision or routine layer, where information alone does not reach.
  • Apply the Monday test: if an element's outcome is "know" or "be aware", it has not been built to change behaviour.
  • A demonstration that only succeeds teaches that the tool succeeds.
  • Attach advice to the moment in the working week when it applies.
  • Report research with its whole finding and its limits; Choi and colleagues found large speed gains and only slight, inconsistent quality gains, among law students using a 2023 model.

Legal example

The fictional Tanager County Bar Association invites Hollis Wyatt-Brenner, a fictional in-house lawyer, to give its annual AI programme. Her existing deck has forty-one slides for sixty minutes. She applies the four-pass audit. Counting, she finds thirty-six information slides, three demonstration slides and two slides of questions. Listing the moments in her audience's week where the session should change what they do, she finds five: pasting client material into a tool, relying on a summary of a long document, sending a draft containing authority, describing AI use to a client, and recording time. Matching, she finds that none of the five is rehearsed anywhere in the deck. Her demonstration shows a tool producing a clean summary of a fictional lease, which, she realises, teaches the audience that the tool summarises leases well. Her slide on research says "studies show AI makes lawyers significantly faster", and when she reopens the Choi, Monahan and Schwarcz article she finds that the quality result was slight and inconsistent, that the participants were law students, and that the model dates from 2023. She rewrites the slide to report the whole finding and its limits. Cutting, she removes twenty-four slides. The released time pays for two tasks: participants receive a fictional two-page AI summary of a fictional commercial lease with one misstated renewal term and one invented obligation, and must decide which passages of the underlying lease to open before relying on it; then they receive a fictional draft letter containing three authorities, one labelled inline (a fictional citation invented for this lesson; it does not exist), and must write down the steps they would take before sending. Her closing slide is the four-row "when you are about to" table.

Responsible-use warning

Do not replace one unsupported claim with another. It is tempting to tell a room that "research proves lectures do not work" or that people "remember only a small percentage" of what they hear. This programme holds no evidence for any such figure, and a room of lawyers will ask for the source. Make the argument from the structure of legal work, which the audience can check against their own week. And never report a study's headline without its limits, its sample and the date of the tool it tested.

Human-review requirement

A licensed attorney reviews every statement of authority in the rebuilt session before it is delivered. The reviewer confirms that billing guidance is attributed to ABA Formal Opinion 512 as guidance rather than law, with an instruction to check the participant's own jurisdiction; that any study is reported with its sample, the date of the model tested and its full result; that every "stop and" step reflects the firm's actual policy rather than a generic one; and that the reviewer is named on the material.

Practical exercise

Audit a deck against Monday. Use a real deck you have delivered or attended, or build a fictional one of at least fifteen slide titles.

  1. Count: label every slide as information, decision practice or routine practice. Record the totals.
  2. List five moments in the target audience's working week where the session should change what they do.
  3. For each moment, identify the part of the session that rehearses it. Write "none" where nothing does.
  4. Apply the Monday test to every remaining slide and rewrite any outcome that ends in "know", "understand" or "be aware".
  5. Identify any research reported in the deck. For each, write the authors, the sample, the date of the tool tested, and the whole result. Rewrite any slide that reports only part of it.
  6. Cut every information slide that serves none of your five moments. Record how many minutes that releases.
  7. Design one short practice task for the released time using fictional material. Label any fictional citation inline exactly: (a fictional citation invented for this lesson; it does not exist).
  8. Replace any list of tips with a "when you are about to... stop and..." table.
  9. Write the one sentence you will use to report any study in the session, including its sample and the date of the tool tested, and check it against the source before delivery.

Knowledge check

  1. What are the three layers a lecture tends to confuse, and which one can it deliver well?

    Answer guidance: Information (the fact), decision (recognising in a real document the moment the fact applies) and routine (performing the sequence fast enough to survive a deadline). A lecture delivers information efficiently; it does little for decision and nothing for routine.

  2. What is the Monday test?

    Answer guidance: Completing "because of this, on Monday a participant will..." for every element of a session. An outcome ending in know, understand or be aware produces information only; an outcome ending in an action at a desk may produce behaviour if the session also rehearses it.

  3. What did Choi, Monahan and Schwarcz report, and what are its limits?

    Answer guidance: Large, consistent speed gains from GPT-4 with only slight and inconsistent quality gains, and the lowest-skilled participants gaining most. The participants were sixty law students and the model dates from 2023.

  4. Why does a demonstration in which the tool only succeeds fail to prepare lawyers for Monday?

    Answer guidance: Because it teaches that the tool succeeds, and rehearses nothing about recognising and handling the plausible but wrong output that the working lawyer actually has to catch.

Sources and further reading