Lesson objective

Decide which explanation a practice task genuinely needs, where it sits in the session, and how to state rules and research within it accurately.

Plain-language explanation

Why this matters now

With the behaviour named and the practice designed, the instructor finally writes the explanation. Done backwards, this is a much smaller job than most instructors expect. The question is no longer "what do I know about this subject?" It is "what does a participant need in order to attempt this practice, and to make sense of it afterwards?"

This is Best practice, and its logic is economy. A sixty-minute session with a twenty-minute practice and a fifteen-minute debrief leaves perhaps fifteen minutes for everything else. Every minute of explanation that the practice does not need is a minute taken from the part of the session that rehearses the behaviour.

Explanation also carries the highest accuracy risk. It is where the rules, the guidance and the research are stated, and where an instructor who states something wrongly states it to everyone at once. Module 3 treats accuracy and currency in depth. This lesson is about deciding what to say and saying it precisely.

A question to sit with

Look at the explanation you would normally give before a practice task on AI verification. Which sentences would a participant actually need in order to attempt the task, and which are there because you find them interesting or because they make you sound prepared?

What goes wrong

1. Explaining everything first. The instructor front-loads all the context "so they are ready". Participants arrive at the practice tired, and the explanation contained answers that would have been worth discovering.

2. Explaining nothing. Having heard that practice matters, the instructor drops participants into a task without the one rule or route they need. They flounder, and the debrief becomes a lecture after all.

3. The history of AI. A tour of how language models work that no step of the practice depends on. Some understanding of why tools produce plausible errors is useful; the full tour rarely is.

4. Rules stated without their shape. "The rule says you must..." without the jurisdiction, without saying it is a model provision, without the exception. A compressed explanation is especially prone to this, because the caveats are what get cut.

5. Research stated as a headline. One sentence from a study, without the sample or the date of the tool, becomes the thing participants remember and repeat.

The practical answer

Best practice: list what each step of the practice requires. Walk through the practice task as a participant would and, at each step, ask what they must already know or have.

Practice step What the participant needs Type of explanation
Recognise that the draft contains an authority that must be checked Why an AI tool can produce a confident citation to something it has not seen Failure pattern — two minutes, one example
Decide whether the material may go into the tool Where the firm's approved-tool list and policy table are Route — one slide or handout
Perform the check How to open the authority in the research service and what to compare Routine — a thirty-second walkthrough, or none if they already know
Record it Where the verification note goes and what it contains Standard — the template itself
Understand why it matters The duty behind the behaviour, with its caveats Rule or guidance — two minutes, precisely labelled

Anything not on this list is a candidate for cutting or for a handout.

Best practice: decide where each piece sits. There are three places for explanation, and each suits different content.

  • Before the practice: only what a participant cannot attempt the task without. Usually the route and the standard.
  • During the practice: available on request, on a handout or from the instructor. Often the routine.
  • After the practice, in the debrief: the failure pattern and the rule, named once participants have met the thing they explain. An explanation of why a tool invents quotations is far more concrete to a lawyer who has just found one.

Putting the failure pattern after the practice also protects the practice. If the explanation describes the exact error in advance, the practice becomes a hunt for it.

Best practice: state rules in a fixed shape. Compressed explanations need a template, because that is where caveats disappear.

  1. Label and source: "Law / rule: ABA Model Rule 1.4..." or "Professional guidance: ABA Formal Opinion 512, 29 July 2024..."
  2. What kind of authority it is: a model provision, an adopted rule of a named jurisdiction, or guidance that does not bind by itself.
  3. What it says, narrowly: only as much as the behaviour needs.
  4. The jurisdictional instruction: "Your jurisdiction adopts its own version; here is where to check it" or "our firm is in X, and its adopted rule is..." once confirmed.
  5. The link to the behaviour: why this duty makes the routine you are about to practise necessary.

For example, for a session on recording AI-assisted time: Professional guidance: ABA Formal Opinion 512 (29 July 2024) addresses fees for AI-assisted work, including billing for time actually expended. It interprets the ABA Model Rules, which each jurisdiction adopts in its own form, and your jurisdiction may have its own rules and guidance on fees. Check it. That is why the routine records time actually spent.

Best practice: state research in a fixed shape, and only when the practice needs it. Research earns a place in the explanation when it changes what a participant does or expects. The two legal-AI trials in this module's sources are a good example of research that must be explained together, because either one alone gives a misleading picture.

Research finding: Choi, Monahan and Schwarcz (109 Minnesota Law Review 147, 2024), sixty law students with a 2023 model, reported large and consistent speed gains but only slight and inconsistent quality gains, with the lowest-skilled participants gaining most.

Research finding: Schwarcz, Manning, Prescott, Barry, Cleveland and Rich (SSRN, 2 March 2025; Journal of Law and Empirical Analysis, 2026), law students with 2024 tools, reported productivity gains of 50% to 130% in five of six tasks from a retrieval-augmented legal tool and a reasoning model, with improved quality and no hallucinations added by the retrieval tool.

The accurate explanation: the studies used different tools at different times, and neither was conducted on practising lawyers. An instructor should not claim to know which difference explains the different quality results. What a participant can take from both is that faster is not the same as more reliable, that results depend on the tool and the task, and that neither study removes the need to check.

Best practice: write the explanation last and time it. Draft each piece, read it aloud with a timer, and add up the minutes. If the explanation takes more time than the practice, return to the list of what each step requires and cut whole topics.

Remember this

  • Write the explanation last, from a list of what each practice step requires.
  • Before the practice goes only what participants cannot attempt it without; the failure pattern and the rule often land better in the debrief.
  • Keep explanations short by cutting whole topics, never by cutting labels, jurisdictions or caveats.
  • State rules in a fixed shape: label and source, kind of authority, what it says narrowly, the jurisdictional instruction, and the link to the behaviour.
  • Explain the two legal-AI trials together, with samples and tool dates; neither removes the need to check.

Legal example

The fictional litigation boutique Strathmore Fenwick PLLC runs a sixty-minute session whose behaviour statement reads: "When an AI-assisted draft contains any cited authority and is about to go to a supervising lawyer, the associate who prepared it will open each authority in the firm's research service, compare the cited passage with the draft, and record the check on the matter file, instead of relying on the citation looking correct." The practice is a twenty-minute "would you send this?" task on a fictional brief. The instructor, Yusra Delacourt-Mbeki, a fictional senior associate, first drafts thirty-five minutes of explanation: how language models work, the history of legal research tools, a survey of sanctions stories, the full text of three model rules, and the two legal-AI trials. She then walks through the practice step by step and lists what each requires. Participants need to know where the verification note goes and what it contains, which becomes a one-page template handed out before the practice. They need the research service open, which they already know how to use. They need nothing about language models before the task. After the task, in the debrief, they need two minutes on why a tool can produce a confident quotation from an authority it has never seen, illustrated by the one they have just found, which in her notes carries the label (a fictional citation invented for this lesson; it does not exist). They need two minutes on the duty, stated in the fixed shape: the model competence rule and its technology comment, described as model provisions that each jurisdiction adopts in its own form, followed by her firm's state and where its adopted rule can be read. Her slide on the two trials reports both with their samples and dates, and says plainly that neither was conducted on practising lawyers and neither removes the need to check. Her final explanation totals nine minutes. The history of research tools becomes a further-reading note.

Responsible-use warning

Do not shorten an explanation by removing the qualifications from a rule or a study. Removing the jurisdiction, the fact that a model rule is a model provision, or the sample and date of a trial turns a short explanation into an inaccurate one, and every participant will repeat it. Do not describe in advance the exact error a practice task contains. And do not claim to know why two studies reached different results when neither was designed to answer that.

Human-review requirement

A licensed attorney reviews every statement of a rule or guidance in the explanation before delivery. The reviewer confirms that each is labelled, identifies the kind of authority, states only what it supports, names where the participant's jurisdiction's adopted rule can be checked, and does not imply that guidance binds. The reviewer also confirms that research is described with its sample, the date of the tools tested and its full result, and that nothing presents the named-reviewer standard as law.

Practical exercise

Cut an explanation to what the practice needs. Use the behaviour statement and practice task from the previous lessons.

  1. Write out everything you would normally explain in the session, as a list of topics with estimated minutes.
  2. Walk through your practice task step by step. For each step, write what the participant must already know or have.
  3. Classify each requirement as a failure pattern, a route, a routine, a standard, or a rule or guidance.
  4. Delete every topic from step 1 that no step requires. Move useful but unnecessary material to a handout or further-reading note.
  5. Place each remaining piece before, during, or after the practice, and justify any piece placed before it.
  6. Write any rule statement in the fixed shape: label and source, kind of authority, what it says narrowly, the jurisdictional instruction, and the link to the behaviour.
  7. If you use either legal-AI trial, write the sentence reporting it with authors, sample, tool date and full result.
  8. Read the whole explanation aloud against a timer. If it exceeds the practice time, cut again.
  9. Ask a licensed attorney to confirm every rule statement before the session is delivered, and record who confirmed it and when.

Knowledge check

  1. How does an instructor decide what explanation a session needs?

    Answer guidance: By walking through the practice task as a participant would and listing, for each step, what the participant must already know or have; anything no step requires is cut or moved to a handout.

  2. Why is the explanation of a failure pattern often better placed after the practice than before it?

    Answer guidance: Because it is more concrete to a lawyer who has just met the error, and because describing the exact error in advance turns the practice into a hunt instead of a rehearsal of noticing.

  3. What are the five parts of the fixed shape for stating a rule?

    Answer guidance: Label and source; what kind of authority it is; what it says, narrowly; the jurisdictional instruction; and the link to the behaviour being practised.

  4. What can a participant accurately take from the Choi (2024) and Schwarcz (2025/2026) trials together?

    Answer guidance: That faster is not the same as more reliable, that results depend on the tool and the task, and that neither removes the need to check. Both used law students, with a 2023 model and 2024 tools respectively, and neither explains the difference between their quality results.

Sources and further reading