Lesson objective

Design an in-session practice task on fictional material that reproduces the trigger, the decision and the routine named in a behaviour statement.

Plain-language explanation

Why this matters now

Once the behaviour is named, the next design step is not the slides. It is the practice: the part of the session in which each participant performs the behaviour, or the closest version of it the room allows. Everything else in the session exists to make that practice possible and to make sense of it afterwards.

This is Best practice, argued from the structure of the work rather than from learning research. A lawyer's Monday behaviour is a trigger, a decision and a routine, performed on a real document under real constraints. A session can reproduce a surprising amount of that with fictional material and a clock. It cannot reproduce any of it with a slide.

For an AI session there is a particular reason to design the practice carefully. The error an AI tool makes is often plausible and well integrated into fluent text. A practice task that announces its errors ("find the three hallucinations on this page") rehearses a situation the lawyer will never meet. At the desk, nobody tells you there is something to find.

A question to sit with

Your behaviour statement says a participant will check the passages an AI summary turns on before relying on it. In your practice task, what tells the participant that this is the moment to do it, and is that the same thing that will tell them on Monday?

What goes wrong

1. The error is announced. The task says what kind of error is present and how many. Participants search for it and find it. They have rehearsed searching, not noticing.

2. The material is obviously fake. Cartoon names, a two-line contract, a brief with no structure. A lawyer's pattern recognition, which is what the session is trying to engage, never switches on.

3. The material is real. To make it realistic, the instructor uses a genuine client document with names removed. That is a confidentiality problem, and removing names is not the same as making material fictional.

4. The practice skips the routine. Participants discuss what they would check but never check it. They leave with an opinion about the routine and no experience of its friction.

5. Everyone gets the same result. The task is so tightly controlled that there is only one answer, found in the first minute. The debrief has nothing to work with.

6. The tool always gets it right, or always gets it wrong. Both teach something false. A tool that is always wrong teaches that checking is about finding errors, when most of the time the check confirms the output and must still be done.

The practical answer

Best practice: match the practice to the behaviour statement, part by part.

Behaviour statement part What the practice must reproduce How
Trigger The moment arriving unannounced, inside ordinary work Embed the task in a realistic assignment ("prepare this for the partner by 3pm") rather than a hunt
Decision Recognising that the moment has come Mix material that needs the check with material where the check confirms the output
Observable action Actually doing it, with the real steps Give access to the research service, the policy, the approved-tool list, or a faithful copy
Standard Knowing when it is done Require the same record the firm requires: the verification note, the review entry
Instead of The pull of the current habit Include realistic time pressure, so the shortcut is tempting

Best practice: seed the material honestly. Most AI-assisted output a lawyer reviews is partly right. The practice material should be too.

  • Accurate content that the check will confirm. This teaches that verification is a routine, not a hunt.
  • A plausible error that is well integrated: a quotation not in the source, a date shifted, an obligation that the document does not contain, a standard stated for the wrong jurisdiction.
  • A fabricated authority, where the behaviour concerns citations, always labelled in the facilitator's copy and in the debrief as fictional. Where the participant's copy contains a citation to an authority that does not exist, the instructor's notes carry the label exactly: (a fictional citation invented for this lesson; it does not exist).
  • A judgement call where reasonable lawyers might differ, so the debrief has something to discuss.

Best practice: do not let a study shortcut the practice. It is tempting to design practice around the idea that some tools make these errors and others do not. Two studies are worth knowing, and both must be reported with their limits.

Research finding: Choi, Monahan and Schwarcz, 109 Minnesota Law Review 147 (2024), reported large, consistent speed gains from GPT-4 on legal tasks but only slight and inconsistent quality gains, with the lowest-skilled participants gaining most. Sixty law students; a 2023 model.

Research finding: Schwarcz, Manning, Prescott, Barry, Cleveland and Rich, in "AI-Powered Lawyering", first posted on SSRN on 2 March 2025 and published in the Journal of Law and Empirical Analysis in 2026, reported a randomized trial in which a retrieval-augmented legal tool and a reasoning model each produced productivity gains of 50% to 130% in five of six tasks and improved quality, with the retrieval tool adding no hallucinations. The participants were law students and the tools date from 2024.

Neither study says the check can be skipped. The second tells you that one retrieval tool, in one trial, on those tasks, added no hallucinations. It does not tell you about the tool your participants use, on their tasks, this year. The practice rehearses the check whatever tool is in use, because the lawyer at the desk cannot know in advance which output is the exception.

Best practice: design for four task shapes, and pick the one closest to the behaviour.

Shape What the participant does Suits behaviours about
Would you send this? Reviews a realistic AI-assisted deliverable and decides what must happen before it goes out Verification and review
Route it Receives several pieces of material and decides which tool, if any, each may go into under the policy Confidentiality and tool choice
Do the routine Performs the check or the recording from start to finish, against the clock Any routine where friction is the obstacle
Tell the supervisor Writes the note a supervising lawyer would need about how AI was used Supervision and communication

Best practice: sequence the practice so it can fail privately. Individual attempt, then comparison in pairs, then a group debrief about what was found and how. The debrief discusses the material, not who missed what. Module 4 develops the facilitation; Module 6 develops full labs. A single well-designed practice task of fifteen to twenty minutes is enough for most sixty-minute sessions.

Remember this

  • The practice is designed before the slides, and every other element of the session serves it.
  • Reproduce the trigger, the decision, the routine, the standard and the pull of the current habit.
  • Never announce the error, and include accurate content so checking is rehearsed as a routine rather than a hunt.
  • Practice material is invented, never anonymised client material.
  • No study licenses skipping the check; the lawyer at the desk cannot know which output is the exception.

Legal example

Continuing at the fictional Corriveau Asplund LLP, Neve Halloran designs the practice for her paralegal session, "Before You Paste". Her behaviour statement says a paralegal will use only the approved tool, for material the policy permits, and note on the task that AI was used and for what. She chooses the "route it" shape, followed by a short "do the routine". She writes the material from scratch. Each participant receives a fictional assignment email from a fictional supervising partner, Dorian Velasquez-Kent, asking for four things by 3pm on a fictional freight contract dispute for an invented client, Harrowgate Mills: a summary of a forty-page fictional deposition transcript, a list of dates from a fictional email chain, a tidy-up of a fictional internal memo, and a first draft of a cover letter to opposing counsel. Nothing in the email mentions AI. The participant's packet includes the firm's newly revised policy with its three-line table of permitted material, and a laptop login to a training copy of the approved tool. Two of the four tasks may properly go into the approved tool under the policy. One contains material the policy restricts. One, the cover letter, is a judgement call on which reasonable paralegals disagree, because it contains no client material yet but will. A fifteen-minute clock runs on the screen. Participants work alone for ten minutes, compare with a partner for five, and then the group debriefs. Two of the fourteen participants route the restricted material into the tool, and both say afterwards that the email's deadline made them skip the policy. Neve does not identify them. She puts the deadline on the screen and asks the room what would make the policy check fast enough to survive it. The room's answer, a one-line prompt at the top of the approved tool, goes to the IT director that afternoon.

Responsible-use warning

Never build practice material from a real client document, even with names removed; redaction leaves facts and context that can still identify a client. Never announce the number or kind of errors in a task, and never design material in which the tool is always wrong. And never tell participants that a particular tool does not hallucinate on the strength of a single study of law students using earlier tools. The check is rehearsed regardless of the tool.

Human-review requirement

A licensed attorney reviews all practice material before use and owns its legal accuracy. The reviewer confirms that the material is invented rather than anonymised, that every fabricated citation and matter name is labelled in the facilitator's notes as fictional and non-existent, that the correct answers reflect the firm's current policy and the participants' jurisdiction, that any judgement call is genuinely arguable, and that studies mentioned in the debrief are described with their samples and dates.

Practical exercise

Design one practice task for the behaviour statement you wrote in the previous lesson. Use only invented material.

  1. Copy your behaviour statement onto the top of the page.
  2. Choose the task shape closest to the behaviour: would you send this, route it, do the routine, or tell the supervisor.
  3. Write the realistic assignment that carries the trigger without announcing it, such as a supervising lawyer's email with a deadline.
  4. Write or outline the material. Include at least one accurate item the check will confirm, one plausible and well-integrated error, and one genuine judgement call.
  5. If the material contains a citation to an authority that does not exist, label it in your facilitator notes exactly: (a fictional citation invented for this lesson; it does not exist).
  6. List what the participant needs access to in order to perform the routine for real: a research service, the policy, the approved-tool list, or a faithful copy.
  7. Specify the record they must produce to meet the standard.
  8. Set the timing for the individual attempt, the pair comparison and the debrief.
  9. Write three debrief questions about the material, none of which asks who missed something.
  10. Try the task yourself against the clock and adjust the timing.

Knowledge check

  1. Why should a practice task never announce how many errors the material contains?

    Answer guidance: Because at the desk nobody tells a lawyer there is something to find; announcing the errors rehearses searching rather than noticing, which is the decision the behaviour actually depends on.

  2. Why should verification practice include accurate content the check will simply confirm?

    Answer guidance: Because most real AI-assisted output is partly right, and the habit being built is a routine performed every time, not a hunt that ends when an error is found; finding nothing wrong and recording the check is a correct outcome.

  3. What did Schwarcz and colleagues report about a retrieval-augmented tool, and why does it not justify skipping verification practice?

    Answer guidance: In a randomized trial with law students using 2024 tools, a retrieval-augmented legal tool and a reasoning model produced productivity gains of 50% to 130% in five of six tasks and improved quality, and the retrieval tool added no hallucinations. It describes one trial and those tools; it says nothing about the participant's tool and tasks now.

  4. What is the difference between fictional and anonymised practice material?

    Answer guidance: Fictional material is invented from scratch; anonymised material is a real document with identifiers removed, which can still reveal a client through its facts, structure and context. Practice material must be fictional.

Sources and further reading