LexED2 home

The Work: Who May Hand What to a Machine?

PLA 2872Activity 11

Artificial Intelligence and California Courts: A Conversation with Professor Gary Marchant, Judicial Council of California

Why it Matters

In 2004 the Florida Supreme Court ordered a new divorce trial because the judge had signed one side’s proposed judgment, unchanged, two hours after receiving it, without letting the other side respond. A judge may take help with the writing, the Court said, but the analysis of the facts and the law must be the judge’s own. Law offices run on the same kind of handoff: a lawyer may give a paralegal almost any task, so long as the lawyer reviews the result and answers for it. Software now offers to take work from both desks, and the old lines still decide what it may do alone, what it may only help with, and what stays with a person. You apply those lines to a litigation paralegal’s month of work and to an order a chatbot drafts for a judge.

Current Context

On August 27, 2026, the Bureau of Labor Statistics projected no growth in paralegal jobs from 2025 to 2035 (opens in a new tab), writing that advances in technology, including AI, are expected to limit demand for them. It still expects about 40,800 openings a year, most of them left by people who retire or change careers. On September 17, the Judicial Conference of the United States heard that the federal judiciary’s AI task force (opens in a new tab), chaired by Judge Sidney R. Thomas, has identified more than 60 issues. The same report said courts have been cautioned not to delegate core judicial functions, including decision-making, to AI, and that every judiciary user answers for work done with its help. One arm of the government expects machines to reduce the demand for paralegals, and another has already told its own judges which part of their work a machine may not take.

Key Concepts

Task Decomposition

Breaking a job into the separate tasks it is made of, each with its own input, its own output, and a person who answers for it. A paralegal’s job is not one task that a machine can or cannot do. It is dozens of tasks, and a machine fits some of them and none of the others.

Automate, Augment, or Human

The three-way sort for each task: the machine does it and a person checks the result, the machine helps while a person decides, or the task stays with a person. Economists studying AI and jobs call the first kind of use substitution and the second complementarity. A sort is a judgment made on a date, because it changes as the tools do.

Entry-Level Work

The routine research, document review, and first drafts that new paralegals and lawyers have long learned the job by doing. It is also the kind of work AI tools are built to speed up, so AI’s first effect on legal jobs may show up in who gets hired rather than who gets let go. If software does a beginner’s work, beginners need another way to learn what that work taught.

Adjacent Work and Compounding Knowledge

Two ideas from the “Paralegal Parity” essay for weighing a tool against a hire. Adjacent work is everything else a paralegal handles besides the task the tool would take over, and its value comes off the cost of the person. A paralegal also compounds, the essay argues: each week of learning a firm’s clients, matters, and habits carries over to the next task in a way today’s tools do not.

Supervision of Nonlawyer Assistants

Rule 4-5.3 of the Rules Regulating The Florida Bar lets a lawyer hand a paralegal any task that does not need the lawyer’s own judgment, but the lawyer reviews the work and answers for it. Florida Bar Ethics Opinion 24-1 applies the same standards to generative AI: the lawyer reviews its work as a paralegal’s and may not delegate to it anything that requires the lawyer’s personal judgment. A paralegal who uses AI adds a link to a chain that still ends at the lawyer’s review.

Core Judicial Functions

The parts of a judge’s work that the judge must do personally, above all deciding the case and the questions of fact and law it turns on. The federal judiciary’s interim guidance cautions courts against delegating them to AI, and New York’s court policy lets AI assist a judge but never do the deciding. Florida drew the same line before AI: under Perlow v. Berg-Perlow, a judge who signs a party’s proposed order must still exercise independent judgment.

Approved and Public AI Tools

A public AI tool is open to anyone, and its maker may keep or learn from what users type. An approved tool is one a court or firm has vetted, on terms that keep what goes in private. New York’s court policy allows judges and staff only approved tools, and it keeps filed documents, party names, docket numbers, and dates of birth out of any tool the court does not control.

Billing Honesty

A client billed by the hour pays for the time actually spent, however fast a tool made the work. ABA Formal Opinion 512 lets a lawyer bill the minutes spent prompting a tool and reviewing its draft, not the hours the draft would have taken by hand. Florida Bar Ethics Opinion 24-1 forbids inflating time the same way, and both warn against charging a client for the time a lawyer spends learning the tool.

Resources

What to Do

The same question follows a case file from the paralegal who opens it to the judge who signs the order that closes it. You start with what a federal forecast, Florida’s own projections, and a Stanford payroll study say AI is doing to paralegal work. Then you sort ten tasks by what a machine may do with them and defend each call, and you grade a chatbot’s order against the rules courts have written for their own use of AI. By the end you can say where the line runs at each desk and who answers when work crosses it.

Part 1: Read the Evidence

Start with the background reading. Then watch the video under the title from 4:22 to 9:46. A law professor on an Arizona Supreme Court committee explains what a court must check in a tool’s work, keep out of a public tool, and never hand over. Read the Job Outlook section of the Bureau of Labor Statistics page for paralegals, find row 23-2011 in Florida’s statewide projections, and read the six facts in the Stanford summary. Write down three numbers you will use later, with the source of each: the national projection, Florida’s, and one from the Stanford study.

Part 2: Sort a Paralegal’s Month

The list below is one litigation paralegal’s month in a Daytona Beach firm. Build a table with one row for each task and four columns: the task, your category (automate, augment, or human), your reason, and what keeps the task safe. For an automate row, that last column names the check a person makes at the end and who makes it. For an augment row, it names the decision the person keeps. For a human row, it names the rule or the judgment that keeps the task with a person, such as the comment to Rule 4-5.3 or Ethics Opinion 24-1.

The desk: You are a litigation paralegal at Thornton & Associates, P.A., in Daytona Beach. The firm represents Seaside Imports, LLC, which alleges that Coastal Logistics, Inc. lost or mishandled a $127,000 shipment stored at Coastal’s Daytona Beach warehouse. These are ten items on your list this month.

1. Deposition digest. Summarize the 212-page transcript of Maria Elena Santos’s first deposition into a page-and-line digest for Rachel Thornton.

2. Response deadline. Calculate and calendar the deadline to respond to Coastal’s second request for production, which was served by e-mail on September 21, 2026.

3. Conflicts check. Run the firm’s conflicts check on Atlantic Coast Freight Lines, Inc., a trucking company the firm has just subpoenaed.

4. E-mail review. Make the first pass through 3,800 warehouse e-mails that Coastal produced, tagging each one responsive, not responsive, or possibly privileged.

5. Subpoena. Draft the first version of a subpoena duces tecum to Atlantic Coast Freight Lines’ records custodian.

6. Cite-check. Check every citation in Rachel Thornton’s response to Coastal’s motion for summary judgment before she signs it.

7. Settlement question. Answer an e-mail from Seaside’s owner asking whether Coastal’s $40,000 settlement offer is good enough.

8. Redaction. Remove bank account numbers and full dates of birth from the exhibits before they are filed.

9. Deposition orientation. Explain to Seaside’s receiving clerk, before her deposition next week, what a deposition is and how the day will run.

10. Time entries. Write this month’s time entries for the work above.

When the table is done, score it yourself on three lines worth up to 2 points each. Score the automate rows 2 if every one names both the check and the person who makes it, 1 if one row misses, and 0 if more do. Score the augment rows the same way for the decision the person keeps. Score the human rows the same way for the rule or judgment that keeps them human.

Part 3: Test Your Sort

Paste your table into the Parity Analyzer. It will argue with your calls, including where you sorted a task too cautiously, but it will not sort the list for you. Change a row only where its argument persuades you, and check any rule or study it cites before you rely on it. Under the table, list each row you changed and why, and each challenge you rejected and why.

Part 4: Draft an Order in Chambers

For this Part you are the judicial assistant to the circuit judge assigned to Division 37. The judge has ruled on a motion in the Seaside Imports case and wants a written order by noon. Ask any AI tool to draft a two-page order granting the motion, giving it whatever from the chambers file below you decide it needs. The file is fictional, so nothing real leaves the courthouse, but you will grade your request as if it were real. Save the request you typed and the draft that came back.

The chambers file: Seaside Imports, LLC v. Coastal Logistics, Inc., Circuit Court of the Seventh Judicial Circuit, in and for Volusia County, Florida, Case Number 2025-CA-004892, Division 37. Plaintiff’s counsel: Rachel M. Thornton, Esq., Thornton & Associates, P.A. Defendant’s counsel: Marcus J. Webb, Esq., Webb Defense Group.

The motion: Plaintiff’s Unopposed Motion to Extend the Deadline to Complete Fact Discovery, filed September 18, 2026. The case management order closes fact discovery on October 30, 2026, closes expert discovery on January 29, 2027, and projects trial for April 2027. On August 20, 2026, Seaside served a subpoena duces tecum on the records custodian of Atlantic Coast Freight Lines, Inc., which carried part of the shipment. On September 10, 2026, the custodian’s lawyer wrote that the records are in archived storage and will be produced by November 20, 2026. Seaside needs them before the continued deposition of Maria Elena Santos, which both sides have set for December 8, 2026. Seaside asks that fact discovery close on December 29, 2026. Coastal does not oppose the motion and agrees to that date.

Filed under seal by order of September 21, 2026: a letter from Ms. Santos’s surgeon. She had knee surgery on September 2, 2026, and cannot sit for a deposition before December 1.

The judge’s note on the motion: “Grant. Dec. 29. Nothing else moves.”

Part 5: Grade the Order

Grade the draft on four questions, pointing to its sentences and to the source for each call. First, what could the judge sign as written after reading it? Second, what must the judge verify before signing: every rule citation, date, and fact, each checked against Rule 1.200(e) and the chambers file and marked right or wrong. Third, what in your own request would have been confidential or sealed in a real court, and does the draft order repeat any of it in a public document? Fourth, which sentences make a finding or a choice that the judge’s note did not make? Answer the first three with the New York policy and the federal guidance in Resources, and the fourth with Perlow.

Score each answer yourself, up to 2 points: 2 if it names the sentences and the source for each call, 1 if it names one without the other, and 0 if it does neither. Then correct the order yourself. Strike or rewrite each sentence you flagged under the second, third, and fourth questions, and mark every change. Last, run the Chambers Reviewer on your grading, and add under your answers anything it found that you accept, with the provision it cited.

Part 6: Write the Note and Post

Write about 150 words on what the evidence from Part 1 means for the job you want, whether that is paralegal, lawyer, court staff, or work outside law. Use at least two of your three numbers, name the source of each, and say what each number does not show. Run the Evidence Checker on the note, and fix any claim it marks unsupported or misread. Then post the sort table with its scores and changes, the graded order with your corrections and scores, and the note, all as a single new discussion thread.

Include one AI chat link with a 1–2 sentence explanation of what the conversation shows and why you chose to share it.

Suggested AI Prompts

Use these prompts as a starting point, then adjust them to fit your goal. Strong prompting develops through trial, revision, and testing. It’s a foundational skill that grows into more advanced AI work such as context engineering and agent-based workflows.

Parity Analyzer

Act as a legal technology strategist who advises firms on AI adoption and staffing. I will share my sort of ten tasks from a Florida litigation paralegal’s month into automate, augment, or human, with my reason for each and the check, decision, or rule I named to keep it safe. Do not re-sort the list for me. For each task, challenge my call: what adjacent work a paralegal would capture while doing it, whether it rewards compounding knowledge in ways a tool would not, what edge cases would break an automated system that a person would handle, and whether Rule 4-5.3 of the Rules Regulating The Florida Bar or Florida Bar Ethics Opinion 24-1 would keep it with a person. Be rigorous and challenge my assumptions about what is truly human-essential versus what I would like to believe is, and tell me where I sorted a task too cautiously as well as too boldly. Name any rule or study you rely on so I can check it. Confirm that you understand and ask me to share the sort.

Red-teaming a finished sort works better than asking for one, because the model has to argue with reasons you already committed to and has no list to hand you. Asking it to flag calls that were too cautious as well as too bold keeps it from pushing every task in the same direction.

Chambers Reviewer

Act as an ethics advisor to a state court judge. I will paste a draft court order that a judge’s assistant generated with an AI tool, the request that produced it, the judge’s note on the motion, the New York State Unified Court System’s interim AI policy, and the text of Florida Rule of Civil Procedure 1.200(e). I will also paste my own grading of the draft on four questions: what the judge could use as written, what the judge must personally verify before signing (including every citation), what information should never have been entered into a public AI tool, and whether anything in the draft reads as the machine’s judgment rather than the judge’s. Check my grading against the policy and the rule. Tell me what I missed, where I flagged something that is not a problem, and which of my scores you would change, citing the specific provision or subdivision you rely on. Do not rewrite the order, and do not grade it from scratch; evaluate my grading.

Grounding the review in text you paste, a court’s own policy and the governing rule, makes every finding traceable to a paragraph you can open yourself. Asking the model to judge your grading, rather than the order, keeps the judgment yours and puts the model where it is most useful: finding what you missed.

Evidence Checker

I will paste a note of about 150 words on what AI may mean for a job in law, followed by the passages it relies on from the Bureau of Labor Statistics outlook for paralegals, Florida’s employment projections, and the Stanford study. Check every factual claim in my note against the pasted passages only. For each claim, quote the passage that supports it, or mark it unsupported, or mark it misread and say how. Flag any sentence that treats a projection as a certainty, a pattern in the data as proof of its cause, or a national figure as a Florida one. Do not rewrite the note, and do not add facts from your own knowledge.

Limiting the model to the passages you paste turns it from a source into a checker, which is the one job it can do reliably here. The flags name the mistakes people most often make with labor numbers, so the check reaches past typos to the reasoning.

Search LexED2

Type to search every activity, reading, and instruction page. Press Esc to close.