Lesson objective
Trace the specific failure in which a firm has a policy, competent people and good intentions, and still produces an unowned output, and build the controls that close each gap.
Plain-language explanation
Why this matters now
The failure this lesson describes is not exotic and it does not require anyone to behave badly. It requires only that a firm have a policy, several competent people, and no single point at which somebody is required to say the words "I own this."
Law / rule: Johnson v. Dunn, 792 F. Supp. 3d 1241 (N.D. Ala. 2025), is a sanctions order of the United States District Court for the Northern District of Alabama. Three lawyers at a firm that had an artificial-intelligence policy were publicly reprimanded, disqualified from the case and referred to the state bar after filings contained hallucinated citations. It is a federal trial-court decision and it binds only its parties; it is not precedent for your court and not a rule anywhere. Use it for exactly one proposition, which is the one that matters at this level: a policy nobody ran on the document protected nobody. The firm had the instrument. The instrument was not applied to the filing. The individuals answered for it.
That is the whole of this lesson at firm level. Everything that follows is an anatomy of how a firm with a policy arrives at an output that nobody ran it on.
Professional guidance: ABA Formal Opinion 512 (29 July 2024) treats managerial and supervisory obligations and firm policies as part of how competence and supervision are discharged with generative AI. Law / rule: Model Rules 5.1 and 5.3 are the underlying provisions, and they are model provisions — your jurisdiction adopts its own numbering and wording and that is the one you answer to.
A question to sit with
Take the last AI-assisted document your firm sent out under pressure. Name the person who would say "I own this" if asked in a deposition. Now ask whether that person knows they would be the answer.
What goes wrong
The unowned output is produced by five gaps, and a firm usually has at least three of them at once.
1. The assumption chain. The associate assumes the supervising partner will check it because partners check things. The partner assumes the associate checked it because the associate is good and the firm has a policy requiring verification. The practice-group leader assumes the partner ran the process. The programme owner assumes the practice group is following the procedure because nobody has escalated anything. Every link in the chain is individually reasonable, and the chain as a whole reaches the client with nobody in it having verified anything.
2. The handoff with no receipt. Work crosses between people — associate to partner, one office to another, the day team to whoever is finishing at nine in the evening. At each crossing, ownership is presumed to transfer. Nothing records that it did, and the presumption runs in opposite directions at either end.
3. The policy that exists at a different altitude from the work. The instrument says AI-assisted output must be verified. The person producing the document at ten at night is not reading the instrument; they are working inside a matter, a template and a deadline. Unless the requirement appears inside the workflow they are actually in, it is not a control. It is a statement of intent stored elsewhere.
4. Urgency dissolving the process. Every AI governance failure worth studying has a deadline in it. Under time pressure the steps that get dropped are precisely the ones whose value is invisible when they succeed, and verification is the purest example: it consumes time and, when the output was fine, produces nothing anyone can see.
5. Nobody is required to say it out loud. Diffuse ownership survives because it is never tested by a moment that forces a person to claim the work in the first person. A tick box does not force it. A name typed into a field by somebody else does not force it. A sentence the owner has to read and affirm does.
The practical answer
Convert each gap into a specific control. This table is the deliverable of this lesson.
| Gap | Control | Evidence it produces |
|---|---|---|
| Assumption chain | Each AI-assisted deliverable has exactly one named licensed owner recorded before it leaves the firm, and the field cannot be completed by anyone other than that person | A name, with an authenticated actor |
| Handoff without receipt | Ownership transfer is an explicit act with an acknowledgement; until it is acknowledged, ownership stays where it was | A transfer log with two names and a time |
| Policy at the wrong altitude | The verification requirement appears inside the matter workflow, template or filing checklist, not only in the instrument | A completed in-workflow step |
| Urgency dissolving process | A written rule that the verification step is never the thing that gives way, and a defined emergency route that changes who approves rather than whether verification happens | A record of the emergency route being used, and how often |
| Nobody says it out loud | An affirmation sentence in the first person that the owner reads and confirms at sign-off | A timestamped affirmation |
Design the emergency route before you need it. Best practice: a firm that has no defined way to move fast will improvise one, and the improvisation will always be to skip the check. Write the alternative: at short notice, the scope narrows (fewer AI-assisted components, more read in full), the approver may change to whoever is genuinely available at the required seniority, and the record notes that the emergency route was used. Then count how often it is used. A route used weekly is not an emergency route; it is your actual process, and the firm should govern it as one.
Test for unowned work rather than waiting for it to announce itself. Best practice: take ten AI-assisted deliverables from the last quarter at random and, for each, ask three people who owned it, without letting them consult each other or the file. Count the deliverables where all three name the same person. That number, out of ten, is your ownership rate, and it is the single most informative figure a firm-level programme can produce in an afternoon.
Use the Map function of a management framework to find the unowned uses, not just the unowned outputs. Best practice, drawn from the NIST AI Risk Management Framework 1.0 (26 January 2023), a voluntary framework that neither creates nor discharges any professional duty: inventory where AI is actually in use, including features embedded in products the firm already licenses. An embedded summarisation feature in a document-management platform is an AI use that nobody chose, nobody assessed, and nobody owns, and it will not appear in any list built by asking which AI products the firm bought.
Remember this
- Law / rule: Johnson v. Dunn (N.D. Ala. 2025) is a federal trial-court sanctions order binding only its parties; three lawyers at a firm that had an AI policy were publicly reprimanded, disqualified and referred to the state bar over hallucinated citations. A policy nobody ran on the document protected nobody.
- An instrument states what should happen; a control stops the work until it has happened and leaves a record.
- The unowned output comes from five gaps: the assumption chain, the unacknowledged handoff, the policy at the wrong altitude, urgency, and nobody having to say it out loud.
- Define the emergency route in advance so that speed changes the approver or the scope, never whether verification happens — then count how often it is used.
- Measure your ownership rate: ask three people who owned ten random deliverables, and see how often all three agree.
Legal example
Stroud Amankwah LLP is a fictional 45-lawyer firm with a written AI policy adopted a year earlier, acknowledged by every lawyer, and reinforced by an annual training session with a completion log. On a Thursday, an invented client calls at four in the afternoon about an injunction application that must be filed by noon on Friday. The matter partner, a fictional litigator named Reuben Stavros, is travelling and takes the call from an airport. He asks a third-year associate, the invented Nkechi Olamide, to prepare the supporting brief overnight and says he will review it in the morning. Nkechi uses the firm's approved research assistant, which is permitted under the policy for first-pass research, and drafts a brief containing nine authorities. She verifies six of them against the primary source. She runs out of time on three and leaves a note in the draft reading "3 cites still to confirm", which she expects Reuben to see. At six the next morning Reuben reads the brief on a train, thinks it is good, sees the note but reads it as a routine drafting marker rather than a stop, and asks a paralegal to finalise and file. The paralegal, the fictional Tomas Widdicombe, formats, converts and files at eleven. Two of the three unverified authorities exist and say roughly what the brief claims. The third does not exist. Reconstructing it afterwards, the firm finds four people involved and not one gap in competence. Nkechi flagged the issue and expected review. Reuben reviewed for substance and assumed the flag was drafting shorthand. Tomas performed a formatting task and had no basis to think verification was his. The practice-group leader had no visibility at all. The policy said in terms that every authority in a filing must be verified against the primary source, and it was acknowledged by two of the four. Nothing in the firm required anybody to say, in their own name, that verification was complete before the document could be filed. The remedy the firm adopts is deliberately narrow: the filing checklist gains a required line, completed by a named admitted attorney, reading that every authority has been opened in the source and confirmed, with the checker named per authority; the document cannot be filed without it; and the emergency route, when time is short, reduces the number of authorities rather than the checking of them.
Responsible-use warning
Never let urgency decide which step is dropped, because it will always drop verification. Verification consumes visible time and, when the output was sound, produces nothing anyone can point to, which makes it the cheapest-looking thing to cut at nine in the evening before a morning filing. Define the emergency route in advance so that short notice changes the scope of what is attempted or the identity of the approver, never whether the check happens. Then count how often the route is used. If it is used most weeks, it is not an emergency route, it is your real process, and it needs to be governed as one.
Human-review requirement
A licensed attorney owns every output that reaches a client or a court, and in this failure mode the whole point is that the attorney must be required to say so in their own name before the work moves. The reviewing attorney confirms that every authority, quotation and factual assertion the document turns on has been checked against the primary source by an identified person on an identified date, that any marker left in a draft has been resolved rather than interpreted, and that they are personally satisfied, not that the process appears to have run. Where the attorney cannot give that confirmation before the deadline, the professional answer is to narrow what the document asserts or seek more time, not to file and rely on the process.
Practical exercise
Find the unowned outputs in an invented firm, then in your own. Use fictional material for anything you would not put in an email.
- Invent a firm, a matter under deadline pressure, and an AI-assisted deliverable. Write the chain of people the deliverable passes through, from first draft to leaving the firm, naming each.
- At each handoff, write what the person handing over assumed about verification and what the person receiving assumed. Mark every point where the two assumptions differ.
- Identify the moment in your chain at which somebody is required to claim the work in the first person. If there is no such moment, write that in bold; it is the finding.
- Now write the five gaps from the lesson down the page and, beside each, the specific control you would add to this workflow. Controls must be steps inside the workflow, not statements in a policy document.
- Draft the affirmation sentence. First person, present tense, naming what is being asserted, short enough that somebody reads it at nine in the evening.
- Design the emergency route: what narrows, who may approve instead, what the record says, and who counts how often it is used.
- Measure your own firm. Take ten AI-assisted deliverables from the last quarter, ask three people who owned each without letting them confer or consult the file, and write your ownership rate out of ten. Bring the number to module 2; it is the best argument you will have for the governance instrument.
Knowledge check
-
What, precisely, is Johnson v. Dunn used for in this lesson, and what is it not?
Answer guidance: It is a 2025 sanctions order of the United States District Court for the Northern District of Alabama, binding only its parties and not precedent for any other court. Three lawyers at a firm that had an AI policy were publicly reprimanded, disqualified from the case and referred to the state bar over hallucinated citations. It is used for one proposition: a policy nobody ran on the document protected nobody. It is not authority for any rule in your jurisdiction.
-
What is the difference between an instrument and a control?
Answer guidance: An instrument states what should happen and is stored somewhere. A control is a step inside the actual workflow that will not let the work proceed until the thing has happened, and that leaves a record showing it did. A firm can have a strong instrument, full acknowledgement and annual training, and no controls at all.
-
Why does an assumption chain reach the client with nobody having verified anything?
Answer guidance: Because every link is individually reasonable. The associate assumes the partner checks, the partner assumes the associate checked because a policy requires it, the group leader assumes the partner ran the process, and the programme owner assumes the group is compliant because nothing has escalated. No one behaves badly and the chain still fails.
-
How should a firm handle genuine urgency without losing verification?
Answer guidance: By defining the emergency route in advance so that short notice changes the scope attempted or the identity of the approver, never whether the check happens, and by recording each use of the route. If the route is used most weeks it is the real process and should be governed as one.
Sources and further reading
- United States District Court for the Northern District of Alabama — Johnson v. Dunn, 792 F. Supp. 3d 1241 (N.D. Ala. 2025) (sanctions order) (2025-07-23). Accessed 2026-09-14.
- American Bar Association — Model Rules of Professional Conduct. Accessed 2026-09-14.
- American Bar Association Standing Committee on Ethics and Professional Responsibility — Formal Opinion 512: Generative Artificial Intelligence Tools (2024-07-29). Accessed 2026-09-14.
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-14.
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (2023-01-26). Accessed 2026-09-14.
- American Bar Association Standing Committee on Ethics and Professional Responsibility — Formal Opinion 08-451: Lawyer's Obligations When Outsourcing Legal and Nonlegal Support Services (2008-08-05). Accessed 2026-09-14.