Lesson objective

Distinguish accountability from blame in a professional partnership, and design a reporting culture in which people disclose AI failures early enough for the firm to act.

Plain-language explanation

Why this matters now

A firm can name an accountable person, write the undertaking, and still learn nothing about what is going wrong, because the person who knows will not say. This is not a soft issue appended to the end of a governance module. It is the mechanism that decides whether governance receives any information at all.

Law firms are unusually hostile environments for admitting error. The work is adversarial, the professional identity is built on being right, compensation is often tied to visible performance, and an error can be a disciplinary matter, a malpractice matter and a client-relationship matter at once. Into that environment a firm introduces a technology whose most characteristic failure mode is producing something confident and wrong, and then asks people to report when it happens. The asking is the easy part.

Law / rule: Model Rules 5.1 and 5.3 require reasonable efforts to ensure measures are in effect giving reasonable assurance of conformity. A measure that produces no reports is not giving assurance; it is producing silence, and the two look identical from the managing partner's office. These are model provisions; your jurisdiction adopts its own.

Professional guidance: ABA Formal Opinion 512 (29 July 2024) treats supervision and firm policies as part of how competence is maintained with generative-AI tools. A supervisory system informed only by the incidents that became visible on their own is supervising the tail of the distribution.

A question to sit with

The last time an AI-assisted output in your firm was wrong in a way nobody outside noticed, who found out, and what happened to them? If you cannot answer, that is your reporting rate.

What goes wrong

Accountability is announced and blame is practised. A firm says it wants a no-blame culture, then the first reported incident is followed by a partners' meeting in which the individual is named, a process change framed as a correction of their conduct, and a marked coolness toward them for a quarter. Everyone learns the real rule in one cycle, and no amount of subsequent messaging undoes it.

The only available response is disciplinary. A firm with no category for "near miss" must treat every report as either nothing or misconduct. Faced with that binary, supervisors classify almost everything as nothing, and the firm loses the entire middle of the distribution, which is where the learning is.

Reporting has a cost and produces no visible benefit. Someone reports, fills in a form, is asked three follow-up questions, and never hears what happened. They will not do it twice, and they will tell their colleagues.

Blame flows down the hierarchy and accountability does not flow up. The associate is named; the partner who approved the workflow, set the deadline and never asked what verification was performed is not. People notice this with great precision, and it teaches them that reporting upward is dangerous and reporting downward is safe.

Silence is read as health. A quarter with no incidents is announced as a success. It is an absence of information, and the two are not the same.

The practical answer

Hold the distinction precisely, and teach it in these words.

Accountability Blame
Points Forward: what now, what changes Backward: whose fault
Attaches to A named person, before the event A named person, after it
Asks What did the system let happen? Who failed?
Produces Information and a change Silence and a defensive account
Ends when The change is made and verified The person is punished or exonerated

Both are legitimate. A firm needs blame — properly called responsibility for misconduct — for the small number of cases involving dishonesty, concealment, or deliberate disregard of a rule the person knew and understood. The error is applying that response to the large number of cases that are ordinary human and systemic failure, because doing so is how a firm buys silence at the price of one satisfying meeting.

Draw the line in advance and in writing. Best practice: publish the small set of things that will always attract an individual response, and commit that everything else is treated as a system finding.

Always an individual matter Always a system finding
Concealing an error, or altering a record Following a written method that turned out to be wrong
Putting client confidential material into an unapproved tool after being told not to Using a tool nobody told you was not approved
Certifying as checked something that was not checked Missing a fabricated citation the workflow provided no way to catch
Repeating the same failure after a specific instruction and training Being too slow to verify at a volume nobody measured

A firm that publishes this table and then honours it has done more for its reporting rate than any amount of encouragement.

Build a near-miss category and use it. Best practice, drawn from the Govern and Manage functions of the NIST AI Risk Management Framework 1.0 (26 January 2023) — a voluntary framework that neither creates nor discharges a professional duty: record events where a control caught something before it left the firm. Near misses are the highest-value data a programme has, because they are frequent, harmless, and describe exactly the failure mode that will eventually get through.

Make reporting cheap, fast and visibly useful. Three commitments, each measurable:

  1. Under two minutes. What happened, which workflow, what caught it, what it would have cost. Prose, not a form with sixteen fields.
  2. Acknowledged same day, by a person. Even if the answer is only that it has been received and read.
  3. Closed with an outcome the reporter sees. The change made, or the reasoned decision not to change anything.

Report the numbers upward in a way that rewards disclosure. AI Coalition Network recommendation: present reporting volume to the partnership as a positive indicator and say so explicitly in the report, so that a practice group with more reported near misses is understood to be better supervised rather than worse behaved. This is this programme's educational standard, not a rule of any jurisdiction. Say it in the meeting, not only in the document, because the meeting is where people learn what the firm actually values.

Remember this

  • Accountability points forward at the system and is assigned before the event; blame points backward at a person and is assigned after it.
  • Publish in advance the short list of conduct that will always attract an individual response, and treat everything else as a system finding.
  • A firm with no near-miss category loses the middle of the distribution, which is where all the useful information lives.
  • Make reporting under two minutes, acknowledged the same day, and closed with an outcome the reporter can see.
  • Zero reports is an absence of information, not evidence of health; say so when you present the numbers to the partnership.

Legal example

Pellowe Nakagawa LLP is a fictional 85-lawyer firm that has just introduced an AI-assisted research workflow with a mandatory citation-verification step. Six weeks in, a fictional second-year associate, Diarmuid Oyelaran, notices at eleven on a Tuesday night that a brief due the next morning contains a quotation he cannot find in the case he is citing. He rechecks, finds the case is real but the quoted sentence is not in it, removes the passage, rewrites the paragraph and tells the supervising partner the next morning. The partner, an invented litigator named Constance Achterberg, has a choice that will set the firm's reporting rate for the next two years. Her first instinct is to say nothing, because nothing went out and raising it would draw attention to her own group. Her second is to ask Diarmuid why he had left the verification so late. She does neither. She reports it the same day as a near miss, names herself as the reporter, and writes four lines: the workflow, the failure mode, the control that caught it, and what it would have cost if it had not. The firm's AI programme owner acknowledges it within the hour, and the finding goes into the monthly report as the only entry. Two things follow. The programme owner notices that the workflow provided no way to distinguish a real case with an invented quotation from a real case correctly quoted, and adds a required step: every quoted passage is opened in the source and the page confirmed, recorded per quotation. That is a system change, and it would never have been found by asking people whether the workflow was working. Second, Constance says at the next partners' meeting that she reported it, that nothing happened to her, and that she would like more of this. Reports go from one to fourteen over the following quarter. Eleven are trivial. Two are the same fabricated-quotation pattern, which tells the firm it is a pattern rather than an incident. One is a partner discovering that his own group had been using an unapproved consumer tool for five months, which surfaces only because the cost of saying so has come down.

Responsible-use warning

Never announce a no-blame culture and then handle the first report as a conduct matter. People do not learn a firm's culture from its policy; they learn it from the single most recent instance, and one badly handled report will close a reporting channel for years. If your firm cannot commit to the published line between an individual matter and a system finding, do not open the channel yet. A channel that people distrust is worse than no channel, because the resulting silence is read upward as evidence that nothing is going wrong.

Human-review requirement

A licensed attorney owns every output that reaches a client or a court, and owns the decisions that follow a reported failure. The reviewing attorney determines whether anything already delivered must be corrected, whether the duties of communication and candour as adopted in the relevant jurisdiction require the client or a tribunal to be told, and within what time. Those are professional judgements and they are not suspended by the fact that the report was voluntary or the culture no-blame. The attorney also owns the classification: whether this event falls on the published individual-matter line or is a system finding, made against the written line rather than against how senior the person involved happens to be.

Practical exercise

Design the reporting half of your firm's accountability model. Use fictional material only.

  1. Write down the last three AI-related problems you personally know about in any organisation. For each, record how the organisation found out, and how long after the event.
  2. For each, ask whether a formal channel would have received it earlier. Write one sentence on what stopped an earlier report: cost, fear, no category, no channel, or nobody thought it mattered.
  3. Draft your firm's two-column table: the short list of conduct that will always be an individual matter, and the kinds of failure that will always be treated as system findings. Keep the first column under six lines and make every line something a person could recognise in advance.
  4. Write the near-miss definition in one sentence, and three example events from a workflow you know that would qualify.
  5. Design the report itself. Write the actual fields, and time yourself completing one for an invented incident. If it takes longer than two minutes, cut fields until it does not.
  6. Write the three commitments: acknowledgement time, who acknowledges, and how the reporter learns the outcome. Name a person and a named deputy for each.
  7. Rehearse the first report. Invent one involving your firm's most senior person and write, in full, what the programme owner says to the partnership about it. Then decide honestly whether your partnership would let them say it.

Knowledge check

  1. What is the practical difference between accountability and blame?

    Answer guidance: Accountability is assigned before an event, points forward at the system, asks what the system let happen, and ends when a verified change is made. Blame is assigned afterwards, points backward at a person, asks who failed, and ends when the person is punished or exonerated. A firm needs both, but applying blame to ordinary systemic failure buys silence.

  2. Why is a published line between individual matters and system findings more useful than a general commitment to no-blame?

    Answer guidance: Because people can see it in advance and rely on it. A general commitment is tested by the first incident and disbelieved if the handling looks like discipline. A short published list of conduct that will always attract an individual response — concealment, record alteration, certifying unchecked work, repeating a failure after specific instruction — makes the rest genuinely safe to report.

  3. What does a quarter with zero AI incident reports tell a managing partner?

    Answer guidance: That the firm has no information. It is produced identically by a firm with no problems and by a firm whose people do not report, and nothing in the number distinguishes them. It should be presented as a finding to investigate rather than as a success, and reporting volume should be framed to the partnership as a positive indicator.

  4. Does a no-blame reporting culture change what the firm must do about an error that has already reached a client?

    Answer guidance: No. The reporting culture governs how the firm treats the person who disclosed it. The professional duties are unaffected: a licensed attorney still decides whether the client must be told, whether a correction to a tribunal is required, and how quickly, under the rules as adopted in the relevant jurisdiction.

Sources and further reading