Judgment Lab

Strengthening human judgment in AI-enabled work.

Failure mode 5 of 7

Invisible delegation

PME means professional military education. The Classroom version is also available for high school.

An analyst asks an AI for feedback on an email telling a client their project will slip three weeks, then for a better structure, then for a more professional version, and revises after each. What can happen next is invisible delegation: a request for help hands over criteria, structure or meaning, and nobody notices. Illustration based on a 2025 study of knowledge workers' AI use (Lee et al.). The analysts, their AI, the email and the firm's standard are constructed.

1:39 · Read the figure

Invisible delegation, an animated figure in nine beats

An animated figure. A red dot stands for an analyst improving a draft email that tells a client their project slips three weeks. The task card carries the firm's standard: bad news first, with the new date. The analyst reads it. Beside the email is a ledger in three lanes: "The analyst decides", "Requests" and "The AI decides". Three undecided cards start in the analyst's lane: the criteria, how the email runs, and what "professional" means. The analyst asks an AI, "Give me feedback." The request carries the criteria card across into the AI's lane, where it turns grey. The AI's notes come back ("Cut the jargon", "Formal phrasing", "Add a recap"), and the analyst ticks all three, keeping the AI's recap. The analyst then asks, "Suggest a better structure," and "Make it more professional." Each request carries a card across, and the analyst revises to draft 4. Then, under each request, a line marked "unsaid" shows what it left out: whose criteria, whose outline, whose meaning of professional. A second analyst reads the same standard. Before each request, they write the decision on its card in red: "New date, and why"; "Date, reason, plan"; "Clear and courteous". Then they ask in the same words. Their cards stay in their lane. Both get the same AI notes. The second analyst judges them against their decisions: they tick two and strike "Add a recap", writing "not in my outline". Both emails meet the standard. The manager asks each analyst, with only the two emails to go on, both marked "Revised with AI": "Better by whose standard?" Both say, "The firm's standard." Then the manager asks, "Which notes did you keep, and why?" The first says, "I kept them all. They made it clearer." The second says, "Kept two that fit what I decided. Struck the recap: not in my outline." The figure ends on the reviewer’s check, with the source under the picture. Each beat is listed below with what is on screen.

  1. Invisible delegation · Failure mode 5 of 7

    On screen: the name, "Invisible delegation", appears large over the stage with "Failure mode 5 of 7" above it, then moves into the top-left corner, where it stays.

  2. An analyst improves a client email to the firm's standard. We'll track three decisions.

    On screen: a task card at the top left: "Example task. Tell the client the project slips three weeks. Standard: bad news first, with the new date." A red dot labelled First analyst appears with a draft email, goes up to the standard, which grows larger and is underlined, and comes back. Beside the email, a ledger in three lanes: "The analyst decides", "Requests" and "The AI decides". Three dashed cards, still undecided, sit in the analyst's lane: "The criteria", "How the email runs" and "'Professional' means". The AI's lane shows three empty slots.

  3. The analyst asks an AI for feedback. The request hands the criteria to the AI.

    On screen: a speech bubble, "Give me feedback", clips onto the criteria card, and the two travel slowly across into the AI's lane, where the card turns grey. The request settles in the middle lane. The AI's notes appear under the email: "Cut the jargon", "Formal phrasing", "Add a recap". The analyst ticks all three, and the AI's recap stays in the email as grey text: draft 2, revised with AI.

  4. Then for a better structure, and a more professional version.

    On screen: "Suggest a better structure" carries "How the email runs" across, and "Make it more professional" carries "'Professional' means". Both turn grey in the AI's lane. The AI's grey headings and paragraph appear in the email, and the analyst revises to draft 4.

  5. Each request left out whose criteria to use. The AI used its own.

    On screen: a key appears, "Left unsaid". Under each request, a line marked "unsaid" with words in a dashed box: whose criteria?; whose outline?; whose meaning? The three grey cards pulse.

  6. A second analyst reads the standard, then decides before asking.

    On screen: the first analyst's ledger moves up, and a second analyst appears below with a draft and the same three dashed cards. Their dot goes up to the standard, "Bad news first, with the new date", which is highlighted exactly as it was for the first analyst.

  7. Before each request, the second analyst writes the decision down.

    On screen: the dot writes on the criteria card, which turns red: "New date, and why". Then it asks, in the same words as the first analyst, "Give me feedback". It writes "Date, reason, plan" on "How the email runs" and asks, "Suggest a better structure". It writes "Clear and courteous" on "'Professional' means" and asks, "Make it more professional". All three cards stay red in the analyst's lane; the AI's lane stays empty.

  8. Both get the same AI notes. The second analyst judges them against their decisions.

    On screen: the same notes appear under the second email: "Cut the jargon", "Formal phrasing", "Add a recap". The second analyst ticks the first two. The outline card ("Date, reason, plan") pulses for two seconds as they strike "Add a recap" in red and write underneath: "not in my outline". The same grey headings and paragraph appear in this email, up to draft 4; it has no recap. Both emails meet the standard.

  9. The manager asks each analyst two questions, with only the emails to go on.

    On screen: the ledgers and notes fade; the two emails stay, both at draft 4 and marked "Revised with AI". "The manager asks: Better by whose standard?" appears, and both analysts give the same answer: "The firm's standard." A second question appears: "Which notes did you keep, and why?" The first analyst says, "I kept them all. They made it clearer." The second says, "Kept two that fit what I decided. Struck the recap: not in my outline."

  10. Reviewer’s check: What did the AI decide for them? Ask: whose standard? Which notes did you keep, and why?

    On screen: both lines of the check appear under the picture at the same size, and the ledgers and notes return: the first analyst's three decisions in the AI's lane, the second's three in their own, in their words. A source line appears: "Illustration based on a 2025 study of knowledge workers' AI use (Lee et al.)."

Reviewer’s check

What did the AI decide for them?

Ask: whose standard? Which notes did you keep, and why?

An analyst who didn't notice what they handed over can honestly answer the check's first line, "What did the AI decide for them?", with "nothing", and both analysts name the same standard when asked whose it was. The follow-up separates them. In the figure, both analysts read the same standard, ask in the same words and get the same AI notes. The first keeps every note because each made the email clearer: the AI's reason. The second wrote their decisions down first and gives their own reasons: they kept two notes that fit those decisions and struck a recap that isn't in their outline. A analyst who decided first could keep every note and still say why each fits. Both emails meet the standard, so the emails can't tell the manager who decided. Asking which notes each analyst kept, and why, can.

The study behind this example

Lee and colleagues surveyed 319 knowledge workers, who shared 936 first-hand examples of using generative AI in their work. In the paper's words, "higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking" (p. 1).

One participant described the kind of request in this figure: "It's a simple task [make a passage professional] and I knew ChatGPT could do it without difficulty, so I just never thought about it, as critical thinking didn't feel relevant" (P275, p. 11). Another checked "whether the AI added irrelevant content and if it changed up my main point of the letter" (P144, p. 8).

The paper calls what the second analyst does stewardship: "translating intentions into queries, steering AI responses, and assessing if the AI response meets their quality standards for work, while retaining accountability for the work" (p. 13).

Its limits: the survey is self-reported, so "critical thinking" means what participants perceived they did, and the links it reports are correlations. Each quote is one participant's. The authors are at Microsoft Research and Carnegie Mellon University.

Source: Lee, Sarkar, Tankelevitch, Drosos, Rintel, Banks & Wilson, "The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers," CHI '25 (2025). Page numbers follow the copy of the paper hosted by Microsoft Research. The analysts, their AI, the email, the notes and the standard are constructed.

From the original PME essay

Invisible delegation occurs when the student does not notice which parts of the work they have handed over. Asking for "feedback" may delegate criteria. Asking for "a better structure" may delegate the argument. Asking for "counterarguments" may delegate the range of imaginable objections. Asking for "a more strategic version" may delegate the meaning of strategic. The language of assistance hides the transfer of judgment. The student believes they are working; the model has already done the framing.

The Irreducible Officer, Section II, What the Problem Looks Like

Go further

All failure modes