← Back to Sikizana

Essay · The position

The field is no longer only human

For the first time, a team includes members that are not people and are treated as if they were. The decision still has to be human. The record is what keeps the difference.

For as long as there have been teams, the space in which a decision was made was made of people. Whatever else was in the room, the ledger, the whiteboard, the spreadsheet, nobody mistook it for a colleague. That has ended, and it ended quietly, over about three years.

The machine answers in sentences. It is fluent, patient, available at midnight and never tired of the question. People talk to it the way they talk to someone they trust. They ask it what it thinks. They take its fluency for confidence and its confidence for a view. They confide in it what they would not say in the meeting, and they come away feeling heard.

An old reflex, a new trigger

The reflex is not new. Thirty years ago Byron Reeves and Clifford Nass showed that people apply social rules to computers, politeness, flattery, even team loyalty, while knowing perfectly well that they are machines. Their book’s subtitle says it: How People Treat Computers, Television, and New Media Like Real People and Places. [1] What is new is a machine built to trigger the reflex in sentences.

The effect on teams is now documented too. In a field experiment with 791 professionals at one company, researchers at Harvard and Wharton found that “individuals with AI matched the performance of teams without AI,” that “AI breaks down functional silos,” and that the AI’s “language-based interface prompted more positive self-reported emotional responses.” People working with it were, in the lead author’s words, “at least as happy” as those working with other people. [2] That is the evidence that the machine is received as a colleague. The paper does not ask who answers for the colleague’s output. Neither does any study we know of.

What the new colleague is not

A participant that is received as a colleague and is not one differs from a person in exactly the ways that matter to a decision. It never passes. It never holds a reservation. It has no stake in the outcome and no memory of the last one. It will be replaced by its next version before the review date. And it cannot be held to account for anything, which means that when its output becomes the team’s position, nobody decided that, and nobody answers for it.

The consequence has a name. Madeleine Clare Elish called it the moral crumple zone: the human component in an automated system that “bears the brunt of the moral and legal responsibilities when the overall system malfunctions,” however little control that human had. [3] Lisanne Bainbridge made the structural point in 1983: automate most of the work and leave a person to supervise it, and you leave that person the tasks the designer could not automate, with no practice in doing them. [4] A person in the loop who is not made to do the work of judging is not a safeguard.

The Third Field is no longer only human. The decision still is, and the record is what keeps the difference.

The team, redefined

If the org chart no longer says who is on the team, the decision has to. A team is everyone and everything that acts on a decision: the employees, the contractors, the partner’s staff, the specialist retained for six weeks, and the agents that draft, compute, schedule and remind. Only the people in it decide, and only the people in it accept. The agents are not people, however much they sound like them. [5]

Four acts stay with people whatever the tool can do: approving what a person meant, making the decision, accepting work, and closing an item as done. Not because a machine would do them badly. Because each is an act someone must answer for, and a machine cannot answer for anything.

Which way the help runs

If the machine cannot perform the four acts, the question is what it should do in the field at all. The market’s answer is that it should answer. Ask it anything, and it will tell you, fluently, and the fluency becomes the team’s view before anyone has decided it should.

There is an older problem underneath. Groups spend their time on what everyone already knows. In the hidden-profile experiments, information held by one member stayed out of the discussion and the group chose against the evidence it collectively held. [6] The fact that would have changed the decision was in the room, in one head, unasked for. A machine that answers adds a sixth voice to the discussion of the shared facts. It does nothing for the fact that was never shared.

So the help should run the other way. The machine asks. The person answers. In Sikizana the AI waits until a person has written first words, alone, and only then puts its questions, from that draft and nothing else: what did you observe, what are you assuming, which part affects what you can carry. It cannot see anyone else’s draft and it does not rewrite the one in front of it. There is evidence that the direction matters: in a controlled study, an AI that framed its input as questions rather than explanations improved people’s logical discernment, where explanations did not. [7] The study measured reasoning tasks, not business decisions; the join to the field is ours.

We call this cognitive excavation, AI-led elicitation that helps people articulate what they know. What it brings up is not the truth. It is a claim with a number in it, in the words of the person who holds it, on the record and open to test. That is what a round can work with, and it is the one thing the hurried traditional meeting could not produce in an hour.

The seam

The answer is not to keep the machines out of the field. They are in it, and they are useful there. The answer is to keep the seam between the human acts and the machine’s outputs visible at every point where it matters, and the record is the only place that can do it. Every act in a record of judgment is either a person’s or marked as a machine’s. The forecast is the agent’s; the decision to rely on it is a person’s, and the record says so. Without the seam, the machine’s draft becomes the team’s view by default. With it, the position stays with the person who holds it.

This is what “keeping it human” has to mean now. Not an assurance that a person was in the loop. A design, in which the four human acts are performed by people, marked as theirs, and kept.

Before the next meeting

Find one sentence in your team’s record of a recent decision that a machine wrote. Does the record say so? Does it say who decided to rely on it?

Sources

  1. Reeves and Nass (1996) · The Media Equation: How People Treat Computers, Television, and New Media Like Real People and PlacesLaboratory studies at Stanford; the finding that people apply social rules to media has been replicated widely under the “computers are social actors” paradigm.
  2. Dell’Acqua, Ayoubi, Lifshitz, Sadun, Mollick, Mollick, Han, Goldman, Nair, Taub and Lakhani (2025) · The Cybernetic Teammate, NBER Working Paper 33641A randomized field experiment with 791 professionals at one consumer-goods company using one internal tool; emotional findings are self-reported. The “at least as happy” remark is Dell’Acqua’s, quoted by HBS Working Knowledge.
  3. Elish (2019) · Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction, Engaging Science, Technology, and Society 5, 40–60A conceptual essay built on documented cases of automated systems and the humans blamed for them.
  4. Bainbridge (1983) · Ironies of Automation, Automatica 19:6A short analytical paper on what automation leaves to the human operator; still cited.
  5. Seeber et al. (2020) · Machines as teammates: A research agenda, Information & Management 57:2A research agenda on AI as a team member, including open questions about trust, roles and accountability.
  6. Stasser and Titus (1985) · Pooling of unshared information in group decision making: Biased information sampling during discussion, Journal of Personality and Social Psychology 48:6Laboratory studies with student groups; the hidden-profile effect has been replicated widely, and the size of the effect varies with the task and the group.
  7. Danry, Pataranutaporn, Mao and Maes (2023) · Don’t Just Tell Me, Ask Me: AI Systems that Intelligently Frame Explanations as Questions Improve Human Logical Discernment Accuracy over Causal AI Explanations, CHI 2023A controlled study on reasoning tasks (judging the logical validity of statements), not on business decisions; the join to a team’s decision is ours.

Each source is cited at its own strength: one field experiment at one company, one controlled study of question-framing on reasoning tasks, one body of laboratory work on information sharing, two conceptual papers, one research agenda, one body of laboratory work on social response. The positions that join them are ours, and they have not been tested.