First principles · Responsibility
Someone Will Get It
A job that everyone can see belongs to no one. AI is about to add one more witness to the rooms you work in, the one everybody assumes is watching. It can also make the only cure that has ever worked free: saying one name.
felt duty ≈ 1 ÷ everyone who can see it
Shared among the witnesses, a duty shrinks in each head. Give it one name and it is held whole again.
A customer has written in the team channel: the order never arrived, the party is Saturday, can anyone help. Under the message sits a little row of faces, eight people who have seen it, and you are one of them. You are on a call. Someone will get it. At 9:15 the next morning the message is still there, and the row has eleven faces.
The smoke in the room
In 1968 two young psychologists in New York, John Darley and Bibb Latané, ran an experiment that has been repeated in one form or another ever since. A student sat alone in a booth, talking over an intercom with other students about college life. Partway through, one of the voices on the line began to sound like a man having a seizure. When the student believed no one else could hear it, they got up and went for help, and quickly. When they believed four others were listening, many of them stayed in their seats, and those who moved moved later.
In a companion study, smoke began to pour through a wall vent into a room where people were filling in a questionnaire. Alone, people walked out and reported it. Seated with two strangers who stayed calm, most kept writing while the room filled.
The newspaper story that prompted the research, a 1964 murder in Queens said to have had dozens of silent witnesses, turned out to be badly wrong. The experiments outlived the story.
Darley and Latané called the effect diffusion of responsibility, and the principle is structural, not moral. A duty that many people can see is divided among them, and it is divided inside each person’s head. Nobody decides to shirk. Each person sees the others, supposes one of them has it, and their own felt share shrinks toward nothing. With no one feeling like the one, the act has nowhere to start.
The cure is just as old, and it is taught in first-aid classes. Do not shout “somebody call an ambulance.” Point at one person: “You, in the blue coat, call an ambulance.” The moment a request carries one name, the division stops, and one person holds the whole of it.
A duty addressed to everyone is held by no one. A duty with one name on it is held whole.
The new witness in the room
Now change one input: the number of witnesses, and what kind they are. Over the next few years most shared places you work in will gain a witness that does not sleep. The channel has an assistant reading it. The inbox has a triage agent. The doorbell watches the porch, the car watches the road, the building watches its own pipes.
The obvious prediction is that more watching means fewer things slip. The principle says the opposite, at least at first. A witness that everyone assumes is watching is the perfect place to set a duty down. It is present around the clock. It looks capable. It does not say “I thought you had it.” In 1968 the student in the booth assumed one of the four others would go. In 2027 the student assumes the system has flagged it.
The foundation I work from places conscious responsibility in the human: agency is relocated, not abolished, and the human is its site (R13). I do not have to settle what a machine is to see the consequence. Whatever the assistant is, assuming it has picked up your part does not move your part anywhere. The duty is still in the room. It has only lost its last address.
Here is the turn most people will miss. The same machine can do the one thing that cures diffusion, and at a scale no person could. The cure was never more watching. It was naming. A good triage system does not announce “a customer needs help” to eight people. It says “Priya, this one is yours, the party is Saturday.” Naming used to be expensive, because it took someone who saw everything, knew who was free, and was willing to be the one who points. That is exactly the work these systems do well. So diffusion does not vanish. It moves to one question: who writes the rule that decides whose name goes on the message?
The same customer writes the same message. It does not land in the channel under a row of faces. It lands on one phone, yours, with one line: “Yours. Party Saturday. Courier marked it delivered to the wrong unit.” No one else sees it unless you hand it back. You call her at 4:55.
Five predictions
- “Someone” disappears from requests at work. Broadcast asks, “can someone take a look?”, will start to read as careless, the way reply-all does now. Tools will send each request to one named person by default, and showing it to the whole group will become the exception you choose on purpose.
- “I assumed the system caught it” becomes a standard line in post-mortems. Before naming catches up, the first wave of AI monitoring will produce a recognizable failure: something seen by several people and one machine, and acted on by none. Incident reviews will need a name for it, and they will get one.
- Every automated alarm gets one person’s name on it. The organizations that learn this will attach a named human to each alert, the way a ship has an officer of the watch. An alert with no name on it will be treated as a defect.
- The useful assistant in a group chat is the one that points. In families, clubs and neighborhoods, the assistant people keep will not be the one that answers questions. It will be the one that says “Dad, you said you’d get the ice,” and nothing more.
- The rule that assigns the name becomes the real power. Whoever decides whose phone gets the message decides who carries the work. Arguments about fairness at work will shift from who did it to who got named, and people will ask to see the assignment rule the way they once asked to see the rota.
Where I stand
I already work this way, from both ends. When I schedule a recurring job for my AI, I do not post it to a shared board and hope. I hand it over in so many words, “It’s yours,” and I do not look at it again. When the work needs a decision, it does not come back to a list. It comes back to one person, me, by name. In The Last Line Is You I wrote that once machines clear the queue, the line moves to the step that cannot be copied. This is the same lesson from the other side. A machine can carry the watching. It cannot be the person the duty lands on, so the most useful thing it does is find that person and say their name.
The claim ledger
- Already true
- The effect has been studied for more than half a century, and the remedy of pointing at one person is standard first-aid teaching. Shared channels with their rows of read faces show the pattern daily.
- What has to happen
- Routing has to become cheaper and easier than broadcasting: tools that know who is free and who knows what, and that send to one name by default.
- Where I am probably wrong
- Real life may not behave like the booth. Studies of street-camera footage found that when public conflicts turned serious, someone usually stepped in, and larger crowds made it more likely that at least one person did. If AI-watched workplaces turn out the same way, so that “everyone saw it, no one acted” is rare when it matters, the wave of failures I describe will not come, and naming will be a convenience rather than a cure. I am also wrong if people come to resent being named and simply hand the message back.
Back to the channel
The message is still there with its row of small faces. You could leave it for someone. Or you could type the oldest fix there is, one word with a name behind it: “Mine.”
Someone is a name nobody answers to.
Background
- John M. Darley and Bibb Latané, “Bystander intervention in emergencies: Diffusion of responsibility,” Journal of Personality and Social Psychology, 1968.
- Bibb Latané and John M. Darley, “Group inhibition of bystander intervention in emergencies,” Journal of Personality and Social Psychology, 1968.
- Peter Fischer and colleagues, “The bystander-effect: A meta-analytic review on bystander intervention in dangerous and non-dangerous emergencies,” Psychological Bulletin, 2011.
- Richard Philpot, Lasse Suonperä Liebst, Mark Levine, Wim Bernasco and Marie Rosenkrantz Lindegaard, “Would I be helped? Cross-national CCTV footage shows that intervention is the norm in public conflicts,” American Psychologist, 2020.
- John Rector, The Last Line Is You; Wrong on Purpose.
Spine v1.72 · Primitives: none · Results: R13 · Registers: DOCTRINE (cited as grounds for where responsibility sits), REPORTED (the experiments and field studies, sources above). The predictions and the claim that naming is the cure at scale are this piece’s own; the falsifier is in the ledger.
Part of John Rector’s first-principles writing. Start with The Immutable Past, then read The Last Line Is You and Wrong on Purpose.