Responsibility Doesn't Compile

How AI Work Actually Plays Out

It’s become clear to me how this plays out. Forget the sci-fi version where AI replaces everyone, and the complacent one where it’s just a tool. Within a few years, most knowledge workers will spend the bulk of the workday talking to AI systems. Delegating, reviewing, redirecting. Every individual contributor becomes, functionally, a manager of a large team. The team just isn’t human.

I’m already living this in miniature. A big fraction of my day goes to briefing agents, reading their output, catching where they went sideways, re-scoping the work. It feels less like using software and more like running a staff of eager, tireless, occasionally overconfident junior employees. Extrapolate that from one exec to the whole white-collar workforce and you have the real transformation: the unit of work shifts from doing to directing.

The job that survives is the foreman’s job

We’ve run this experiment before. The Industrial Revolution didn’t eliminate work. It changed what a worker was, brutally and unevenly.

Before mechanization, a weaver owned the whole task. His pace, his tools, his standards, his cottage. Power looms made cloth cheaper, but the deeper effect was decomposing his craft into machine-tending: loading, watching for faults, stepping in on exceptions. Skill that took a decade to build was suddenly embedded in capital equipment. One worker went from operating one loom to overseeing many. The Luddites of 1811 get remembered as anti-technology cranks. They were skilled artisans who correctly saw the value of their craft being transferred into machines owned by someone else.

That’s the analogy for knowledge work, and it’s precise. The craft (writing the code, drafting the brief, doing the analysis) is being embedded into the machine. What’s left for the human is the foreman layer: specify the work, judge the output, handle the exceptions, own the consequences. The industrial foreman oversaw a room of machines. The knowledge worker of 2030 oversees a fleet of agents. Same job, and the same promotion path: the factory foreman was usually the best craftsman on the floor, the weaver who stopped weaving to supervise. Doing to directing, run once already.

The obvious objection: if AI can write the code, why can’t it evaluate the code? Won’t the foreman layer get automated too? On capability, probably yes. But the foreman’s job was never about capability. It was about accountability. He survived because he was the one you fired when the shipment was bad. Someone has to answer for the work, and a machine can’t be fired, sued, or shamed in front of its peers. When an agent ships a security hole or misquotes a contract, the org chart still needs a human name attached to the failure. Judgment stays human because responsibility doesn’t compile. The word has carried this for five hundred years: before the mills borrowed it, a foreman was the juror who stood up and spoke the verdict for the twelve. The one who owns what the group decided.

Accountability is also where adoption will bottleneck. A company’s AI capacity isn’t its model budget; it’s headcount times the amount of blame each person can credibly carry, and that second number has limits. The railroads learned both halves of this. After a head-on wreck on the Western Railroad in 1841, the investigation produced one of the first modern management structures: defined divisions, exact timetables, a named person responsible for every stretch of track. And the first durable rail unions, the engineers’ brotherhoods, formed from below to cap what one name could be blamed for. The Brotherhood of Locomotive Engineers began as a mutual aid society, insuring the men in the cab before it ever bargained for wages. Unlimited accountability isn’t a structure. It’s a scapegoat, and skilled people refuse the cab until it’s bounded.

That negotiation has already started. The 2023 writers’ strike was the brotherhoods’ modern echo: the WGA didn’t ban the machine, it bargained the boundary of credit and blame. AI output can’t count as source material, and a writer handed a machine draft to fix gets paid for a screenplay, not a polish. No cleaning up the loom’s output at machine wages. SAG-AFTRA ran the same play for digital replicas: consent and payment before your likeness works a shift without you. Expect a version of this in every knowledge profession, unionized or not, because the leverage doesn’t come from the picket line. It comes from the fact this piece keeps circling: someone’s name has to be on the work. How many agents can one person be asked to own? What is her sign-off worth? What happens to her when one of them is wrong? The companies that scale AI first won’t have better models. They’ll have the org design that distributes blame across enough people, with limits each of them can live with.

Three more lessons from that era matter here.

The gains lag the disruption. Economists call it the Engels’ pause: from roughly 1790 to 1840, British output per worker climbed steadily while real wages stayed nearly flat. Two full generations absorbed the dislocation while the gains accrued to capital. Living standards did eventually rise, dramatically. But “eventually” was fifty years. Longer than a working life. No law of economics says the people disrupted by a technology are the ones who benefit from it, or that the benefit arrives on a schedule that helps them. Expect a knowledge-work Engels’ pause: output per worker explodes, headcount growth stalls, and wage gains concentrate in the small group who can direct machine labor well.

The transition created a new class, and that class won. The factory system invented the professional manager. Before about 1850 there was no managerial career track; coordination happened through markets and family firms. Once production meant orchestrating hundreds of machines and workers, the coordinator became the scarce resource. Alfred Chandler called it the visible hand replacing the market’s invisible one. The same move is happening now: the scarce skill is shifting from producing work to specifying and evaluating it. Taste. Judgment. Knowing what good looks like. The people who thrive won’t be the best writers or coders; they’ll be the best editors and architects. And the leverage changes: a factory manager directed 200 workers, and an AI-era IC can plausibly direct the output of 200. Management stops being a rank and becomes the baseline job description.

Discipline systems follow the machines. Factories changed time itself. The clock, the shift, the whistle: industrial work invented schedule-discipline because machines demanded synchronization. AI inverts this. Agents run continuously, so the bottleneck is human review bandwidth. The dark version is knowledge workers on call to their own agent fleets, checking on the runs at 11pm the way a mill owner walked the floor. The tempo of work reorganizes around the machine’s tempo. Again.

The part nobody’s pricing in: we’ll talk to AIs more than to people

The workplace consequence is legible. The social consequence is stranger, and I think bigger.

Do the arithmetic on a workday in this world. Eight hours, six of them in dialogue with AI. Add AI tutors for your kids, assistants in every consumer app, voice agents handling every customer interaction. For most working adults, the majority of daily conversational turns will be with machines. That’s never been true of any technology. TV displaced conversation but didn’t simulate it. This does.

Some second-order effects I’d bet on.

The conversational skills that atrophy first are the ones AI doesn’t require. AI is infinitely patient, never bored, never carrying a competing agenda. It doesn’t need to be persuaded, only instructed. Humans are none of those things. Negotiation, reading resistance, tolerating being misunderstood, waiting your turn: these are muscles, and the reps disappear. We may end up with a generation superb at specifying what they want and terrible at getting it from someone who can say no.

Status inverts around human attention. When machine dialogue is abundant and free, undivided human attention becomes the luxury good, and it’ll be marketed as one: human-taught classes, human-staffed service, “human-in-the-loop” as a premium tier. Mass production made handmade goods scarce, and “artisanal” went from default to status symbol. Handmade conversation is next.

Some communication survives automation for a harder reason: the message is the cost of sending it. An apology that took nothing to produce is worth what it cost. Think of it as spending empathy tokens — a layoff delivered face to face, a doctor walking a family through a diagnosis, a founder calling the customer whose data got leaked. The words are almost beside the point; what’s communicated is that a person showed up and paid something. Route it through an agent and the recipient doesn’t hear the message, they hear that you didn’t come. The machines will draft better words than we do. They can’t spend them.

The relational default shifts. Everyone effectively gets a staff, the thing that used to come only with seniority. But talking to your staff isn’t peer conversation. It’s directive, asymmetric, and consequence-free for the speaker. When most of your daily interactions can’t contradict you in a way that costs you anything, that posture leaks. Managers have always struggled with this — the CEO-brain problem of forgetting how to be disagreed with. Now it’s everyone’s occupational hazard.

And loneliness gets more ambiguous, not less. AI conversation is real enough to satisfy the itch and unreal enough to not build the thing the itch is for. The industrial migration to cities broke village social structures decades before urban institutions grew back to replace them: unions, clubs, churches, pubs adapted to factory schedules. The social technologies that make an AI-saturated life healthy don’t exist yet. They’ll get invented. Late, if the pattern holds.

What will they look like? The industrial era’s answers were oddly specific inventions: the weekend, the corner pub, friendly societies, Sunday football leagues. Institutions built to give factory-scheduled people somewhere human to be. The AI-era equivalents will rhyme. A human sabbath: agent-free time, first a wellness fad, then a norm, the way the weekend went from labor demand to law of nature. Third places whose membership pitch is verified human presence — the early version is run clubs and board-game bars filling up right as screen time peaks. Conversation taught as a practiced skill, Toastmasters for the atrophied muscles: negotiation reps, disagreement reps. And etiquette, the cheapest social technology there is: norms about disclosing when an agent wrote the message, when it’s insulting to send one, when showing up yourself is the point. All of it sounds small. So did the weekend.

Truth, when answers are free

There’s an epistemic version of the same pattern, and it may matter more than either of the above.

The Industrial Revolution industrialized information too. The steam press cut the cost of a newspaper by an order of magnitude, the penny press was born, and what followed was decades of sensationalism, fabricated stories, and yellow journalism, because when distribution is cheap the incentive is volume, not accuracy. What restored trust wasn’t less information. It was new institutions built on top of the flood: professional journalism with editorial standards, wire services staking their business on being right, peer review. Verification became an industry because raw information had become worthless on its own.

AI runs this again at a steeper grade. Everyone now has an oracle that answers any question instantly, fluently, with total confidence. Fluency was our main heuristic for credibility. When every answer arrives polished, the signal that separated someone-who-knows from someone-who’s-guessing is gone. Access to information is now free and infinite; the scarce good is knowing whether any of it is true. The machine makes the output abundant, and the human’s remaining job is judgment about the output. Epistemics turns into the foreman’s job too.

Two forks from here. The centralized one: a handful of models become the de facto arbiters of what’s true, a level of epistemic concentration no newspaper baron ever approached. Hearst could shape opinion, not autocomplete it. The fragmented one: everyone’s AI gets tuned to their priors, and the shared factual baseline that mass media accidentally built over a century dissolves back into a thousand villages, each with its own oracle. Probably we get both at once. As before, the verification layer this era needs (provenance standards, models staking their reputation on calibration the way wire services staked theirs on accuracy) will arrive late, after real damage is done. Betting on the institutions eventually emerging is safe. Betting on them arriving on time never has been.

Where I land

The Industrial Revolution’s endpoint was good. Shorter hours, longer lives, the modern middle class. I’d take that trade for AI in a heartbeat. But the path ran through fifty years of stagnant wages, broken crafts, and dislocated communities, and the workforce that bore the costs wasn’t the one that collected the gains — their grandchildren were. It works out, eventually. How fast it works out for actual people depends on choices made during the transition: who owns the leverage, how the gains get shared, how quickly the new institutions get built. And last time, those weren’t gifts from the market. That gap closed only when workers organized and Parliament legislated. The weekend was a demand before it was a norm.

The foremen did fine in the factory era. The weavers didn’t. The strategic question, for individuals and companies and countries alike, is the same one it was in 1811: get to the supervising side of the machine, of its labor and of its claims, and don’t assume the market will move you there on its own.