Coexisting with Machines That Think

My grandfather operated a telephone exchange in a fishing town in northern Norway. For a stretch of the last century, that was a respectable technical career: he sat at the switchboard where every conversation in the district crossed, and he physically made the connections that let one person reach another. The automation of his profession is so complete that most people alive today do not know it was ever a profession. Nobody mourns it. The conversations, which were always the point, continue.

I think about him often now, because I have spent my career doing to office work what the automatic exchange did to his switchboard, and because the machines have recently crossed a line he would have found harder to wave away. The systems I work with today do not just connect or calculate. They read, summarise, argue, draft, and decide — provisionally, imperfectly, but unmistakably. They perform acts that everyone in my grandfather’s world would have called thinking. The polite fiction that machines handle the mechanical while people handle the mental has quietly expired, and we have not yet agreed on what replaces it.

The wrong question

The public conversation about this is stuck on one question — will the machines take our jobs? — and I have come to believe it is the wrong question, not because the answer is reassuring, but because the question smuggles in a picture of society where a job is the only channel through which a person can receive three very different things: income, structure, and standing. Bundling them was always an accident of industrial history. The machines are now unbundling them, whether we have a plan or not.

Ask instead: what do we owe each other in a world where cognition is abundant? That question has the advantage of being answerable by choice rather than by prediction. Nobody can tell you with confidence which occupations exist in twenty years; forecasts about this from very serious institutions have aged like fruit. But what a society owes its members — a floor under their dignity, a route to contribution, a say in the systems that govern them — is not a forecast. It is a decision, and decisions are things you can be early on.

Of the three things a job bundles, income is the one our politics knows how to argue about, and it is honestly the easiest of the three. Transfers, dividends, shorter weeks — the mechanisms are known even where the will is missing. Structure is harder: human beings are creatures of rhythm, and a life without the scaffolding of obligation tends to sag in ways that money does not fix. But standing is the hardest, because standing — the experience of being needed, of your competence mattering to someone — cannot be issued by a ministry. A society that solves the income problem and ignores the standing problem has not avoided the crisis. It has scheduled it.

I am impatient with both of the standard moods on offer. The doomers treat every capability gain as a countdown to human irrelevance, which is a strange way to describe tools that are, so far, overwhelmingly operated by and for humans. The boosters treat every concern as friction on the road to abundance, which is a strange way to describe the largest uninvited redistribution of economic leverage in living memory. Both moods share a fatalism — it is coming, adjust — that I find corrosive, because it converts citizens into spectators of their own future. The historical record says otherwise. Weekends, universal schooling, workplace safety: none of it emerged from the technology. All of it was negotiated by people who refused to be spectators.

Cognition as infrastructure

Here is the frame I find most useful, as someone who installs this technology into institutions for a living. Machine cognition is becoming infrastructure — a utility, like electricity or roads — and the history of infrastructure tells you where the real battles are. Not in whether the thing works, but in who gets connected, who sets the price, what standards it must meet, and who answers when it fails someone.

Infrastructure has a second property that the current discourse keeps missing: it disappears. Nobody describes themselves as an electricity user; the miracle dissolved into the walls within a generation, and what remained visible were the things electricity made possible. Machine cognition will follow the same arc. The chatbots and copilots that feel like the story today are the story’s scaffolding; a decade out, the thinking will have receded into the plumbing of every institution, and the visible questions will be the ones infrastructure always leaves behind — who is connected, on what terms, and who answers when the lights go out for someone.

When electricity was young it electrocuted people, burned districts, and enriched whoever controlled the wires. It became civilisation-grade infrastructure through decades of profoundly boring work: codes, meters, inspection regimes, liability rules, universal-service obligations. Nobody writes poetry about the electrical code, but the electrical code is why you plug in a kettle without thinking about death. Machine cognition is at the electrocuting-people stage of this arc — confidently wrong diagnoses, hallucinated case law, biased screening — and the answer is the same as it was: not to unplug the grid, but to do the boring civic work that makes a wild capability into a dependable one.

The infrastructure frame also clarifies the stakes of access. If thinking-as-a-service becomes as economically necessary as electricity, then exclusion from it becomes a new kind of poverty, and the terms of access become a new kind of politics. A society where the best cognition is reserved for those who can pay the most is not a neutral market outcome; it is a constitutional choice made by default. We have faced this exact structure before — with literacy, with libraries, with telephone lines to fishing towns — and our better answers were public: schools, lending institutions, universal service. I do not know precisely what the public library of cognition looks like. I know that deciding not to build one is also a decision.

The values gap

People in my industry talk about the alignment problem: how to make powerful AI systems reliably pursue what their operators intend. It is real, it is hard, and I am glad serious people work on it. But watching these systems land inside organisations has convinced me that there is a second alignment problem that gets a fraction of the attention: aligning what the operators intend with what the rest of us can live with. A perfectly obedient system pointed at a corrosive objective is not a safety success.

This is what I mean when I say the values gap. Our technical capability compounds on a timescale of months. Our collective agreements — law, regulation, norms, the shared sense of what is acceptable — move on a timescale of years or decades. The gap between those two clocks is where the damage happens: not because anyone is evil, but because in the absence of agreement, the default values that ship are whatever maximised a metric somewhere. Engagement, throughput, cost per case. Metrics are not villains. They are just very fast, very literal, and constitutionally incapable of noticing what they trample.

I will risk a paragraph of patriotism here, because I think the Nordic countries are sitting on an underexploited advantage in exactly this. The tripartite model — employers, unions, and the state negotiating change at the same table — was built for precisely this shape of problem: a technology arrives, work must be reorganised, and the costs of transition need distributing before they curdle into resistance. High trust, flat hierarchies, and a workforce that expects to be consulted are not soft cultural garnish. They are adoption infrastructure. The countries that can negotiate their way through this transition will move faster, not slower, than the ones that try to bulldoze it — because consent, it turns out, scales better than coercion.

Closing the gap is not a job for the AI labs alone, and honestly it is not mostly their job. It happens in procurement meetings where a municipality decides what transparency to demand from a vendor. In union negotiations that establish what monitoring is off-limits. In professional bodies deciding what a doctor or an auditor must personally verify. In the drafting rooms of legislation like the accessibility and privacy regimes that already, imperfectly, encode a floor of dignity into software. That work is unglamorous, slow, and absolutely decisive. The values that govern machine cognition will be set by the people who show up to those rooms.

What coexistence looks like on a Tuesday

Because I distrust essays that stay in the stratosphere, let me land this in the rooms where I actually work. Coexistence, at ground level, looks like a handful of design commitments that any organisation can adopt this quarter, without waiting for parliament or the singularity.

It looks like legibility: every consequential machine decision carries an account a layperson can read — what it looked at, what it weighed, why it landed where it did. Not because the account is always faithful to the mathematics, but because accounts are what make contestation possible, and contestation is the immune system of any institution that wields power over people.

It looks like appeal: a route from every automated outcome to a human with genuine authority to overturn it, staffed and timed so the route is real rather than ornamental. The measure of a society was never how it treats the average case. It is what happens to the exceptions, and machine systems manufacture exceptions at scale.

It looks like measuring the things you actually care about, because in an automated institution, the metrics are the management. A human organisation runs on judgment with metrics as advisors; an automated one inverts that, and whatever you chose to count becomes, operationally, what you chose to want. If the caseworker’s dashboard counts throughput but not the number of people who gave up mid-application, the system will be brilliant at throughput and blind to despair. Choosing metrics used to be a reporting question. It is becoming the closest thing an institution has to writing its own character, and it deserves the same seriousness we give to hiring its leaders.

It looks like honest accounting of the transition. When automation removes a role, the surplus it creates is real money, and where that money flows is a choice made by named people. Reinvesting some of it in the people displaced — retraining with content, bridges to new roles, honest severance — is not charity. It is the price of maintaining the social permission on which every future automation project quietly depends. Organisations that pocket the surplus and externalise the people are spending down a commons they did not build.

It looks like rethinking what we ask schools to produce. For a century, education has been optimised to supply the middle of processes — reliable executors of defined procedures — and that is precisely the territory the machines are claiming. What cannot be claimed is the ends of processes: framing problems worth solving, judging whether a solution actually serves the people it touches, and carrying responsibility in front of other humans. Those capacities are teachable. We mostly do not teach them; we assume they will be absorbed, someday, on the job. In a world where the job’s lower rungs are automated, “someday, on the job” is no longer a plan.

And it looks like keeping humans on the purposes even as machines take the tasks. Every process exists because some human, somewhere, needs something. Machines are becoming superb at the middle of processes. The two ends — deciding what is worth wanting, and standing accountable for what was delivered — are not technical functions, and any organisation that automates them has not become efficient. It has become unmanned.

A wager on shared vision

I will not pretend to certainty about where this settles. The honest range of outcomes is wide, and the people who quote you a single confident future, radiant or ruined, are selling something. But uncertainty about outcomes does not mean uncertainty about posture. You can hold a wide distribution about the future and still know exactly which way to lean into it.

My lean is this: the machines are not the other side of this negotiation. There is no other side; there are only people — building, deploying, regulating, adapting — and the machines are the terrain the negotiation crosses. Terrain matters enormously. It rewards some strategies and punishes others. But terrain has no goals, and the fatalists on both wings who talk as if the technology itself has already decided our arrangements are handing their agency to a landscape.

People sometimes hear all this and ask whether I am an optimist, and I have stopped treating the question as one about temperament. Optimism, in a transition like this, is not a mood; it is a discipline with a daily practice. It means building the transparent version when the opaque one ships faster. It means writing the appeal route into the contract when nobody in the room would have noticed its absence. It means telling a client that the system they want will manufacture a signature machine, and losing the engagement, and sleeping fine. The pessimist and the optimist read the same forecasts. They just file different work into their calendars.

My grandfather’s switchboard went away and the conversations continued, because the conversations were always the point. Most of what fills our working days now is switchboard: coordination, translation, retyping, routing. It can go, and much of it should, and letting it go will be wrenching in all the ways honest people admit and dishonest people wave away. What must not go is the part that was always the point — people needing things from each other, promising things to each other, and answering for those promises face to face.

So this is the project as I understand it, the one worth showing up for in every room where these systems are being wired into the world: build the amazing tech, and build, with equal seriousness, the agreements that let us live well beside it. Dream loudly, spar honestly, and refuse the spectator’s seat. The exchange is being rewired either way. The only open question is whether the people it connects will have written any of the connection rules — and that question is answered by who shows up.