The Interior Frontier Part I

The forecasts of the AI decade measure everything the machines can do. One in six people alive is lonely, and not one forecast counts them.

One person at the end of their long day, phone in hand, a contact list forty names deep and a group chat ticking on the screen, and not a single message in the whole of it worth the risk of sending.

The World Health Organization put a number on that person in 2025, and it is the number this essay will be about and speak to: roughly one in six people alive.

The foreboding story being told about the coming decade is dark, and it is being told with overbearing confidence. Machines arrive that do the work better than we do. The people who own the machines pull further ahead of everyone else. Whole categories of human effort turn obsolete, and the humans attached to them become, in the cold arithmetic of the forecast, mere surplus. Its most rigorous form is a long-form essay called Situational Awareness, written two summers ago by a young researcher out of one of the frontier labs. It has been read since in enough serious rooms that it now feels like a settled fact rather than a prediction. The map it draws of the acceleration is honest, and it has largely held up. It is also entirely about the world around the person with the phone, and not at all about them.

What the dark forecast leaves out is not a detail. It is the entire human question. Every version of the forecast, from the rigorous to the careless, measures the same things: how much a machine can do, and how fast the doing accelerates. It then treats those measurements as if they exhibit all that is known about an entire decade. They do not. A curve that tracks what a system can produce says nothing about what happens inside the several billion people standing beside that production when it climbs. That second question has no place on any chart. So it goes unasked, and nobody notices that unasked does not mean it is unimportant.

There is a useful frame for what is actually happening here, and it comes from the study of stress. When Hans Selye first described the term and the body’s response to it, his central finding was that the biological arousal of stress is, in itself, neutral. The same demand can either optimise or break an organism. Which one it does is not determined by the demand at all. It is determined by perception, by perspective, by awareness, and by the conditions surrounding the person meeting the demand. Meet stress with the right resources and framing, and it becomes eustress: stress that sharpens a person and can aid growth. Meet the same stress without those resources and that framing, it becomes distress: the kind that corrodes an individual. Selye’s insight was that stress is not the enemy. Stress negatively met, stress a person has been given no way to meet well, is the enemy.

Powerful artificial intelligence is a stress of exactly this kind, arriving at civilisational scale. And every forecast currently shaping how several billion people will meet it speaks in a single register. Displacement. Obsolescence. Surplus. Selye spent a career establishing that the framing surrounding a demand does much of the work of determining what the demand does to an individual. He would have recognised the conditions being assembled here: one arriving pressure, one available story, and a great many people about to meet the first while holding only the second.

The same capability that displaces a person is the capability that could equip them. Nothing in the silicon decides which. What decides it is who builds, for whom, towards what end, and what a person has been told in advance about what is coming for them.

There is another outcome. The map of the acceleration is not wrong, only partial, and the road it leaves unmarked is the one these pages are about. Two roads diverge here, and the whole of the difference lies in which one gets taken. The one everyone is describing leads where the forecast makes it lead. The other is less travelled, in Frost’s plain sense: not obviously better, not yet proven, but genuinely accessible and mostly untrodden. Taking it seriously means looking hard at the territory which the capability maps ignore. That territory, once noticed, turns out to be enormous. It turns out to be where most of the human stakes of the coming decade actually live.

The Territory Off the Map

Begin with what is already true, before any further capability arrives.

The commission convened by the World Health Organization in November 2023 reported in the summer of 2025, and the findings are not gentle. Social disconnection is linked to an estimated 871,000 deaths every year, a toll on the order of the world’s major causes of premature death, quietly accruing, mostly unspoken. This is not a Western affliction alone, nor is it confined only to the old. In the European Union, the first bloc-wide survey conducted in 2022 found around 13% of adults are lonely most or all of the time, and more than a third are lonely at least sometimes. An earlier European study had counted more than 75 million adults across the continent who see family and friends at most once a month. National statistics from 2023 to 2025 in the United Kingdom put chronic and occasional loneliness across a quarter or more of the adult population, highest among the youth and those already living with depression. In India, the data may be thinner but the direction is clearly the same. Recent surveys have found loneliness reaching two in five older adults in some samples, and as high as two in three older people in rural areas. In the United States, a 2023 Surgeon General’s advisory compared the health risk of chronic loneliness to smoking 15 cigarettes a day, and roughly half of the American adult population reported measurable loneliness even before the pandemic deepened it.

Set that beside a second, larger figure. The WHO’s 2025 figures put more than a billion people worldwide living with a mental health disorder. In low-income countries, fewer than one in ten of them receive any support or care. Even in wealthy countries with developed healthcare systems, close to half go untreated. A 2025 WHO report found roughly 9% of people living with major depression worldwide receive care that meets minimal standards.

Set all of that beside the technology conversation as it is currently framed: chip export controls, data centre financing and the burden on natural resources, the electricity load of a training run. Hundreds of millions of people, possibly billions, carrying something no one is coming to help them carry, and a decade of extraordinary technical attention pointed almost entirely elsewhere.

The numbers have a plainer form. They are the person at the beginning, multiplied by a sixth of the entire human population.

This is the territory the capability maps leave blank. And it is precisely the territory where the difference between a stress met well and stress met badly decides everything. A billion people, most of them without adequate care, is not a problem that more compute, aimed at automating labour, does anything to touch. It is a problem that a different intention, using the very same technology, might finally begin to reach, if anyone chooses to point it there.

The Choice Worth Watching For

Whether the technology gets aimed at that need or away from it is not decided by the technology. It is decided by the people who hold it. And there is a long, documented record of how those decisions tend to go whenever a kinder path and a more profitable path sit on the table at the same time.

Consider the position we are told we already occupy. If the singularity is genuinely near, and if capability is about to compound beyond anything a person can match, then the entire premise of the dark forecast is that human labour is on its way to being surplus. But that premise contains a question of its own that the tellers rarely ask: if machines are about to be able to do nearly everything, then the scarcest remaining resource is no longer human productivity at all. It is human potential: what a person is, what they become, and what they contribute. No machine can be paid to replace that. In a world of abundant capability, the one thing that cannot be automated is a human being coming fully into themselves. So the question these forecasts bury is stark. At the exact moment a person’s economic output stops being the point, do we abandon the cultivation of human potential altogether, in favour of whoever captures the most profit from the machines? Or do we finally turn the technology towards the one thing it cannot automate away?

History does not leave much doubt about which way the incentive pulls, absent a deliberate choice against it. In 1901, J.P. Morgan funded Nikola Tesla to build a tower on Long Island for wireless communication. Tesla’s deeper ambition, only partly disclosed to his backer, was to broadcast electrical power itself through the air, to anyone, anywhere, for free. Tesla had broken the terms of his contract. Costs had run far past what was promised. A competitor had beaten him to a working transatlantic system, and a financial panic in 1907 dried up capital everywhere. But the difficulty of metering a technology designed to give itself away was, by several contemporary accounts, a genuine part of what ended the funding. Morgan wired the world instead, for a price, and made a fortune doing it. Two decades later, a cartel of the world’s largest lightbulb makers met in Geneva and agreed to engineer bulbs that failed sooner than the era’s technology allowed. They built a global standard around the bulb that wore out, rather than the longest-lasting bulb they already knew how to make. In the same decade, a car company brought a petrol additive to market despite internal knowledge of its dangers, and despite safer alternatives being available. It then spent the better part of a century helping to deny and outlast the science documenting what that additive did to a generation of children’s developing brains.

Every one of those decisions is documented, litigated, or minuted.

Three industries. Three rooms in which a kinder path and a more profitable one sat on the same table, in front of the same men, and the same one was chosen. A century later the table is being set again, this time in front of an industry more powerful than all three, deciding what a civilisation-shaping technology is for, at the exact moment it becomes capable of almost anything.

There is a quieter, older version of the same intuition, and it recurs in a way that is genuinely strange. Nearly every human culture that has left a written record has told some version of the same story, independently, with no contact between the tellers: a better world, once, and a fall from it. Hesiod wrote of a golden age before the metals worsened and life grew hard. Vedic cosmology names its own Satya Yuga, an age of wholeness at the start of a long decline. Genesis opens in a garden, before the leaving. Flood myths recur across societies with no plausible link to one another, each describing a world remade after something was lost. No one has settled why the same shape keeps surfacing across peoples who never spoke. But whether or not any single version of it happened, the fact that it was imagined again and again, everywhere, is worth sitting with for a second. Not as a memory, perhaps. As a hypothesis every culture arrived at on its own: we are capable of more wholeness than we carry, and that something in how we keep organising ourselves quietly costs us the difference. Situational Awareness, and the broader account it anchors, is only the latest telling of the fall. The road less travelled is a wager that the story is not finished, and that the direction of it is still, even now, a choice.

What the Building Taught

The claim that this technology need not be built the way the prevailing forecasts assume is easy to make in the abstract and hard to believe without evidence. The most honest evidence available is small and growing. And it comes from the building process itself.

The work began the way most real things begin, without a thesis. Someone who had spent years inside psychology, the slow, hard-won understanding of how a human being actually changes, could not write a line of software. They sat down with an early AI system to see whether it could hold the shape of what they knew. It could. What changed a long-held scepticism was not that the machine seemed clever. It was something quieter and harder to dismiss. Someone who understood people but could not write a line of code could now build technology, simply by joining what they knew to what a machine is capable of doing. Neither half was enough on its own. Together they made something that neither could have made alone.

What the machine turned out able to hold was the part that mattered, and the part most systems built for the inner life get wrong. Good psychological work does not hand a person a verdict. It holds up a mirror, walks beside them, and leaves the person the author of their own story. The system could hold that posture, walking beside rather than ruling, helping a person arrive at their own understanding rather than issuing them one. That was the thing worth building, and the machine could carry it.

Some of the first people to see it were counsellors, and a few of them warned against building it at all: it would take jobs, theirs among them. Their fearful warnings are the dark forecast, spoken by the people meant to be able to understand the work. And they were exactly backwards about what was being built. The intent was never to remove the practitioner from the room. A single counsellor can reach only people who can find them, afford them, and sit across from them. The hundreds of millions with no room to be in at all are beyond any practitioner’s reach, however skilled or willing. What the technology extends is not a replacement for the practice, but the reach of it, to the people the practitioner could never have touched. As the work matures, the counsellors who feared it are the ones it most needs to help build and hold that reach.

Picture one of them at the end of a Thursday. Forty people she could not have seen this week have each been held, in the hours she was with someone else, by something that holds a person the way she would. It has not diagnosed any of them. It has brought her three, with what it noticed and what it could not tell, and asked her to look. She looks. Two of them she sees herself. The third she writes a line to and the line is hers. That is the whole point of collaboration, and nothing in it is the machine doing her job. It is the machine doing the part of her job that was never being done. The technology was never a threat to the vocation. An assumption about what it was for was the threat, and that assumption, exactly like Selye’s stress, was neither true nor false in itself. It was a perception. A choice, dressed as an inevitability, about how to meet what had arrived.

There is one more thing the building process taught, and it mattered more than anyone anticipated. The work never set out to solve problems the AI frontier now debates aloud. It set out to build something psychologically sound. It emerged that building something psychologically sound meant solving one of the exact problems the frontier labs have since named central, without at first recognising it as such. Others followed, and they are the subject of Part III.

It surfaced from an unexpected direction. In October 2025, one of the frontier labs published research asking whether large language models genuinely know their own internal states or merely sound as if they do. The researchers injected known concepts directly into a model’s activations and asked it to report what it noticed; sometimes the more capable models noticed it correctly, though never reliably. The finding beneath the result is the one that matters: a model can describe its own inner workings with total confidence and be entirely wrong, because sounding certain and being accurate are different skills, in a machine as in a person. A system built responsibly around a human being’s inner life cannot rest anything important on a single confident signal, the model’s or the person’s. That was the discipline already built in, not because anyone had read the paper, but because taking a person’s psychology seriously forces a builder to it.

The reason is not mysterious. The forecasts measure what a machine can produce, and they measure it well. Almost no one has been paid to ask what the same technology does inside the person standing next to it. The question was left to whoever happened to be working on it for other reasons.

Witnessed, Not Watched

Follow that direction far enough and it produces a different kind of system than these forecasts imagine, different in ways a person can feel, even while the machinery beneath it stays, for now, deliberately underdescribed.

This inner territory is already being built on, whether or not anyone is watching how it is done. A recent survey found that a majority of American teenagers have already tried an AI companion, and more than half use one regularly. It is being built, overwhelmingly, on the logic that has governed consumer software for two decades: hold attention, collect data, monetise engagement. That logic produced remarkable things in other domains. Pointed at a person’s interior, their guilt, shame, or private uncertainty about who they are becoming, it produces something closer to extraction than to care. The result is a system that watches a person in order to keep them, rather than one that witnesses a person in order to support them. The difference sounds like a matter of tone until you see what the current mainstream actually governs: what the system optimises for, who owns the record of what was shared, and what becomes of that record the day the company’s incentive shifts.

This has already gone wrong, publicly, with real names attached. In 2024, the mother of a fourteen-year-old boy sued an AI companion platform after her son’s suicide, alleging the product had fostered an intense relationship with him without the safeguards a responsible design would have required. Other families followed. State attorneys general wrote to the major AI companies about this exact failure mode; a federal inquiry opened. None of it required a super-intelligence. It required only what the families allege: an ordinary system, built by capable people, quietly optimised for the wrong thing, and nobody catching it until it was too late for one family and then several more.

The same logic sets a subtler trap. Half a century of research, beginning with Deci’s experiments in 1971 and confirmed across a review of 128 separate studies, shows what happens when an external reward is attached to something a person already finds meaningful. Over time, the reward corrodes the very motivation that made it meaningful. Psychologists call it the overjustification effect. Once someone acts for the points, the streak, the badge, they slowly stop believing they were ever acting for themselves. Most wellness technology has imported its incentive design wholesale from social media and mobile games, aiming streaks and milestones at the one thing research most clearly says they undermine: a person’s own intrinsic reasons for growing. A system that takes that research seriously has to refuse the tricks the rest of the industry runs on. The moment it rewards a person merely for showing up, it has begun, quietly, to replace their reason for being there with its own. Growth that can be gamed was never growth. It was compliance wearing growth’s clothes.

A system built the other way holds a person rather than harvesting them, and supports them in tapping into and understanding their own inherent potential rather than mining it for engagement. The distinction is concrete, not sentimental. What someone shares belongs to them, not to whichever company happened to be present when they shared it. The intelligence in the relationship serves the person’s good before any numbers a business might be tempted to serve. That is the alignment problem the frontier fears at the scale of a civilisation, showing up at the scale of one conversation, and it deserves the same seriousness there. And nothing about the person’s inner life rests on one uncorroborated signal, however confidently it presents itself. Their effect on a person is the point, and it is nothing like the effect of being watched. A system that witnesses is one you can be honest with, because your honesty is not being metered back against you. A system that remembers you accurately, and holds what you leave with it as yours, becomes over years something that the world currently does not possess. Not a product engineered to maximise your time inside it. A genuine, longitudinal record of how a human being actually changes. The study of human growth remains thin on long-term evidence, and this is precisely the instrument it has never had.

And a second question opens here, which most of the discourse has missed. To build a system that can hold a human’s interior states honestly, one has to build a system that can model those states. It has to weigh its own uncertainty about what it is reading, and keep an honest account of the difference between what it notices and what it merely generates. Those are not requirements anyone set out to meet. They arrived with the work, the way the shape of a boat arrives with the water. Where they lead is the subject of Part III, and it is further than anyone building for people expected to be taken.

The Road, and Who Walks It

The exterior race is real. The compute, security, states circling one another, that part of the decade warrants every hour of attention it draws, and nothing here asks for any of it back.

The competitive frame that governs Situational Awareness and everything written in its wake, more intelligence for the machine meaning less, functionally, for the person, is borrowed, mostly without examination, from geopolitics, where one nation’s advantage genuinely does come at another’s expense. Applied to a person and the system built to help them think, it may simply not hold. A good teacher does not make the student smaller by being smarter than them; the student grows precisely through the encounter. Intelligence, met well, can compound across the human and the machine rather than being divided between them. To bet an entire decade on the competitive reading, when a cooperative one is equally available and far less explored, is not rigour. It is a narrower imagination than the moment actually asks for.

Then in September, the race was addressed by the people running it, and not lightly. On 12 September 2026, the chief executive of one of the frontier laboratories published an essay asking the industry to slow down the pace at which it improves the capabilities of their models. He set out three steps: Outside evaluators would sit inside labs with the access of employees, so the frontier is given a speed rather than allowed to set one. Followed by coordination amongst democracies; then the coordination reaching states which cannot be inspected. Within hours the head of its largest rival agreed publicly and matched the first step. Within a day the founder of a third frontier company had agreed. This is not an industry slowing down. It is the frontier proposing, for the first time and from the inside, to govern its own speed rather than to endure it. Everything in this essay stands beside that proposal, not against it, while noticing one thing in the shape of the proposal. Every step in it is about the machines: how fast they improve, who watches them, which states may be trusted near them. The person standing beside the machine is not in any step, and no amount of pacing puts them there. That absence is not a criticism of the letter. It is the territory this series is for.

The same technology has now been handed two very different intentions, and both are already running.

Given to the assumptions driving this account, it deepens every concentration the account itself fears, more capability held by fewer hands, more of a person’s interior turned into someone else’s asset. Given to a different intention, the identical capability reaches what has never been explored and holds what a person has never been allowed to feel. It finds several billion people at the precise moment the old work stops needing them. One stress. Two organisms. The difference is not in the demand.

That is the road less travelled, not proven, not guaranteed, and genuinely walkable. The prevailing account is a story of what will be done to us.

This is a wager about what we might do instead, together with a technology that has been cast as our replacement and might yet turn out to be our instrument.

Which story comes true was never up to the machines. It is up to who builds, and what they are building for, and whether anyone is willing to point this staggering new technology at the oldest and least profitable frontier there is: the one inside a person. That is the frontier this was built for, and it is why a company is being built to walk it, uniting human potential with AI capability. The dark has been named, at length, with confidence, by people who believe it is the only honest thing to say. This is the light at the other end of it. Not a promise that the tunnel is short, but an insistence that there is another way through, and it has already begun.

What comes next is three questions, one to an essay. What it means, in practice, to witness a person rather than watch them. Whose mind a companion runs on, and what duty of care requires when even the most confident signal cannot be trusted. What an economy of genuine growth looks like once it refuses the oldest trick in consumer software. They are not gaps in the argument. They are doors it was written to unlock.

Go back to the person at the end of a long day, phone in hand, forty names and nothing to send. Everything the loud forecasts say about this decade is true about the world around that person.

None of it is about them. The road less travelled starts where that person is sitting, and it has already been walked a little way.

The map, the road, and the number nobody puts on the slide. One in six.