The Catalogue of Collapses

In the autumn of 1830, farmers across the south of England began to receive letters. They were short, they were signed by a man who did not exist, and they were almost embarrassed by their own occasion:

“Sir, This is to acquaint you that if your thrashing machines are not destroyed by you directly we shall commence our labours. Signed on behalf of the whole, Swing.”

A letter signed ‘Swing’, 1830

The threshing machines were the immediate target. The grievance underneath was that mechanised farming had been eating the winter wages of agricultural labourers for several years before anyone wrote a letter about it, and that whole families were being pushed towards the parish or the workhouse. The men who rode out under the name of Captain Swing were not protesting unemployment in the modern sense. They were protesting the destruction of a wage that had once supported a household. Most of them lost. Hundreds were transported, 19 were hanged, and the machines stayed.

I am asked, more often than I would have predicted, whether artificial intelligence is about to bring on a social collapse. The question arrives at dinner parties, in airport queues, in late calls with old colleagues, and most pointedly from parents whose children are studying computer science. After 30 years in which the only career advice anyone gave a clever child was to learn to code, the suggestion that the same career might be the first to go has a flavour of grim comedy to it. I have written about the modern particulars, the school, the working week, the force multiplier and the habit of thinking for yourself, in Both Readings True at Once. This essay is about the catalogue that sits behind them, because I have read this script before, and the people who delivered it in earlier acts were sincere, articulate, often very clever, and largely wrong about the parts that worried them most. They were not wrong about everything. Some of what they feared came to pass, and some of the work and the lives that were lost were not replaced for the people who lost them, even while the country around them prospered. Holding both of those things in mind at once is the only way to read the catalogue seriously.

The Wage, Not the Job

The shape of the worry, when it is described carefully, is not usually that a new tool will eliminate work outright. The shape that history produces is more particular, and harder to see. When the mechanised loom arrived in a Yorkshire valley, the weavers did not lose their jobs on a Monday morning. They lost the value of their skill over the course of a decade. Wages compressed. Apprenticeships dried up. The number of men holding identical training crept higher than the number of looms that needed them. A man could still call himself a weaver, and still find weaving to do, at a wage that no longer fed a family. The job had not, in any clean technical sense, disappeared. The wage had.

This is what economic disruption looks like in its first cycle: not the dramatic vacancy of a profession, but a slow compression in the price of skills that find themselves in oversupply against a market that is ceasing, in tightening increments, to need quite so many of them. The whole sector does not have to be replaced for this to bite. A shift of 10 or 15 per cent in demand, against the same supply of trained people, is enough to change the lived experience of an entire trade.

I labour the point because the public conversation about artificial intelligence treats job elimination as the relevant variable and largely ignores wage compression, which is the variable that does most of the work in practice. The real worry is not that everyone in a category loses their employment. It is that the value of certain skills falls faster than the people in those skills can retrain into something the market still pays for, and the adjustment is felt as a personal loss long before it shows up as an aggregate statistic. Daron Acemoglu and David Autor have spent the last two decades building the formal version of this argument, Acemoglu with Pascual Restrepo around what they call task displacement, Autor around skill-biased technical change and the polarisation of the wage distribution. What I set down here is the founder’s version of their work, the form in which it has been useful to me in practice, and every entry in the catalogue that follows is an instance of it.

The Oldest Version of the Worry

The catalogue is long, and a fair part of it is comic in retrospect. The oldest entry I have read is in Plato’s Phaedrus, where Socrates tells a story about an Egyptian king, Thamus, refusing the god Theuth’s gift of writing:

“This invention, O king, will create forgetfulness in the learners’ souls, because they will not use their memories… You have found a specific, not for memory, but for reminiscence, and you give your disciples not truth, but only the semblance of truth.”

Plato, Phaedrus (trans. Benjamin Jowett)

The judgement is Thamus’s, but Socrates tells the story approvingly, and it is the closest in spirit to what is being said now. Thamus was not stupid. He was right that something about memory was about to change. He was wrong about the consequences, and Socrates could not have known how much of what he himself said would survive only because his student wrote it down.

The early printing presses provoked an almost identical reaction in compressed form. The scribes’ guilds in the cities where Gutenberg’s machines arrived saw the threat to their craft clearly, and were not always above smashing the presses when they could. The Church and the secular authorities had their own version of the worry, that information unbound from its custodians would corrupt, mislead and inflame, which in various ways it did. Both objections were sincere, both turned out to be partly correct, and neither in the long run kept literacy or the press from spreading.

The Machine Breakers

The Luddites are the textbook case because they gave the language its word, and the word was contemporary with them: the machine breakers of 1811 signed their threats in the name of a General Ludd who was as fictitious as Captain Swing, and the newspapers of the day were already using the name against them. Between 1811 and 1816, framework knitters and croppers in Nottinghamshire, Yorkshire and Lancashire broke into mills and smashed the new machines that were destroying the wages of skilled hand workers. They were not the village idiots that “Luddite” has come to mean. They were tradesmen whose decades of skill were being arbitraged away by capital that could now buy a machine and hire a child to feed it. They lost. Some were transported, some were hanged at York, and the mills kept running.

Captain Swing’s letters came 20 years later, and the machine breaking of 1830 was the largest rising of rural labourers England has seen. Four years after that, in March 1834, the six Dorset labourers now called the Tolpuddle Martyrs were transported to Australia for the crime of swearing secret oaths in forming an agricultural workers’ friendly society. They were not machine breakers in the strict sense. They were trying to combine against the wage collapse that the machines had helped to bring about. The collapse was real for them, and it was not redressed in their lifetimes.

Gaslight and the Family Meal

The objections to electricity were less violent and more decorous, and they came from the same place. Thomas Edison spent the late 1880s prosecuting what came to be called the war of the currents, backing public demonstrations in which dogs, calves and eventually a horse were electrocuted to prove the dangers of his rival Westinghouse’s alternating current. Newspapers carried fervent letters about the loss of gaslight intimacy, the danger of light bulbs in bedrooms, and the moral hazard of being able to read after dark. The telephone, when it arrived, was going to destroy letter-writing, conversation and the family meal. The motor car would kill horses, which it did, and civility, which it did not, exactly, though something was lost there. Television would finish reading. The internet, in the version of the argument that ran in the late 1990s, would atomise communities, hollow out civic life and rot the attention of the young, and a fair part of that has happened, though not in the ways the early critics imagined.

There is a gap in the catalogue. It is dense with named figures from the historical disruptions and almost silent on the modern equivalents. There is no contemporary Captain Swing. No business has yet gone publicly bankrupt and named artificial intelligence as the proximate cause of its collapse. The reason is not that the disruption is not happening; it is that the disruption has not finished. The Swing letters were written in 1830 because the threshing machines had been chewing at agricultural wages for several years before the protest crystallised. This is the middle of those years. The named modern figures, when they arrive, will arrive late, after the disruption has run far enough to be undeniable, and the reader who wants the modern names will have to wait for them.

I find the catalogue useful for the same reason I find old company minutes useful. It is a reminder that smart, serious people, in good faith, with real stakes, have looked at a profound new tool and concluded that the society around it would not survive. They have been doing this for at least two and a half thousand years. The society has, every time, survived. It has not always survived in the form they wanted to preserve, which is a different matter, and it is the matter the rest of this essay is about.

Twelve Years from No Such Thing

Rene Haas, who now runs Arm, tells a story I like about a conversation he had with some interns in 2019. He had been thinking about how much had changed in the 12 years since the iPhone was introduced in 2007, and he ran through the inventory aloud. Twelve years earlier there was no such job as an app developer. There was no such job as a full-stack developer, in anything like the present sense of the term. There was no cloud infrastructure of the kind that runs the modern internet, and no industry of tools and frameworks and observability platforms that joined the pieces together. None of the businesses that people now use without thinking, Uber, Airbnb, Netflix in its streaming form, Spotify, the rideshares and food deliveries and second-hand markets that have remade urban life, existed in the form they now take. By 2019 the people working in those categories, in those companies, on those platforms, numbered in the millions.

The point of the story is not that the iPhone caused all of this; it did not, by itself. The point is that in 2007, standing on the platform of what was then known, no honest observer could have predicted the shape, scale or number of the new categories that would arise. They were not discoverable from where the worry was. The worry of 2007 was about the death of the desktop, the locked-down handset business, the loss of the keyboard. The categories that arrived were almost completely orthogonal to it.

Artificial intelligence is not a smaller change than that. It is larger. That is why I am sceptical of the confident claim, currently fashionable, that there will not be enough work for human beings to do. The claim has the same texture as the claim, in 2007, that the iPhone would not need a software ecosystem because the App Store would only ever host games. It may turn out to be right. It has not been right yet, in any previous instance, and there is no particular reason to expect this case to be the exception.

The Time Dividend

The catalogue has a second lesson, quieter than the first and harder to see from inside any single disruption, and it concerns what happened to the gains.

The working week in Britain in 1800, averaged across the trades that the new factory system was beginning to absorb, was close to 70 hours. The week was six days. The day was twelve hours and often more. The Factory Acts that would, over the next 60 years, slowly bring the day below twelve and then below ten had not been passed, and the worker who protested the conditions of his employment was protesting, in the literal sense, for the right to live in his own house for a few hours each evening rather than at the bench.

By 1900 the figure had moved. The typical British industrial worker was still on a six-day week, but the legal day was capped at ten and in practice averaged closer to nine. The week sat at around 55 to 60 hours, depending on trade and region. It was an enormous improvement in the time available to the worker for his own life, accomplished across three generations, by a combination of legislation, organised labour and the slow rise in productivity that mechanised manufacture had brought.

By 2000 the figure had moved again, and further. The 40-hour week was the norm for full-time work across the developed economies and the five-day week was almost universal in them, and the hours actually worked, averaged across those economies, came in somewhere between 35 and 42 depending on the country. France legislated its 35-hour week in two laws, in 1998 and 2000. Germany was at 38. The United Kingdom, for the average full-time worker, was at 37. The hours per worker had been cut by roughly half over two centuries.

This was not foreseen, in 1800, by anyone in the factories. They were not building the 40-hour week. They were trying to keep their wages from collapsing further as the looms multiplied around them. The 40-hour week is an outcome: an artefact of two centuries of technological capacity being slowly, partially and unevenly converted into time off rather than into pure additional output. The conversion was not automatic. It was political at every step. The Ten Hours Act of 1847 did not happen because mill owners volunteered. The Saturday afternoon free did not arrive because employers decided. When Henry Ford gave his factories the five-day, 40-hour week in 1926 he did it without a strike, and he did it because he wanted workers with the time and the wages to buy cars, which is a reason and not a gift; it took the Fair Labor Standards Act of 1938 to make the 40-hour week American law, and organised labour to make the law bite. Every step of the curve was contested, and each step was eventually accepted by capital because the productivity gains argued over in the previous decade had, by the time of the next argument, made the previous step affordable.

John Maynard Keynes, writing in 1930, gave the worry the name it has carried ever since:

“We are being afflicted with a new disease of which some readers may not yet have heard the name, but of which they will hear a great deal in the years to come, namely, technological unemployment.”

John Maynard Keynes, Economic Possibilities for our Grandchildren (1930)

Keynes was wrong about the timing. He thought his grandchildren, by which he meant the people now reading this, would be working 15-hour weeks by now. He was right about the mechanism. The 40-hour week is itself the artefact of the mechanisation he was responding to; the productivity surplus was eventually banked partly as wages and partly as time. What the long series makes visible is that human time has been a capacity all along. The technology of the eighteenth, nineteenth and twentieth centuries was, in the part of the world that managed to capture its gains domestically, a slow and imperfect mechanism for converting that capacity into time the worker did not have to sell. The hours-per-worker line moves down because the productivity-per-hour line moves up. The relationship is not exact. The conversion rate is set by politics, not by economics. The long trend, in the developed economies, is unmistakable, and one does not need to be optimistic about human nature to see it in the figures.

The quieter worry belongs in the same paragraph. The curve has only worked where the politics allowed it to work. The Industrial Revolution reduced the working week in Britain. It lengthened it in the parts of the world to which Britain exported its industrial model under conditions where the local politics had been removed. There is no law that says a surplus reaches the worker rather than the shareholder, or reaches the country that produced it rather than concentrating somewhere else. The reason for hope is that the conversion has been fought for, and largely won, in stages, over 200 years. The reason for caution is that the conditions under which it was won were not automatic, and the people who would prefer it not be won again are at least as well organised as their nineteenth-century predecessors.

What the Catalogue Cannot Say

Every disruption in the catalogue is informative about the shape of disruption: wage before job, the long tail of wrong predictions, the survivors and the silent losses, and the gains taken as time only where someone fought for them. None is informative about the scale of this one, because the scale of artificial intelligence may turn out to be categorically larger than the scale of any tool the catalogue contains. The analogy carries the shape. It does not carry the scale.

The line “this has all happened before, calm down” is one of the more annoying voices in this conversation, and I do not want to add to it. This has all happened before in its general shape, and never before in its particular shape, and the particulars are where both the harm and the opportunity will sit. I have tried to be specific about them elsewhere. What the catalogue offers is narrower and, I think, more durable than reassurance. It suggests, gently, that there is no reason for despair. It also suggests, in a quieter voice, that despair is a poor reason to be careless about the people for whom the collapse was real.