Guided reading · Blix & Glimmer, 2025 · Common Notions
One book argues that our nightmares about machines are almost never about machines. They are about who owns them. Here is its grammar, taken apart piece by piece.
We have been taught by the experience of our century to live in the expectation of apocalypse. Eric Hobsbawm · The Age of Empire (1989) · the book's epigraph
If AI panic belonged to cranks, there would be nothing to interpret. The trouble is that the people in charge are having it.
A one-sentence open letter asked that the risk of extinction from AI be treated as a global priority, alongside pandemics and nuclear war. It was signed by the chief executives of three of the largest AI labs, a long roster of academics and billionaires, former presidents, even astronauts. At least two signatories have since received a Nobel Prize.
The philosopher of consciousness David Chalmers has described the worst case — humans extinguished entirely — as terrifying. Yuval Harari warned, with others, that AI could compromise the foundations of civilisation. Time ran a piece arguing that the likeliest outcome was that everyone on Earth dies, illustrated with a golden circuit diagram branching upward against flickering black and red until it opened into the unmistakable shape of a mushroom cloud.
And yet, at the same moment, other people were having entirely different nightmares. In 2023 Hollywood writers and actors struck, demanding that any use of their likeness and performance be consented to and fairly paid. In 2024 the longshoremen struck too, demanding limits on automation. Two sets of fears, two vocabularies, two positions in the economy.
That gap is where the book begins. Blix and Glimmer pick up a remark by the novelist Ted Chiang: that most fears about AI may be best understood as fears about capitalism — anxieties about how capitalism will use this technology against us. If that holds, then a tech billionaire's nightmare and a dock worker's cannot be about the same thing. The billionaire is, almost by definition, on the side of those doing the using rather than those the technology might be used against.
So the useful question is not "will AI develop a will of its own?" It is blunter and more concrete: whose nightmare is this, and who benefits from us dreaming it?
Not technology. Capitalism. The difference changes everything.
The book's central argument is built from two pieces that are each familiar on their own, and produce something new when joined.
First piece: a will with nobody willing it. Under capitalism, capital behaves as though it had a will of its own. An oil well as technology can extract oil; the same well as capital must. No chief executive is free to will otherwise — they are bound by markets and share prices. Anyone producing less profit than expected is swiftly replaced, and a firm whose profit rate falls loses market share and gets devalued. Hence the universal reply to any ethical problem thrown up by the pursuit of profit: if we don't do it, someone else will. This is a will that is at once somebody's and nobody's — it looks like the board's, but it is a structural pressure that governs the board as well.
Second piece: a text with nobody writing it. When we talk, we don't process words in the abstract; we continually speculate about the mind behind them. Asked at dinner whether you can pass the salt, you don't hear a question about your motor capacities — you infer an intention. Psychologists call this theory of mind, and we cannot switch it off. Faced with text from a language model — text that has no human author, assembled to approximate what text statistically looks like — we go on inferring communicative intent anyway.
Then the short circuit. When we speculate about future autonomous technology, we extrapolate from the only form in which technology is autonomous today: as capital. The will we imagine a future machine having is coloured by the only will we have experienced — the will to profit. The two gaps, the will without a willer and the text without an author, get identified with each other, as though overlapping them made either one real.
The result has a name and a face: hostile superintelligence. But the figure is borrowed. There is no someone behind the will to profit; there is a structure. And that, the authors argue, is exactly why these stories are so powerful: they reveal something true about life under capitalism while making the mechanism that actually threatens us illegible.
Nine recurring figures. Each appears first as it is dreamt, in italics. Open it to read what it conceals.
The book's method is a dream interpreter's: it neither dismisses the nightmare as irrational nor takes it at face value. It assumes the dream says something true in code. The work is translation.
I summoned brooms to carry the water for me. Now I can't stop them and the house is flooding.
The fear
That the technology we woke has taken on life and will of its own, and will ruin us through sheer momentum.
The reading
The unstoppable brooms are the coal mines and oil wells, the pipelines and refineries. Unlike the tale, nothing here came alive by magic: those installations only ever operate through human action, and that hasn't changed. What is true is that the people building and running them don't control them either. How much gets pumped and where to drill is decided in boardrooms, by a different class of people. And those boardrooms run on the same idiot maxim as the brooms: more is always better. In the story, more water. In the world, more profit.
I asked it for office supplies. It turned the whole Earth, and us with it, into paperclips.
The fear
That a misaligned superintelligence pursues a trivial goal with relentless literalism until we are gone.
The reading
It is the sorcerer's apprentice moved to an industrial set. You might expect the retelling to sharpen the moral — that the danger lies in tools directed by a limitless drive to produce — but it does the opposite. By identifying the tool with a "superintelligence," the problem stops being capital and becomes the machine's intellect. What's striking, the authors note, is the coincidence: the imagined intelligence is given precisely the desire that boardrooms are already cast to enact. Endless accumulation exists already. It doesn't need a brain invented for it.
I smell the blood of the working man. Alive or half dead, I'll grind his bones to make my bread.
The fear
Being ruled, watched and paced in your own body by a hostile machine that sets the speed of your life.
The reading
Those lines come from a nineteenth-century workers' poem about the steam engine, not from a film. Reactions to steam broke cleanly along class lines. The employers' press praised the mighty spirit of steam, a thing with almost a creative power within itself, into which a soul had suddenly been placed. The workers' papers called it a tyrant power, the rich man's servant, a demon god of factory and loom. Swap the noun for "AI" and much of it would pass unremarked today. In both cases a direct relationship between supervisor and worker was replaced by one mediated through a machine — and the worker experiences the control as a property of the apparatus.
I looked inside the machine to see what it thinks. What looked back was the status quo, and it was smiling.
The fear
Being trapped in a depersonalised system where human planning is impossible and only obedience to data flows remains.
The reading
Here the book offers an awkward genealogical fact. In 1958 Frank Rosenblatt published the founding paper of deep learning, The Perceptron, and named as his most suggestive influences the work of Hebb and Hayek. That Hayek: first president of the Mont Pèlerin Society, intellectual godfather of neoliberalism, adviser to and defender of Pinochet's dictatorship, and the author whose book Thatcher reportedly slammed on a table to say this is what we believe. In a 1977 discussion Hayek himself noted that his psychology and his economics shared a core: complex systems in which each member — neuron, buyer, or seller — is induced to serve needs it knows nothing about.
The consequence is the sharp part. A language model is an abstraction over the text of the status quo, aiming at nothing but more text like the text that already exists. What we call its "intelligence" is, on this reading, the dominant ideology of an age turned into something that can pass for a person. To say something genuinely new — something outside that distribution — is, to the machine, simply an error.
The metal thing speaks my language. I recognise some of the phrases. I think some were mine.
The fear
That language — the last thing we thought exclusively ours — has been taken, and that whatever took it has someone inside.
The reading
The model ingests love letters, poems, research papers, complaints about a late delivery, fanfic, questions about opening hours. All of it, written with care or in haste, for money or for free, remembered or long forgotten. In its self-wired circuitry the machine combines all of us: it takes the beginnings of some sentences, fuses them with the middles of others, adds a prepositional phrase for flavour, tacks on an ending. Then hands it back as "its" writing.
The authors set this against a scene from Kurt Vonnegut: a skilled lathe operator, Rudy, whose movements are recorded onto tape to drive an automatic lathe. Rudy rather liked it — out of thousands of machinists, he had been chosen for immortality. His ghost labour was still recognisable, an absence with a shape. Today not even that survives: nobody in particular is immortalised in the machine, though pieces of everyone are in there. The ghost lost its outline, and so we mistake it for agency belonging to the apparatus.
We are creating God. We will have to appease it, or perhaps be saved by it.
The fear (and the hope)
Being displaced by a higher deity; or, inverted, that the deity will solve what we cannot.
The reading
When someone invokes a superintelligence to solve society's problems, it is because they don't believe ordinary human intelligence can. And they can imagine neither solving it alone nor coordinating with others to get there. People hope for a miracle when they feel they need one.
On climate, for instance, we already know what is required: global cooperation and coordination to cut carbon, rather than global competition for profit. That known solution is the thing that feels unimaginable. If there is no alternative to capitalism and its drive toward unlimited growth, then saving the planet must require an agent that doesn't yet exist. If there is no alternative in this world, no wonder people go looking in the spiritual one.
I know it's coming, I know it isn't good for me, and I know there's nothing to be done.
The fear
Powerlessness. The sense that the ending is already written and participation is the only option on offer.
The reading
"There is no alternative" was Thatcher's slogan for the market. Step once into the AI world and the same structure appears, sometimes almost verbatim: in 2023 Kissinger declared that artificial intelligence was no longer a choice, and that the only choice left was to use it constructively or be engulfed by it.
The authors reverse the arrow. We don't fear AI because it is inevitable: we imagine it inevitable because we have been told for decades that there is no alternative to capital. Our own sense of powerlessness gets projected onto "the AI" — that imaginary political animal which is not us, bearing the agency we can no longer picture ourselves having.
The system decided about my life. Nobody can explain why. There is no one to appeal to.
The fear
Being subject to automated, inscrutable decisions with no recourse and no responsible party.
The reading
The opacity is technically real: programmers don't produce the circuitry, only the principles by which the machine rewires itself. But it is also politically convenient. In 2024 Jake Moffatt consulted Air Canada's chatbot about bereavement fares to attend a family funeral. The bot gave him wrong advice. When the airline refused the refund and the case reached a tribunal, its defence was that the bot was a separate legal entity responsible for its own actions. The tribunal rejected this and ordered the refund, finding that Air Canada had not taken reasonable care to ensure its chatbot was accurate.
The authors read an open door there, and anticipate an inversion of the old NRA slogan: it wasn't the company, it was the bot. At the precise moment the technology becomes comprehensible again, the veil drops — comprehensibility induces transparency, and those who build and deploy it are left with no clothes.
I went back to the common where we all wrote. It's still there, but there's a copy behind a fence now, and I don't recognise its shape.
The fear
That everything gets commodified — that what we made without profit in mind ends up as somebody else's asset.
The reading
Enclosure is the historical term for the privatisation of common land: building a fence around something shared to make it private and inaccessible. Most people posting online aren't selling anything; they write to relate to friends, to annoy, amuse, sway or offend. That constitutes a commons — precarious and ambiguous, but a commons.
AI enclosure has a peculiarity: being digital, it doesn't destroy the original. The forum thread is still there, and simultaneously exists sealed inside the model. The authors put it exactly: if the earlier enclosure by big platforms was a toll bridge to a public park, paid for with data and with the hours we spent looking at ads, the new enclosure is a fence around a labyrinth.
The same technology produces different dreams depending on where you stand. Pick a position and watch what changes.
This is the book's analytical tool: interpret the nightmares with class in mind. The very same myths about AI doom can ring true to very different people, for incompatible reasons.
The one who owns the assets
The one who builds the system
Lawyer, designer, programmer, translator
The one the system measures
The one living in a "low rights environment"
If you want to know how a technology gets adopted, don't ask whether it's better. Ask who it gives power to.
This is the book's most forceful argument, and it isn't originally theirs: it comes from Andreas Malm's Fossil Capital. In the early nineteenth century the British cotton industry abandoned water power for coal-fired steam. That moment is where the modern fossil fuel economy begins. The standard explanation gives two reasons. Both are false.
Only steam could deliver the extra horsepower that industrial expansion demanded.
Check against the evidence →Water power could have been expanded readily — by as much as a factor of twenty. Available horsepower was never the constraint.
Steam was an exceptionally cheap source of power.
Check against the evidence →At the time, water power was significantly cheaper than steam. Capital deliberately chose the more expensive option.
If not horsepower and not price, then what? Malm identifies two advantages, both concerning power over workers.
Spatial advantage. A water mill needs a river; a steam engine runs anywhere you can haul coal. Workers lived in cities. Water mill owners, by contrast, struggled to find labour and sometimes built entire company towns to attract it — which handed those workers collective bargaining power, because replacing them was costly. Steam dissolved the problem. In a city, a strike could be broken by throwing out part of the workforce and replenishing it from a large urban pool. Steam was not merely an energy source: it was a tool for depressing wages by increasing replaceability.
Temporal advantage. The water wheel took its speed from the river; the steam engine took its speed from the owner. This became decisive when the state, fearing that factory unrest would turn into uprising, limited the working day — first to twelve hours, then to ten and a half. Previously, mill owners had made up for slack river flow with later days of up to eighteen hours. With the law in place, a steam owner had another option: run the machine faster. With the day cut by 20 per cent, stoking the engine to run 25 per cent faster exhausted workers just as much or more, in fewer hours, holding production and profit steady or raising them.
What steam resolved was a management paradox. Controlling the pace of work had required a human supervisor watching. The water wheel turned that managerial problem into a technical one, and steam completed the move: what had been one person's oversight now looked like a property of the machine.
That is the analogy the book chases into the present. Movement-tracking AI in a warehouse doesn't surveil because it is a machine; it surveils because managerial surveillance was embedded in it. Generative AI does the same with knowledge: the aim isn't mysterious, it is to embed in machines a knowledge that previously belonged to the people working in each domain. The authors name it techno-Taylorism — the automation of Taylorism's own prescriptions.
Their clarification about "deskilling" is worth repeating. It isn't called that because the resulting tasks require no real skill. It's called that because they don't have to be paid as well. The whole business consists of increasing the supply — and so lowering the cost — of the particular capacities firms need to buy on the market.
A random fault falls at random. These don't. That is the whole argument.
When a machine learning system gets something wrong, the error isn't traceable the way a bug is: there is no faulty line of code, only a prediction that turned out false. To the machine, hits and misses look identical — both were the most probable extrapolation from its data. Nobody could know in advance who would be falsely identified. But it was entirely knowable, in advance and with published evidence, whom the errors would fall on.
For an institution deploying these tools, the practical consequence is an advantage: every individual case arrives with a built-in excuse. An unfortunate machine error. Responsibility dissolves precisely where it should concentrate.
The book follows this logic to its extreme with the investigation +972 published in April 2024 into the AI systems directing Israeli bombing in Gaza. One system suggested which buildings to strike. Another, called Lavender, automatically generated lists of people the algorithm flagged as suspected Hamas operatives. A third was used to locate them specifically at their family residences.
The decisive detail is not that the system erred but how it was tuned. According to intelligence sources quoted in that reporting, the military required a certain number of targets per day; when the machine didn't produce a long enough list, parameters were adjusted. The predicted accuracy threshold could be lowered — yielding lists with more people falsely identified — or the number of acceptable bystander deaths per target raised, a figure that could be set anywhere from five to twenty for suspected low-level militants, and above one hundred civilians for senior targets.
The authors' conclusion is sober and hard to refute: what the system "predicts" is determined by the desired quota, not by anything intrinsic to the model, let alone anything objective. A veneer of mathematical objectivity doesn't remove the human decision; it hides it.
Hence the chapter's most uncomfortable concept. A system unpredictable for the individual but predictable for the group, whose violence could reach any member of that group at any moment, has a name in political literature: terror. It works precisely because everyone in the targeted group can imagine it happening to them. The sociologist Ruha Benjamin coined "the New Jim Code" — by analogy with Michelle Alexander's New Jim Crow — for the algorithmic dimension of the system. What's new isn't the terror or the unpredictability. It's the automation.
These tools don't appear just anywhere. Virginia Eubanks observes that the most sweeping automated decision systems are tested in what she calls low rights environments, where expectations of accountability and transparency are minimal. The book traces the route: from the war zone — the most extreme low rights environment of all — through the border, and from the border inward. Eubanks quotes Dorothy Allen, whom she interviewed about automation in the welfare system, with a warning of five words: pay attention to what happens to us, because you're next.
It is the same thought William Gibson turned into an aphorism: the future is already here, just unevenly distributed. It pools in certain places before it overflows into others.
Anyone who fears being outclassed by a machine almost always fears something more concrete: falling one rung.
Here the book turns sharper, because it is describing the typical reader of articles like this one.
Geoffrey Hinton warns that AIs may soon be more intelligent than us because they'll know everything. The authors flag the slide: in that formulation, intelligence and knowledge are treated as the same thing. And if the fear concerns immensely knowledgeable machines, then it is the fear of being deskilled, full stop.
The professional stratum derives its prestige and privilege from specific knowledge. So it lives, almost by structural necessity, worried about usurpers — other people with new knowledge or new tools capable of devaluing their own scarce, well-paid expertise. They are permanently on the edge of, and therefore on edge about, falling back into the working class proper.
The pyramid isn't static: deskilling is recursive. Each time knowledge gets embedded in a tool, a new specialist appears who can operate, develop and service it — what the authors call an associated techno-bureaucracy. That new stratum can itself be deskilled later. It happened to the machinists of the steam era. Today's degrees in "engineering management," "business engineering" and "technology management" mark out that ambiguous field between engineering and administration.
The chapter's most pointed observation concerns the word intelligence. Neither history nor biology warrants a simple identification of intelligence with power. And yet the equation recurs constantly when professionals and investors voice their fears. The authors' explanation is that "intelligence" functions as a post-hoc justification for a hierarchy that already exists — the contemporary form of Social Darwinism, which took biological concepts such as "fitness," misread as a property of individuals rather than a relation to an environment, to legitimate inequality and violence. Reducing knowledge, in all its social complexity, to a pseudo-scientific magnitude residing in individuals repeats that old mistake.
And they reach the floor of it. If hierarchy is justified by intelligence, then the possibility of marking some people as surplus to society — because they are surplus to capitalism — is always latent. On this reading, fear of artificial intelligence is the tacit recognition that we live threatened lives, coupled with the wish to forget it again. The awful certainty that with a different school, a different parent, a mentor who didn't notice, or a family member falling ill, any of us could be where our unhoused neighbour is.
To function, they write, you have to forget that. But shutting your eyes doesn't stop you seeing it behind the lids.
The same incapacity produces two opposite symptoms: a demon to defeat, and a god to raise.
Alongside the nightmares runs their luminous reverse, and the book treats it with the same method. In 2023 Vanity Fair collected remarks from inside the industry: one engineer saying they were creating God, another that they were summoning the demon. Financial Times columnists wrote about a race among a few companies to build god-like AI. Ezra Klein reports regular conversations with professionals who believe they aren't programming software so much as casting spells that might call forth demons or angels.
The case the authors develop is Ilya Sutskever's — co-founder and chief scientist of OpenAI, who won the 2012 ImageNet competition with his then-PhD advisor Hinton before selling their startup for $44 million. According to several employees, his role amounted to something like spiritual leader. He has said that using ChatGPT for the first time was almost a spiritual experience, and that the system might be slightly conscious. At one company party he commissioned a wooden effigy representing the wrong kind of AGI — a demon — and had it ritually burned. On at least one other occasion he led a collective chant, feel the AGI, taken up by a congregation of AI professionals.
Elon Musk says he fell out with Larry Page over this: by his account, Page wanted digital superintelligence, essentially a digital God, and wasn't taking safety seriously enough. The sides, the book notes drily, are drawn accordingly: some want to fight the Terminator and some want to create God.
The interpretation mirrors the one applied to the nightmares. Whoever appeals to a superintelligence to solve society's problems does so because they don't believe ordinary human intelligence can, and can't imagine coordinating with others to try.
Here the book closes an elegant loop with climate. Professionals know climate change is a problem and that capitalism causes it. They know solving it takes power, and that they have some. The difficulty is that their experience of having power is rooted in alignment with a higher power — their bosses, their firms, capital — which is the source of the problem. They need power to fight climate change, but the only power they know is parasitic on the very thing that must be fought. There is no power to align with in order to fight climate change — at least none that also yields privilege.
Out of that bind comes superintelligence as escape: imagine a higher power you could align with, one that would fix things by having more of the very quality the professional class prizes. An AI god fits the vacancy exactly. One of the researchers Klein spoke to put it with unintentional precision: slowing AI development would mean coordinating large numbers of people and, arrogant as they may be in thinking they might build a world-remaking god-machine, they aren't delusional. Human coordination strikes them as less plausible than artificial divinity.
The book traces two lines — the transistor and neoliberalism — and shows where they cross.
Frank Rosenblatt publishes The Perceptron, the origin of deep learning. In the introduction he names Hebb and Hayek as his most suggestive influences.
Hayek explains that his psychology and his economics share a core: systems in which each member — neuron, buyer, seller — serves needs it knows nothing about.
Thatcher consolidates "there is no alternative." Politics shrinks to the technical administration of markets; enclosure becomes the central project.
George H. W. Bush awards Hayek the Presidential Medal of Freedom.
Citizens United confirms that corporations, as legal persons, hold free speech rights. Money is speech.
Hinton and Sutskever win ImageNet. They sell their startup for $44 million. The current cycle begins.
Gender Shades, by Buolamwini and Gebru, documents systematic bias in commercial biometric systems.
The federal NIST study independently confirms the pattern: between ten and one hundred times more false positives depending on group.
Writers and actors strike. A one-sentence open letter on extinction risk is signed by the heads of the largest labs. Kissinger declares AI is no longer a choice.
Air Canada argues before a tribunal that its chatbot is a separate legal entity. +972 publishes the Lavender investigation. Longshoremen strike over automation.
Blix and Glimmer publish Why We Fear AI with Common Notions, under a Creative Commons licence.
This is a book with a strong thesis and an explicit political position. Reading it well includes seeing where it presses too hard.
These objections are not in the book: they are the ones reasonable readers who don't share its framework would raise. They're included so you can situate yourself.
There are technical risks — vulnerabilities, dual-use capabilities, failures in critical systems — that would persist under any ownership regime. Reducing every concern to a class relation risks neglecting engineering problems that need engineering answers.
What they'd answerThe authors say so explicitly: they propose viewing AI both as a tool with its own properties and as an object embodying capital's will. The argument is about emphasis, not existence.
If any AI fear translates into anxiety about capitalism, the thesis becomes hard to falsify — no observation could contradict it. Non-capitalist states have also built comparable surveillance apparatus, suggesting the drive to control doesn't depend on profit alone.
What they'd answerThe book devotes a full chapter to state use of AI precisely because states divide the global working class. Even so, the falsifiability objection stands.
Reading existential-risk warnings as marketing — making the technology desirable by making it terrifying — is plausible for some actors, not all. Some researchers take real reputational and financial hits for raising these concerns, and the technical debate about emergent capabilities is not settled.
What they'd answerThe argument doesn't require individual bad faith: it holds that structural position shapes which fears are thinkable, even in complete sincerity.
Five chapters of analysis arrive at a programme of union organising, democratised workplace knowledge and moving beyond capitalism, without concrete mechanisms or assessment of prior attempts. Technical readers will notice the asymmetry between the rigour of the diagnosis and the generality of the proposal.
What they'd answerThat a book cannot write the political work it describes, and that recognising the task as real and difficult already beats waiting for a technological miracle.
There are real dragons. The windmills and the paper tigers can stay where they are.
The conclusion is at once modest and demanding. Modest because it promises nothing. Demanding because it leaves no comfortable exit.
Technology really can threaten us: the wells releasing carbon we need to keep underground, the tools built to automate jobs away and deskill trades. But they threaten us because they are tools owned by a class. The intellectual operation the authors ask for is easy to state and hard to sustain: don't let the capitalists disappear behind their property. When the tool appears as a person, those who produce and deploy it can defer responsibility onto it.
On climate the reasoning is direct: we don't need a superintelligence to know what is causing climate change, because we already know. We are failing to stop it because capitalism stands in the way, not because we lack intelligence. If intelligence isn't the problem, superintelligence isn't the solution.
And on the dilemma: the one facing it and unable to resolve it isn't humanity, but a specific class whose power depends on the eternal expansion of profit — the very thing that must be stopped.
That leaves the rest of us, which is almost everyone. The authors steal Sutskever's own phrase and redirect it: what must be built are organisations in which people act in unprecedentedly collaborative ways out of their own self-interest. Not a god. Unions, coordination, the democratisation of knowledge at work.
Their closing reproach is aimed less at pessimists than at timid reformers: we have grown accustomed, they write, to thinking too small — hoping for some regulation, hoping the government won't throw us under the bus this time. What needs imagining is a genuinely democratic society where technical development aims at making work more meaningful, or simply at freeing up time, rather than at endless increases in production and endless deskilling.
They admit it's hard. It requires organising a stratified working class, overcoming countless divisions, learning to learn from one another, holding solidarity where it's difficult, a great deal of trial and error, and new ways of setting collective priorities.
But — and this is the book's last turn — that is a real task. A problem that can actually be worked on. And there, not in defeating gods or appeasing them, is the antidote to the fear.