Governing Cyclopes Starts in the Classroom

I read with great interest the essay Arundhati Roy published in Equator on 15 October, “Apocalypse Now and Then”. Like many authors I love, Roy takes a hard line on the machine, and that line challenges the engineer in me. I understand her view and I think it is legitimate. I fear that her gaze on the machine stops at the harm it causes, while I persist in looking at it from the place where it is built, perhaps fooling myself that there is room there for productive iteration. Each view has a blind spot the other can see, so on the same object they can complete each other. I will try that here.

I start from a sentence that arrives halfway through her essay and that looks to me like the front door. Roy pictures Cyclopes holding the Machine God, still an infant, and inviting us to bow down in return for its razzle-dazzle. Then she writes: “Colonisation always comes bearing shiny gifts.” AI, she adds, is colonisation of a kind few of us can imagine, “capitalism’s final destination.”

The first half of that sentence holds, and the place where it shows best is the school. In July Anthropic launched Claude for Teachers, free for US K-12 teachers; OpenAI has offered ChatGPT for Teachers free since November 2025, plus discounted plans for universities; Google has folded Gemini into the cloud contracts of schools. According to eMarketer’s summary of a Financial Times report, all three companies aim to win over the people who will pay tomorrow. The gift comes first and the bill after. Roy knows the script: in her Kerala story the rayon factory on the Chaliyar river arrives with a promise of 3,000 jobs, and downstream the fish die and thousands of fishermen lose their living.

The second half, “capitalism’s final destination,” says where the road ends before anyone has walked it, and like Cyclopes it cannot be refuted. There is also a difference of substance. Colonisation presupposes a colonised party that cannot refuse, whereas a school, a teacher or a state can say no, or say yes on conditions. Whether the gift becomes a colony depends on the competence of the people who receive it. That makes the education system the crucial knot, because it has to evolve to teach people how to govern the tool.

The ground of knowledge therefore shifts onto the responsibility of those who teach: schools, universities, teachers. For generations the aim of schooling was to compute the result of a complex equation correctly. Today the machine does it in a thousandth of a second, and school has to teach students to set up the problem and to interpret the result critically. The task is to train directors of thought. An assignment built on reproducing facts or mechanical calculation leads to a dead end, because the machine will solve it in the student’s place, and what is needed instead are new kinds of tests, built on logic, oral discussion, creative problem-solving and the breaking down of problems.

The experimental data show how much the design of the test and of the tool matters. In an experiment with nearly a thousand high-school students in Turkey, those who used an assistant that gave the answer improved by 48% in practice sessions and did 17% worse on the exam without it. Those who used a version that gave only hints, under the teacher’s supervision, matched on the exam the students who had only books and notes. The tool is the same and the outcome changes, depending on whether it works as a shortcut or as a gym.

Giorgio Parisi, Nobel laureate in physics, has run the hardest test. In July he solved, with Francesco Zamponi and with the help of Claude, a physics problem that had been open for over ten years: the machine found the decisive idea, and he steered it, correcting its wrong turns. We risk losing our capacity to think, he reportedly told the Italian daily La Repubblica, explaining as well that without the calculations he had done himself he would not have been able to understand the problem in depth, or to guide the machine. Here the thesis about school meets its constraint: setting up a problem requires having done, at some point, the sums. School can move the weight of the test from the result to the defence of the reasoning, but a generation that has never calculated will not recognise a wrong turn. If those who teach turn AI from a shortcut for not thinking into a gym for thinking bigger, Parisi’s fear shrinks, and in that gym the manual work remains the exercise nobody can do in your place.

The role of the teacher becomes showing the limits of the tool. The machine produces plausible answers from past data and speaks in the same tone when it is right and when it is wrong, so only someone who knows the field notices. In five trials with 3,132 people, access to an AI answer that was almost always wrong cut the share of questions people declined to answer from 17–44% to 1–6%. What disappears is the “I don’t know.” Roy speaks of brains that atrophy, and I largely agree, though I might call it a design parameter, because the order in which things are presented changes the outcome. I like to think that writing down your own answer and your own doubt first, and only then looking at the machine’s, keeps the “I don’t know” alive. That is a hypothesis to test, and the classroom is the place to test it.

This also changes how we should read Cyclopes. The image captures the effect and misses the mechanism, because a single head presupposes a single will, and the bodies do fight. After the US government shut Anthropic out of federal contracting, it won on appeal on 25 September; two days later President Trump had dinner with the company’s chief executive, and the day after that he said he might take a stake in the labs. Roy has her answer ready, that the quarrels are theatre too, but a theory that every quarrel confirms cannot be refuted, and a description has to be able to lose.

Shared inputs are enough to make a group move together. Redundant components, for example, fail at the same instant when they share one cause: a power supply, a cable, a supplier. The bodies of Cyclopes share chips that come from a handful of suppliers, the same capital, the same state clients and the same metrics for measuring themselves. A second model that agrees with the first is not a second witness, because it can be wrong in the same way. The difference settles what can be done: a head can be cut off, whereas a dependency shared by many bodies is defused by adding diversity, and the diversity no supplier controls is people trained to judge with their own heads.

A creature with one eye sees without depth, because depth comes from comparing two images taken from different places. The industry is a structure where every evaluation happens from the same side. The labs test their own models on a calendar of their choosing, and Anthropic and Google have put some recent models in the hands of US testers before Britain’s AI Security Institute. On 5 October the New York City Council questioned four companies under oath; each chose its own witness, and none gave a plain answer to three yes-or-no questions. What is missing is a second eye. Engineering has a name for the principle: jidoka, the rule that anyone on a Toyota line can stop it. A check that cannot stop the line is a receipt.

Roy chose the Cyclops, and I keep thinking about how that story ends. Odysseus escapes with a name: he says he is called Nobody, and when the blinded giant shouts that Nobody is hurting him, the other Cyclopes go home. Roy describes the same trick at Minab, in southern Iran, where a US strike hit a girls’ school; she attributes it to automated planning and writes that no one apologised because no one did it. An industry that sells a god has a good legal defence, because a god keeps no access log and cannot be summoned. The remedy is dull, which is why it is realistic: make it reconstructible who authorised what, with permissions, logs and a name under every decision, as aviation does with black boxes. A machine can move a queue forward; only a person owns a verdict. School has its own access log in the oral examination, where the student has to defend, under their own name, what they handed in.

Roy quotes Sam Altman imagining intelligence as a utility, like electricity or water, sold on a meter. Take him at his word: electricity and water are regulated services, with supervised tariffs, universal service obligations and liability for blackouts, because a network is a natural monopoly. The people proposing the metaphor want the meter and leave out the regulator. Roy looks with suspicion at Sanders’s public fund and asks whether getting shares in the apocalypse is our only option, and she is partly right: a share is not a say. A utility gives the public its say by law.

Roy wants to put an end to these companies: “terminate with extreme prejudice.” A structure with many bodies and no head has nothing to decapitate, and it changes when you change what its bodies share: who tests them, who signs for them, who regulates the meter, and how many people can set up a problem and doubt an answer.