AI, Empire and the Living World — Arundhati Roy and Naomi Klein, with Karen Hao

Arundhati Roy, Naomi Klein · 2026 · 50:52

Summary

The central question. What is AI doing to us, and why is it spreading so fast? The panel gives two answers, and they overlap. Roy’s answer: AI attacks the one thing that holds a society together, language, and it is a new form of an old pattern (dams, mining, the East India Company) in which the harm is hidden and the benefit is advertised. Klein’s answer: AI spreads because modern life already feels “impossible and inhuman”, so people want to “subcontract” their lives to bots. The bots are “the closers in a bad deal” that is decades old.

The main thesis, in five claims.

The AI firms are one project. The big labs are “tuning differences” on the same engine (Klein,). They fight each other, and they coordinate. Roy’s image is a single-headed, many-bodied Cyclops (named by Hao,).
The prize is everything. Klein reads the firms’ own claims (a “$28 trillion” or “$30 trillion” market) as a plan to swallow the whole economy. The people left over become “surplus humanity”, whose function is “to be surveilled” and “contained”. Roy adds that these firms seek to “fuse themselves with the state” like the East India Company.
The harm is already here. Gaza, Minneapolis, France, India. Not “something bad might happen in the future”.
The pro-AI case is real but is the old drug pitch. AI does help people. Roy: “you can’t produce morphine and say have this, you’ll feel better”. Klein: the demand for chatbot tutors only makes sense once the teacher has 40 students.
The way out is a broad movement plus a plan for the crisis. A “welcoming” movement with few purity tests (Klein,); a plan for the “window” when the bubble pops (Klein,); and for Roy, a refusal to look away, plus joy.

How it closes. Klein uses Milton Friedman’s “ready ideas” theory against the right: “We’re going to have another window and we need democratic socialism”. Roy reads her “writer’s manifesto”. The room chants When people rise, empires fall.

Stance to note.

Roy is the more cultural and moral voice. She writes about language, mystery, and stupidity. She mocks the AI founders (“low IQ”,; “lizard blood”).
Klein is the more political-economic and strategic voice. She reads investor documents, sets out the bubble risk, and asks how to build a movement.
They agree on the enemy. They differ in emphasis: Klein wants a movement that does not shame users; Roy says the benefits “are nothing compared to the costs that generations are going to pay”.
Both quote the AI leaders’ own doom talk (a 10% chance you’ll be annihilated, Roy; the Coxon resignation) but read it as ideology or pathology of capitalism, not as a neutral risk estimate.

Map — talking points

#	Point
1	Hao introduces the panel; two “living legends” on anti-imperialism and anti-capitalism
2	Roy reads Apocalypse Now and Then: “Think or sink. Comply or die”; the em dash as a “tell”; locked out of language
3	Roy: the first verse of John (“the word was God”); AI disrupts the human-language bond and “social coherence”; the taxidermist’s Shangri-La
4	Roy: mass psychosis? “Moral panic” objection raised and left open. Coxon quits Anthropic; CEOs’ “tie me up” plea for regulation
5	Klein reads: March 2020 relief; the wish to live so that cancelled commitments would be a loss
6	Klein: seminar on fascism; Arendt’s inner dialogue; a student says “everyone wants their own slaves” (AI agents)
7	Klein’s thesis: people use AI to escape lives that were already “impossible and inhuman”; “the bots are the closers”; we already talk like chatbots
8	Klein: regulate, nationalize, or abolish? A narrow window when the bubble pops. The larger question: make life livable
9	Klein quotes End Times Fascism: apocalypse culture and social crisis are twins; “to beat the tech oligarchs, we will all have to become more human”
10	Hao to both: why write when language is “frictionless”? Roy: “to stay sane”; Berger: “I will never be zero”
11	Klein: intelligence is not a vending-machine utility; the point is to have made the map. Berger and “calm”
12	Roy: art is mystery; the AI men treat it as toothpaste; “low IQ”; “lizard blood”
13	Hao: one ideology, a Cyclops. Klein: Coke vs Pepsi; coordinated; TAM claims; AGI = doing most human activity; the “permanent underclass”
14	Klein: surplus humanity must be surveilled; Gaza as test bed; France; the state crushing protest
15	Roy: companies fuse with the state; East India Company; surplus people; cameras on foreheads; voter-roll deletions
16	Roy and Hao: all fronts are one front; the opponent is also more unified
17	Hao Q1 (to Klein): “I feel guilty using AI at work. Am I a bad person?” Klein: fast organic movement; no purity tests; make life worth living
18	Hao Q2 (to Roy): “But AI helped me.” Roy: no anti-AI movement in India; Oxycontin; the tsunami passage; cognitive decline; morphine
19	Klein: the over-leveraged build-out, “$8 trillion”; Green New Deal “no money”; the teacher with 12 vs 40 students
20	Roy: privatization pattern; “privatize history and sell it back to us by the meter”
21	Hao asks for joy. Klein: Friedman’s theory of crisis; “we need democratic socialism”
22	Roy’s writer’s manifesto; call-and-response chant

Putting it together

The mental model in one paragraph. Roy and Klein offer two linked stories. The supply story (Klein, Roy) says the major AI labs are one project: a coordinated bid, financed by a possible bubble, to own the means of producing language, art, and decisions, and to fuse with the state. The demand story (Klein) says people reach for AI because ordinary life, work and public services have already been hollowed out, so the bots arrive as “closers in a bad deal”. Roy adds the cultural cost: language, trust, and the work of making meaning are damaged first, and the harm shows later, like a tsunami behind a retreating sea. The political answer is a broad, non-moralising movement; policy ideas ready for the crash (regulate, nationalise, or abolish); and a refusal to look away. The best objections are: the labs differ more than “one Cyclops” allows; the harms are partly older than AI; the evidence on cognitive decline is thin; the crash may help incumbents; and AI’s real benefits fall hardest on people the critics most want to protect.

Open tensions to keep in mind

Cause or instrument? Is AI creating these harms (voter roll politics, surveillance, precarity) or speeding harms with older roots? Klein’s “closers” phrase implies the second; some of Roy’s examples sound like the first.
Hype or hazard? Both guests are unsure the technology will work as promised, yet both cite its leaders’ doom warnings. The two stances need a way to be told apart.
Purity or critique? Klein wants a welcoming movement; Roy says the benefits “are nothing compared to the costs”. How strong a message can a welcoming movement keep?
Window or trap? A crash is both a chance and a risk. The same event may help the organised side, and that is not always the left.
One empire or many? The “single Cyclops” frame is strong for finance and ideology. It is weaker for the open-weight and non-US parts of the field.
Local vs global. Klein speaks about a US-based organising wave. Roy says India has none. A movement for people “all over the world” has to answer this gap.

Reading path

Start: Karen Hao, Empire of AI (2025), the first chapters and the OpenAI-nonprofit story. Klein and Taylor, End Times Fascism (2026), the introduction. Roy’s essay once published in full (the title and venue were not confirmed).
Then: Naomi Klein, The Shock Doctrine (2007), the introduction and the Friedman chapters, alongside the libertarian critique (Libertarianism.org). Karl Polanyi, The Great Transformation, chapters 6 and 12. Hannah Arendt, “Thinking and Moral Considerations”. Kate Crawford, Atlas of AI.
Test it: Arvind Narayanan and Sayash Kapoor, “AI as Normal Technology” (Princeton). Carlota Perez, Technological Revolutions and Financial Capital. Daron Acemoglu and Simon Johnson, Power and Progress.
Deeper: Mike Davis, Planet of Slums; Arendt, The Origins of Totalitarianism (the section on superfluousness); Dalrymple, The Anarchy; Couldry and Mejias, The Costs of Connection; Lewis Hyde, The Gift; Benjamin, “The Work of Art in the Age of Mechanical Reproduction”; Roy, “The Greater Common Good” (1999) and Azadi (2020).

Part 1 — Before you read

Key terms

Empire. In Hao’s book and in this talk, “empire” means a power that takes land, labour and resources from others, and justifies it with a story about progress. Hao uses it for AI firms. Roy uses it with an older, literal case: the East India Company.
Cyclops. Roy’s image (as named by Hao) for the AI industry: one head, many bodies. The claim is structural: firms look separate but share one logic and one goal.
AGI (artificial general intelligence). Klein gives the industry’s own working definition: reaching the point where the technology “performs the vast majority of human activities”. She says that “literally means we’re out of luck”. Other definitions exist; this is the one she attributes to them.
TAM (total addressable market). The most revenue a company could earn if it sold to everyone who could possibly buy. It is a pitch number. It is not a forecast. Klein reads it as a statement of intent.
Surplus humanity. People the economy no longer needs as workers or as buyers. See idea 4: this phrase has a history (Marx’s “reserve army”, Arendt’s “superfluous”, Mike Davis’s Planet of Slums). Klein says Silicon Valley’s own term is the “permanent underclass”.
Permanent underclass. A 2026 Silicon Valley meme: if AI replaces most labour soon, class positions freeze, with capital owners on top. Its spread is described in Big Technology, which also gives sceptical views (see the debate section).
Fascism (as used here). Klein and Taylor’s book frames the current political-tech alliance as “end times fascism”. Klein’s seminar asks whether AI is “a fascist technology seeking to replace human thought and agency and tightly control what passes for truth”. Watch the word: fascism has competing definitions (idea 4).
Shock doctrine. Klein’s 2007 theory. Elites use crises (wars, disasters, crashes) to push through policies the public would reject in calm times. In this talk Klein turns it around: the left should have ideas ready for the next crash.
Purity test. Demanding that people be fully free of a harm (fossil fuels, AI use) before they can join the fight against it. Klein warns against it. Roy jokes about it in terms of Gandhi’s loincloth.
Inevitability narrative. The claim that AI’s spread cannot be stopped. Hao calls such stories “hegemonic narratives of progress and inevitability”; the panel treats them as a power move.
Two uses of “intelligence”. Altman’s “intelligence as a utility” (a metered product, see below) versus Roy’s and Klein’s sense of intelligence as the result of effort and relationship (Klein: “there’s nothing better than to have written”,). They are not arguing about the same object.

People named (a reading list)

Person	Why they come up	Start with
Hannah Arendt	Klein: the inner “dialogue between me and myself” protects against evil	The Life of the Mind (1978), “Thinking”; “Thinking and Moral Considerations” (1971)
John Berger	Roy: “I will never be zero; Klein: his blurb said the book made him calm”	Ways of Seeing (1972); his letters and essays
Milton Friedman	Klein: “only a crisis… produces real change”	Preface to Capitalism and Freedom (1982)
Sam Altman	Klein: “intelligence is a utility”	His March 2026 remarks (see debate section)
Jacob Coxon	Roy: the Anthropic resignation	News reports, Sep 2026
Astra Taylor	Klein’s co-author	The Age of Insecurity (2023); End Times Fascism (2026)
Emily Dickinson	Roy: the dash as signature	Her poems; the em dash is a mark of her manuscripts
Gandhi	Roy’s joke on purity	Hind Swaraj (1909) for his critique of machine civilisation
Elon Musk, Anthropic, Nvidia, Palantir, Meta	The “Cyclops” bodies	Company filings (S-1 for SpaceX)
Karen Hao	Moderator; frames “empire”	Empire of AI (2025)

The speakers’ own frame

Roy writes as a novelist-activist. She has a record of fights against large dams (the Narmada), mining and displacement, and wrote The Algebra of Infinite Justice (2001) and Azadi (2020). She says here: “the science of AI is new, but the way it’s been forced on us is very, very old”. So she reads AI as the next item in a series. Her mode is literary and moral: language, stupidity, mystery, courage.
Klein writes as a political economist of crisis. Shock Doctrine argues that free-market policy spreads through disaster. Doppelganger (2023) argues that the internet makes us into copies of ourselves. End Times Fascism argues that a billionaire-religious-nationalist alliance is preparing for collapse and expects to ride it out (Democracy Now). Her mode here is empirical (investor documents) and strategic.
Hao frames the whole evening with her book’s thesis: the AI companies as an empire. Her questions steer toward what people should do (guilt, benefits, joy).

The current debate (2025–26)

These are the live arguments the talk touches. I give both sides. Sources are linked. Some are secondary summaries; I say so where it matters.

Is the money real? The bubble question. Klein says the industry is “incredibly overleveraged” and has found “around eight trillion dollars” with no matching market. Sceptics: a Bloomberg analysis (republished by ThePrint) reports investor worry that AI revenue may not justify trillions in data-centre funding, and cites Bain’s estimate of under $2 trillion in AI revenue by 2031. Defenders: KKR’s Anthony Wanger argues that “bubble” is misplaced because today’s spenders are the largest, most stable firms (Bisnow). A middle view, from economic history: bubbles often leave useful infrastructure behind (Carlota Perez, idea 5).
Is “TAM $30 trillion” a plan or a pitch? Reports say Anthropic would tell IPO investors its market tops $30 trillion, counting the full range of human work AI could in principle do (The Next Web). Commentators call a TAM a ceiling, not a forecast, and note Wall Street scepticism. Klein reads the same figure as a statement of goals.
Self-regulation. On 29–30 Sep 2026, leading AI firms signed a voluntary “White House Accord on Superintelligence” at an event with Trump, who called it “almost like a Constitution” (Axios). Critics say it has no enforcement (Computing). Roy’s “tie me up, tie me down” joke targets this pattern: a call for rules from the people who can write the rules.
Insider warnings. Coxon’s resignation drew very wide attention. Reports say Anthropic’s alignment lead Evan Hubinger said he shared the concern and put the chance of AI killing everyone within a decade above 10% (AI Weekly). (Secondary report; I did not confirm with a primary post.) Roy and Klein treat the form of these warnings as part of the problem. Others take the risk estimates at face value. Part 3, idea 7.
“AI as normal technology.” Princeton’s Arvind Narayanan and Sayash Kapoor argue AI should be seen as a normal, general-purpose technology like electricity: slow to spread, controllable, and best governed by ordinary institutions, not as a separate superintelligent agent (Princeton CITP). This is a third position: neither “AGI is coming” nor “it’s all hype”. It is useful for testing both Roy and the lab leaders.
Brain and learning evidence. The MIT Media Lab study (“Your Brain on ChatGPT”, 2025) is the best-known evidence for “cognitive debt”. It has 54 participants and was a preprint, and a published comment raises concerns about sample size, EEG methods and reporting (arXiv 2601.00856). I found no 2026 replication. Roy’s “studies show” claim is therefore suggestive, not settled.
Energy and climate. Klein says AI’s climate impact is “disastrous”. The IEA’s 2025 Energy and AI work (as summarised by Carbon Brief) puts data centres at about 1.5% of world electricity and about 0.5% of CO₂ emissions today, expected to roughly double in electricity by 2030. That is growing fast but small today. Which description fits depends on the time frame and on local effects (water, grid strain), which global shares hide.
The anti-AI movement. Klein says she has never seen a movement “rise up this quickly this organically”. Sept–Oct 2026 events I found include a “Stop the AI Race” rally outside Dreamforce in San Francisco and the Circuit Breakers tech-worker gathering in New York (17–18 Oct) (from search results; not checked in detail). Roy says there is no such movement in India. Both can be true.
India’s Gen Z “cockroach” protests. Roy mentions them. They began over leaked exam papers and spread into a wider youth movement in 2026 (NPR, PBS). Roy’s point: the same tools that carry the protest can be switched off by internet shutdowns.

The ideas

Roy and Klein speak for 50 minutes to a room that knows these traditions. They do not name most of them. Here they are laid out. Eight families of ideas meet in this talk. None of this is the speakers’ claim unless a timestamp in the ideas says so.

 POLITICAL ECONOMY OF TECH ────┐                    ┌──── EMPIRE, COLONIALISM, DISPOSSESSION
 Marx · Polanyi · Zuboff       │                    │     Luxemburg · Harvey · Said · Fanon
 Srnicek · Crawford · Hao      │                    │     Couldry & Mejias · Dalrymple
                               ▼                    ▼
                       ┌─────────────────────────────────┐
 CRISIS & SHOCK ─────► │  ROY + KLEIN: AI is the closer  │ ◄──── SURPLUS PEOPLE & FASCISM
 Friedman · Gramsci    │  of an old bad deal, run by one │       Arendt · Bauman · Davis
 Polanyi · Klein       │  coordinated empire; the answer │       Mbembe · Paxton · Eco
                       │  is a broad movement + a plan   │
                       └─────────────────────────────────┘
                         ▲             ▲             ▲
   LANGUAGE, ART, MYSTERY│   WORK, BURNOUT, CARE     │ MOVEMENTS & STRATEGY
   Benjamin · Hyde       │   Marx · Han · Graeber    │ Luddites · Narmada · Climate
   Adorno · Bender       │   Fraser · Aristotle      │ Chenoweth · Polanyi
                         │                           │
                         └──── THE COUNTER-CASE ─────┘
                           Perez · Acemoglu · Narayanan · Amodei · Andreessen

Where each school of thought is covered:

Political economy of technology — who owns the machine, and what is it for?: idea 6
Crisis and shock — “keep the ideas ready”: idea 6
Surplus people — from “reserve army” to “permanent underclass”: idea 4
Fascism and technology — three different questions: idea 4
Empire, colonialism and dispossession — “this is East India”: idea 3
Language, art and mystery — what is lost when words become supply?: idea 1
Work, burnout and care — why people reach for the bot: idea 2
Movements and strategy — and the strongest counter-case: idea 5

1. Language, writing and what is lost (Roy’s essay; the writers’ exchange)

In short. Roy opens by asking whether we are being gradually locked out of language. She argues that disrupting the relationship between humans and language will undermine social coherence. She names two harms: distrust in schools and publishing, and a loss of knowing who we are. She says she writes to stay sane, as “language is the skin on my thought”.

🎙 In the talk.

Roy opens with a mock-alarm: “Think or sink. Comply or die. Look up or be left out”. She notes she has “already used two dashes in this opening paragraph. That, as they say, is a tell”, and so “perhaps I haven’t written this essay at all”. “You can no longer use your instincts. You have to use the app.” Her question: “Are we being gradually locked out of language?“.
Roy: “I believe that disrupting and vitiating the fundamental relationship between humans and language will undermine all semblance of social coherence” (caption reads “visiating”; likely “vitiating”).
Roy names two harms: distrust in schools and publishing (“detectives inspecting a potential crime scene”) and a loss of knowing “who we are or even if we are”. Her image: “A taxidermist’s Shangri-La in which everything that was once alive has been skinned and stuffed”.
Roy then asks the sceptic’s question herself: is this “the inevitable moral panic” that met the industrial revolution and the computer? “Is that all it is?“. She does not answer it directly. She moves to “Or is this indeed a catastrophic rupture?“.
Hao asks why write “in a moment when language is supposedly so frictionlessly abundant”. Roy: “I write to stay sane… so that the blood flows easy in my body”; “language is the skin on my thought”. She retells Berger: you write “to tell them I will never be zero”.
Klein: Altman’s idea that “intelligence is a utility” misses that writing is hard and that “there’s nothing better than to have written, to have figured it out” because “you re-enter the world with that sanity that you earned often through insanity”. Berger’s blurb said her book “made him feel calm”: “when somebody like you makes the effort to order it and tame it, that lets your readers feel like, okay, someone’s got this”.
Roy: “at the heart of it is mystery… you cannot say why the world loved this particular band or this particular book”. The AI founders “look at it like a product no different from mosquito repellent or toothpaste”, so “I really find them low IQ”. “AI is like asking me to drink lizard blood”.

📚 Deeper theory

Two ideas of language. One treats language as transmission of content (a pipe). The other treats it as a medium in which thought forms (Benjamin 1916; Roy’s “skin on my thought”). If language is a pipe, a machine that moves content faster is an improvement. If it is a medium, then outsourcing the writing outsources some of the thinking. The talk assumes the second (Roots).
Why the em dash is a good joke. Detection by style is a proxy test. The point of Roy’s gag is that proxy tests fail and cost trust: Dickinson used dashes for a century and a half before any model did. AI-text detectors are known to be unreliable, which makes “distrust everyone” the likely outcome (general knowledge; no source checked). The stronger claim is about trust, not style.
Writing as thinking. Writers and cognitive scientists (Flower and Hayes, 1980s; Didion, Orwell) treat the act of writing as how one finds what one thinks. Klein’s “there’s nothing better than to have written” says this in a sentence.
The “moral panic” test. Roy raises it herself. Historical panics (Socrates on writing in the Phaedrus; the printing press; television) are often cited. Socrates’ worry that writing would weaken memory was partly right and writing still won. The right test is not “was the panic wrong before?” but “what did the loss cost, and who paid?”
“Taxidermy” and the Chinese Room. John Searle’s Chinese Room (1980) argues that symbol manipulation is not understanding. Roy’s “meaningful glint in its bright, spurious eyes” is a literary form of Searle’s point. The reply (the “systems reply”, and modern work on emergent abilities) is Part “push back” below.
Hao’s own question carries a premise: that language is now “frictionlessly abundant”. Text is abundant. Trust in text is not.

📚 Roots: Language, art and mystery — what is lost when words become supply?

The problem. Roy’s essay starts here. Is language only a tool for passing information, or is it the thing that makes us a we?

The point. The two opening speeches set language against supply. Altman’s line, “intelligence is a utility… people buy it from us on a meter” (reported from a March 2026 BlackRock event; Gizmodo via bioethics.com), is the thing they answer. The two sides treat writing differently: for the industry it is a product, and for Roy and Klein it is a process that changes the writer.

The chain

The Gospel of John, 1:1. “In the beginning was the Word.” Logos means word, reason and order. Roy says she is “enthralled” by it “for reasons that have more to do with the writer’s ego than with faith”.
Walter Benjamin, “On Language as Such and on the Language of Man” (1916). Naming is the human share in creation; language is not a tool but a medium in which things become knowable. And, in the 1935 essay, the aura: the unique presence of an original, which mechanical copying wears away. Roy’s “mystery… you cannot say why the world loved this particular band or this particular book” is close in spirit to the aura. (My link; Roy does not cite Benjamin.)
Theodor Adorno and Max Horkheimer, “The Culture Industry” (1944). Art made as a standard product, to fit standard needs. Their phrase is “enlightenment as mass deception”. Roy’s toothpaste or mosquito repellent is the same insult.
Lewis Hyde, The Gift: Imagination and the Erotic Life of Property (1983). Art is a gift economy, not a market. It passes between people and creates bonds. Commodifying it destroys something. This is the best fit to Roy and Klein’s “you can’t turn on a tap” claim.
Bender and Koller, “Climbing towards NLU” (2020) and the “octopus test”; Stochastic Parrots (2021). The scholarly form of the claim that fluent text is not understanding. Noam Chomsky, Ian Roberts and Jeffrey Watumull made a similar claim in a 2023 New York Times essay (from memory). The counter-view: scholars such as Geoffrey Hinton say models do build a kind of understanding (idea 1).
George Orwell, “Politics and the English Language” (1946); Victor Klemperer, LTI: The Language of the Third Reich (1947). Degraded language enables degraded politics. Roy’s “locked out of language” and “truth from falsehood” belong here.
John Berger, Ways of Seeing (1972) and his letters to Roy. Berger is the figure both women quote. The two quotes give two ideas about why to write: defiance (“I will never be zero”) and order (a blurb saying the book made him “calm”). Berger’s essays on seeing and ownership of images are the closest English-language cousin to Roy’s language argument.
Joan Didion, “The White Album” (1979). “We tell ourselves stories in order to live.” It is the same claim as Roy’s “I write to stay sane” and Klein’s “ordering the world”. (Not cited by either; offered as a parallel.)

Where others push back.

Understanding is not binary. Geoffrey Hinton and others argue that models trained to predict text must build internal models of the world to do it well, so “they only predict” says little. Whether this amounts to understanding is disputed among researchers. Roy’s image (“skinned and stuffed”) takes a side.
Tools extend writers. Many working writers use AI for translation, editing, or access (dyslexia, a second language). Roy’s own line, that AI can “translate [a novel] into any language you want”, is acknowledged as a benefit and then set aside. The critic’s reply: those uses are real and not what drives the investment.
Dickinson’s dashes and detectors. Others say the em-dash panic proves humans are the weak link: people over-read signals. That is an argument for better practice, not for fearing the tool.
“Low IQ.” is a rhetorical flourish. Whatever else one thinks of the AI founders, their technical and financial record does not support “low IQ”. Read it as an argument about taste and wisdom, not intelligence tests.

What they leave out. The gains in access: speech-to-text, translation into Indian languages (Roy comes from a country of many languages,), low-cost education tools. And the interplay with Indian languages: models are weaker in most of them. That is a claim for Roy’s worry (flattening) and also a claim against it (no mass flattening yet).

Check yourself

What is the difference between “language as a pipe” and “language as a medium”? Which does Roy hold, according to her own words?
Roy asks whether this is a “moral panic”. What would count as evidence one way or the other?
Hao asks why write if language is “frictionlessly abundant”. Why do both guests answer about the writer, not the reader?

2. Why people use AI: “the bots are the closers in a bad deal” (Klein)

In short. Klein asks why people use AI. Her answer is that so many people are desperately trying to get out of their lives, and with good reason. The bots are just the closers in a bad deal stretching back decades. So the task is to make life more livable, so we are less tempted to subcontract our lives to bots.

🎙 In the talk.

Klein tells of March 2020: “intense relief” at cancelled work, then “a creeping concern about my mental health”. Her resolution: “I want to live differently in such a way that if all my commitments suddenly vanished once more, I’d be pretty disappointed”.
Klein on her seminar: a student says young workers in her home country cannot resist AI agents because “work hours are so grueling and pay so inadequate… everyone wants their own slaves”. An agent is “a duplicate version of yourself who you put to work performing tasks to earn you money”. “This phrasing stops our conversation cold.”
Klein’s thesis: “AI is surging through our culture in part because so many people are desperately trying to get out of their lives and with very good reason… The bots are just the closers in a bad deal stretching back decades”.
Klein quotes Roy: “on social media, we all talk like chatbots… the bots just arrived to finish the job”.
Klein: “How do we make life more livable… so we are a little less tempted to subcontract our lives to bots”.
Klein, quoting End Times Fascism: “Apocalypse culture and societal crises are age-old twins. Many of the people caught up in efforts to speed up the apocalypse are facing urgent earthly… challenges”. Her closing line: “to beat the tech oligarchs, we will all have to become more human, more loving, more animate, more present in our bodies”.
Klein on teaching: with 12 to 15 students and a living wage, “the idea that a chatbot could do her job… is absurd”; with 40 students and GoFundMe supplies, “a chatbot tutor… sounds like maybe that’s a good idea”. “It did not start with AI. The body was already weak”.

📚 Deeper theory

A demand-side theory. Most AI criticism is supply-side (what firms do). Klein offers a demand-side one (why users want it). The nearest theories are Marx on alienation (1844), Han on the burnout self (2010), and Fisher on capitalist realism (Roots). The claim has a clean test: where life is harder, adoption should be faster, and for pressure rather than pleasure. Some survey evidence supports “adoption driven by workload and employer pressure”, but I did not find a study that tests Klein’s claim directly (not confirmed).
The “slaves” thread. The student’s phrase and Aristotle’s Politics both link machine labour to the wish for servants (Roots). A modern point: an “agent” earning income for you is a rentier fantasy; whether that is realistic or just a story people tell themselves is unclear.
Chatbot-ification of speech. We all talk like chatbots has a long ancestor in Adorno and Horkheimer (standardised culture) and in Goffman’s idea that social media is performance. It is also a testable claim: studies of vocabulary convergence exist, but I did not check them.
Arendt again. Klein asks whether AI replaces the “dialogue between me and myself”. The link between Arendt’s “thinking” and outsourcing drafts to a model is Klein’s extension, not Arendt’s.
The structural point in the teacher example. It is a conditional claim: a technology’s appeal depends on the baseline it is compared with. This is the “shock” logic in miniature: damage a public service, then offer the private fix (Klein, 2007). Roy says it directly.
Tolstoy-like “Why do we do what we do?” question. Klein’s 2020 story is a personal ethics claim: design your life so that you would miss your commitments. It echoes Aristotle’s eudaimonia and, in modern form, Self-Determination Theory (Deci and Ryan): people thrive when work feels autonomous and meaningful.

📚 Roots: Work, burnout and care — why people reach for the bot

The problem. Klein’s most original move is to ask why AI is wanted. Her answer is exhaustion. This is a long tradition.

The point. Klein’s central claim (“The bots are just the closers in a bad deal stretching back decades”) is the only part of the talk that offers a causal theory of AI adoption. The rest is critique of the industry. idea 2 covers it in depth.

The chain

Aristotle, Politics I.4 (1253b). He wrote that if tools could do their work on command, “master-craftsmen would have no need of servants, nor masters of slaves”. He says this to show why slavery exists, since tools cannot. A student’s phrase, “everyone wants their own slaves”, has the same logic. Karel Čapek’s play R.U.R. (1920) coined “robot” from the Czech robota (forced labour). The dream of mechanical servants is an old dream of getting out of work.
Marx, Economic and Philosophic Manuscripts (1844): alienation. Work becomes something the worker does for another’s purposes and not for their own life. People then feel themselves “at home” only when not working. Klein’s relief in March 2020 describes this.
Oscar Wilde, “The Soul of Man under Socialism” (1891); John Maynard Keynes, “Economic Possibilities for our Grandchildren” (1930). Both expected machines to free people to live. Wilde argued (in paraphrase) that dull, unpleasant labour should be done by machines, but only under the right ownership. Keynes expected a 15-hour week. Neither happened. That gap is Klein’s subject.
Byung-Chul Han, The Burnout Society (2010). The “achievement subject” exploits itself. Mark Fisher, Capitalist Realism (2009): the sense that no alternative is imaginable. David Graeber, Bullshit Jobs (2018). Kathi Weeks, The Problem with Work (2011): the demand to “work less” as a political aim.
Nancy Fraser, “Contradictions of Capital and Care” (2016); Silvia Federici. Capitalism depends on unpaid care and social reproduction while eroding the conditions that support it. Klein’s “make life more livable… built to meet the needs and desires and deep pleasures of our living selves” is a care-economy argument.
Roy’s counterpoint. She does not engage the burnout argument directly. She answers with the effects: “there’s a lot of talk about what AI can do for human beings, but not much about what it does to human beings”. Klein and Roy therefore disagree on the order of the argument. Klein says: understand the cause of use. Roy says: look at the result of use.

Where others push back.

Revealed preference. People adopt tools because they work for them. Surveys suggest rapid uptake across income groups and in countries with less grinding work. Klein’s causal story may explain some adoption and miss most of it.
The “why not both?” point. Better schools and AI tutors are compatible. Evidence is mixed: some trials of AI tutors show learning gains (e.g., the World Bank pilot in Nigeria reported in 2025; from memory, check), and others show cheating or shallow use.
“Bad deal” is not unique to AI. The same account could be given of the smartphone, television or the car. It then explains any technology and so none in particular. Klein’s answer is that AI is the first technology marketed as a replacement for the person.
Agency. The student’s story is of workers choosing an agent to earn more. Calling that “closers in a bad deal” can read as denying that people make rational choices under constraint.

What they leave out. Who gets the income from “your” agent: the user or the platform? And what happens when everyone has one (competition erases the gain). Also: older users, disability, and caring work, where assistive tools matter most.

Check yourself

State Klein’s demand-side thesis in one sentence. What would count as evidence against it?
How does the teacher example differ from a simple “AI is bad” claim?
Why does Klein say the book passage “applies to end times religion as much as it does to end times tech”?

3. One Cyclops: coordination, “the prize”, and the bubble (Klein, Roy, Hao)

In short. Klein says the AI companies are not distinct. Their differences are tuning differences, and the guts are essentially the same. They loathe each other but they coordinate, and she points to circular financing. They race to be the one company left standing, and she says they found around eight trillion dollars for the buildout without a market for their products.

🎙 In the talk.

Hao (question): these “are not distinct entities”; Roy’s image is a single-headed, multi-bodied Cyclops.
Klein: OpenAI and Anthropic are “caught in like a Coke and Pepsi battle”; Musk and Grok are “the monster energy drink”; Meta is “the pumpkin spice latte”. “These are tuning differences… The guts are essentially the same”.
Klein: “They loathe each other… but… they do coordinate.” Evidence: the joint signing with Trump; “Nvidia is doing circular financing with all of them”.
Klein: they believe “eventually there will just be one company standing… in a race to swallow the entire economy”. She cites SpaceX’s market of “$28 trillion” and Anthropic’s “$30 trillion”. Her guess: “probably some sort of merger of like Nvidia, Palantir, Anthropic, and SpaceX, judging from who hates each other least” (a guess; she says so).
Klein on AGI: they define it as the point at which… their technology performs the vast majority of human activities… that literally means we’re out of luck… I don’t know whether they can do it or not, but that is what they’re racing towards.
Klein: “they are incredibly overleveraged but they’ve somehow managed to find somewhere around eight trillion dollars to finance the AI buildout without a market for their products”. What happens “when that bubble bursts”?
Roy: they are trying “to fuse themselves with the state and eventually become the state”. “We were colonized by the East India Company and then that was taken over by the British crown… maybe Columbus was right, this is East India”.
Roy on the founders’ warnings: they “informed us in somber … tones” that “there was a fair chance the machine god… could bring about the end of human civilization” and made “a plea for government regulation… a sort of tie me up, tie me down vibe. Is it that or is it just the inherent nature of capitalism? Helplessness in the face of untethered greed”.

📚 Deeper theory

Competition vs coordination. Economics has two views. Oligopoly theory says a few firms can compete hard on features and coordinate on the rules (lobbying, standards, capacity). Competition theory says rivalry itself forces innovation. Klein’s claim is the first. Evidence of coordination she gives (the Accord signing; chip-financing loops) is real but does not show coordination on goals. “Hating each other” is also consistent with a winner-take-most race, where no one can stop.
Circular financing. In 2025–26, chip and cloud firms invested in AI labs which then bought chips and cloud from them. Analysts are split between “normal vendor financing” (a long-standing practice) and “round-tripping that inflates demand”. I did not check the specific deals Klein means.
TAM as story. A total addressable market is investor storytelling: “if we capture 1% of $30 trillion…“. Its bigness is a signal that the story is about replacing labour at scale. Anthropic reportedly derives its figure from the range of human work AI could do (The Next Web). Klein’s reading (“so they believe there’s really only one player that can win”) is an interpretation: a TAM can equally be read as marketing.
Bubble mechanics. Carlota Perez (2002) describes installation-phase bubbles that overbuild infrastructure; Hyman Minsky (1986) describes how stability breeds leverage. Klein’s “when the inevitable pop happens” assumes a pop. The Bain estimate of under $2 trillion in AI revenue by 2031, against much larger planned data-centre spending (via ThePrint), is the best-known quantitative form of the worry. Defenders say demand is real and revenues are growing fast. Anthropic’s revenue is reported to have more than doubled to $11.6 billion in Q2 2026 (The Next Web, secondary).
Firm-state fusion. Historians of the East India Company stress that it was a state in a company’s clothes, with an army and tax rights, and that this ended through a crisis (the 1857 rebellion) and nationalisation (Dalrymple 2019, Roots). Modern parallels people cite: defence contracts (Palantir), cloud contracts with governments, and tech executives in advisory roles. The strongest version needs evidence of capture. The Accord is a candidate: a voluntary set of commitments, written with the firms, with no enforcement (Computing).
The “regulatory capture” reading of doom talk. Roy’s reading of the founders’ warnings follows a theory of regulatory capture: ask for rules you can shape, which also raise barriers for smaller rivals. Critics of AI-risk talk (Gebru, Whittaker, Bender) make this argument. The reverse reading: some researchers do fear the risk (see idea 7).

📚 Roots: Empire, colonialism and dispossession — “this is East India”

The problem. Why do the same patterns (extraction in the South, benefits in the North, a language of progress) keep reappearing?

The point. Roy’s central claim: “the way it’s been forced on us is very, very old”. It is this tradition. The evidence that fits best is material (data centres, labour, land, water), not the cultural one.

The chain

Rosa Luxemburg, The Accumulation of Capital (1913). Capitalism needs a non-capitalist “outside” to feed on (land, labour, resources) and keeps seizing new ones.
Frantz Fanon, The Wretched of the Earth (1961); Edward Said, Orientalism (1978). Colonialism worked through knowledge and language as well as force: the colonised are described, classified and ruled through the coloniser’s categories. This links Roy’s language theme to her empire theme. A system that decides what counts as “language” and “truth” has a very old precedent.
David Harvey, The New Imperialism (2003): “accumulation by dispossession”. Privatisation, financialisation and enclosure of commons are modern versions of seizure. Roy’s line (“first you undermine public infrastructure and then say privatization is the answer”) is a plain statement of Harvey’s idea.
William Dalrymple, The Anarchy (2019). A history of the East India Company: a private company with its own army, which took over a subcontinent, and which Parliament later had to nationalise after a crisis (1857–58). Roy’s joke “Maybe Columbus was right, this is East India” draws on exactly this: a firm that becomes the state. The history is solid; the analogy to AI is Roy’s argument, not a fact.
Nick Couldry and Ulises Mejias, The Costs of Connection (2019); Michael Kwet on “digital colonialism” (2019). Data extraction as a new colonial relationship: the “Global South” supplies data, labour and sites, and the “North” owns the models. Hao’s book gives fieldwork: Kenyan data labellers, Chilean water activists.
Cedric Robinson, Black Marxism (1983): racial capitalism. Capitalism has always run on racial hierarchy. In Klein’s examples (Gaza, Minneapolis) the tools are tested first on groups already treated as outside full citizenship.
Surveillance tech as an export. Reporting by +972 Magazine in April 2024 described Israeli military systems “Lavender” (marks people as suspected militants) and “The Gospel” (suggests buildings), with reportedly brief human review. The Israeli military disputes the “automated kill list” description and says such systems assist analysts (+972). Klein’s “AI-aided genocide” is a political-legal characterisation that Israel rejects; I did not verify its current legal status. The factual question (how much AI is used and how) and the legal question (is it genocide) are separate.

Where others push back.

They are not all one thing. Anthropic, OpenAI, Google, Meta, xAI and Chinese labs differ on safety policy, openness (Meta, DeepSeek, Mistral release weights) and business models. Open models weaken the “one Cyclops” picture: anyone can run them. Klein’s “tuning differences” claim is hard to square with the open-weight ecosystem.
Coordination is thin. The Accord is two pages and voluntary. Some firms have lobbied against each other’s preferred rules. A “joint signing” can be a photo opportunity.
A merger of Nvidia, Palantir, Anthropic and SpaceX is speculation (Klein says so). Antitrust law, shareholder structures and personal rivalries are large obstacles.
The bubble could pop and AI still matter. The dot-com crash did not end the internet. If so, “window when it pops” is real, but so is a durable industry after it.
China. The “one project” frame is a US-centred one. If a rival bloc builds AI, Coxon’s call for cooperation with China shows that the geopolitics is not simple (Crypto Briefing summary, secondary).

What they leave out. Public funding and state AI programmes (India, the EU, China); the role of open-source communities; workers inside the firms (Coxon is one); and a numerical test of “overleveraged”: who holds the debt, and how much is cash-funded by profitable parent firms (Alphabet, Microsoft, Meta).

Check yourself

What evidence does Klein give for coordination, and what would the opposing evidence be?
Why is a “total addressable market” not a forecast? What does Klein read into it?
Roy’s “fuse with the state” claim: what historical case does she use, and what would count as a current example?

4. Surplus humanity, the state and the front line (Klein, Roy)

In short. Klein asks how the system will treat surplus humanity once the bots are trained. Her answer is that people’s function is to be surveilled, contained and scanned. She points to Gaza and to France as glimpses of the future. Roy says the same is happening in India, with workers wearing cameras and millions removed from voter lists.

🎙 In the talk.

Klein: after “we finish training the bots with our labor and our creativity… how do you treat surplus humanity?” The answer: “your function… isn’t to consume and it isn’t to work. It is to be surveilled. It is to be… contained, to be scanned”. (Klein says “as Arundhati just said” before the list “killed… contained… scanned”, so she credits Roy for the idea.)
Klein: Israel has been running an experiment since they sealed Gaza… testing these technologies… We already have an AI-aided genocide in Palestine. Also “Gaza, Minneapolis… look at what is happening now in France”: students asked “to cool our classrooms instead of spending more… on the military and… the French state crushed them”. “That is a glimpse of the future.”
Roy: “This is the same thing that’s happening in India now”. In India “a lot of labor is now required to wear a camera on their foreheads that is documenting their movements so that they will be replaced by robots”. “Millions of people, starting from the state of Assam but now in the whole of India, who’ve been removed from the voter lists… are not… given rights or benefits. Some of them are building their own detention centers”.
Roy: You can’t really separate things… the battle against data centers… the battle about climate… the battle of the arts… all of us are on the front line.
Hao: the current fight “is part of a very long continuum of frontline fights, environmental, labor, anti-colonial” and the opponent is “also more unified”.

📚 Deeper theory

Test beds. Experiments on marginal groups that later spread are a known pattern: colonial policing methods returned to the metropole (Foucault’s “boomerang”, from Césaire; Stoler on imperial “laboratories”). Klein’s claim for Gaza, and her list of cities, follow this logic. A careful form: technologies of control are tried where rights are weakest (Mbembe, Roots).
What the reporting says on Gaza. +972’s Lavender and Gospel reporting (April 2024) describes systems that flag people and buildings and, according to its sources, gave brief human review; the Israeli military disputes the “kill list” description (+972). Separate reporting on Microsoft’s cloud use is cited in my search but I did not check details. “Genocide” is a legal term under contest; Israel rejects it. I did not check the current state of international legal findings. This page does not rule on it.
India’s data-systems politics. Aadhaar (the 12-digit ID), the National Register of Citizens in Assam (2019 final list excluded about 1.9 million people; from memory), and voter-roll “special revisions” form the setting for Roy’s claim. Assam’s revision removed about 10.56 lakh names in the Dec 2025 draft, and the chief minister said some deletions followed BJP complaints (The Statesman). The nationwide “Special Intensive Revision” (SIR) removed about 6.5 crore names from draft rolls in 12 states and union territories in phase 2, per a Deccan Herald headline I could not check (Deccan Herald; the URL’s “65 crore” looks like a typo for 6.5 crore). The Election Commission’s stated purpose is to remove dead, duplicate and migrated voters. Whether AI is involved in these deletions is a separate question the talk does not answer.
“Camera on the forehead”. Egocentric video capture to train humanoid robots is a real industry practice (from memory). I did not verify its use on Indian labourers as Roy describes it (not confirmed).
Arendt’s “right to have rights”. The power to remove a person from a register is the power to remove their standing (Roots). That is why Roy’s point on voter lists is more than an administrative complaint.
The “France” example. Klein says the French state “crushed” student demands about classroom heat and military spending. I did not find the specific event and cannot assess the claim (not confirmed).
The unified-front argument. Roy and Hao’s claim that all fights are one fight is the “intersectional” or “movement-of-movements” strategy (Seattle 1999; the World Social Forum; Klein’s own No Logo). Its risk: a coalition so wide that it cannot agree on a demand.

📚 Roots: Surplus people — from “reserve army” to “permanent underclass”

The problem. What happens to people the economy does not need? Capitalist societies have always had unemployed people. The question is whether they are temporary, useful (as a threat that keeps wages down), or disposable.

The point. This tradition is why “surplus humanity” is not a throwaway phrase. It is a concept with a lineage that runs through Arendt’s camps. Klein uses it with full knowledge of that weight.

The chain

Marx’s “reserve army of labour” (Capital vol. 1, ch. 25). The unemployed are a pool that holds wages down. They are useful to capital.
Hannah Arendt, The Origins of Totalitarianism (1951). She writes of people made “superfluous”: not exploited, but unneeded. She sees this as the precondition of the camps. Rightless people, stateless people, “surplus” populations. This is the deepest root of Klein’s phrase “surplus humanity”. Roy’s point on people removed from voter lists fits Arendt’s other idea: the “right to have rights”. If the state removes your name from the roll, you lose the right to claim any others. (Arendt’s phrase; the match to Roy is mine.)
Zygmunt Bauman, Wasted Lives: Modernity and Its Outcasts (2004). Modern order produces “human waste” and needs places to put it.
Mike Davis, Planet of Slums (2006). He uses the term “surplus humanity” for the billion-plus urban poor who are not needed by the global economy. Klein’s phrase echoes it. The new claim in 2026 is that middle-class workers will join them.
Achille Mbembe, “Necropolitics” (2003). The power to decide who may live and who may die; surplus people are managed through exposure to death. Klein’s list “to be surveilled… to be killed… to be contained” is close to Mbembe’s list of techniques.
Tania Murray Li, “To Make Live or Let Die?” (2010). Studies surplus populations in Indonesia: land is taken, but jobs do not follow. People are left to survive without being absorbed into the new economy. This is closer to Roy’s India than any European source.
The 2026 “permanent underclass” meme. A Silicon Valley belief that AI will freeze class positions within a few years (Big Technology). Its critics (summarised there): Alex Kantrowitz thinks it is “impossibly pessimistic about human nature”; Scott Alexander noted that it is spread more by well-off tech workers afraid of slipping than by the poor.
Labour economists, the sceptics. Aaron Benanav, Automation and the Future of Work (2020) argues that “surplus” is produced by slow growth, not by robots. If he is right, AI’s role in making people redundant may be more modest than the “closers” story suggests, while the political use of the story is real. Daron Acemoglu and Simon Johnson, Power and Progress (2023): technology’s effect on workers depends on which direction it is developed; automation-only paths hurt workers, but other paths exist.

📚 Roots: Fascism and technology — three different questions

The problem. The talk uses “fascist” for AI in three ways: a technology that favours fascist control (Klein’s seminar); an alliance of tech and far-right power (End Times Fascism); and Roy’s Indian setting, where Hindu nationalism and data systems have already merged. These need different evidence.

The point. Klein’s seminar question sits at 2.4(6) (Roots, idea 4). The end-times claim sits at 2.4(4) and (7) (Roots, idea 4). Roy’s India sits at 2.4(8) (Roots, idea 4). The audience hears one word. The scholarship has three different tests.

The chain

Walter Benjamin, “The Work of Art in the Age of Mechanical Reproduction” (1935). “Fascism… the aestheticizing of politics”: spectacle replaces participation. Roy’s “taxidermist’s Shangri-La” is an image of an image-world with no life behind it.
Hannah Arendt, Eichmann in Jerusalem (1963) and “Thinking and Moral Considerations” (1971). Eichmann’s crime was a failure to think: clichés, officialese, no inner dialogue. Arendt later describes thinking as a silent dialogue “between me and myself” (the “two-in-one”; see The Life of the Mind, 1978), and says that people who lack it can do evil without noticing. Klein invokes this. The step from there to “AI replaces that dialogue” is Klein’s, not Arendt’s. Arendt never wrote about AI. Treat it as an application.
Umberto Eco, “Ur-Fascism” (1995). A list of 14 features (cult of tradition, rejection of critical thinking, fear of difference, “newspeak”, and so on). It is a checklist. Pro-AI voices ask whether chatbots are “poor vocabulary and elementary syntax” (Eco’s phrase on newspeak) or the opposite.
Robert Paxton, The Anatomy of Fascism (2004). Fascism is best understood as a process with stages, and a political alliance between a mass movement and traditional elites. This is the frame that makes sense of “tech-state-far-right alliance”, and also of why scholars disagree about whether the label fits the present day. Paxton himself wrote in 2021 that he now thought the label fit Trump-era politics (from memory; check).
Jason Stanley, How Fascism Works (2018). Looks at tactics: mythic past, propaganda, anti-intellectualism, hierarchy, victimhood, law and order, sexual anxiety, appeals to the heartland, and “Sodom and Gomorrah” (decay). Used by many US critics.
Lewis Mumford, “Authoritarian and Democratic Technics” (1964); Langdon Winner, “Do Artifacts Have Politics?” (1980). Some technologies need centralised control to work (Mumford’s example: large systems), while others can be run in a distributed way. Winner’s example is Robert Moses’s low bridges. This is the proper academic home of the question is AI a fascist technology?. The strongest version asks whether frontier AI, which needs huge capital and compute, is structurally authoritarian (concentrated control) or merely currently so.
Quinn Slobodian, Crack-Up Capitalism (2023), Peter Thiel’s writings, and the “Dark Enlightenment” of Curtis Yarvin. The Silicon Valley right argues for exit, enclaves and rule by CEO-kings. This is the intellectual ground under End Times Fascism. (I name these as background on who the book is likely to mean; not checked against the book.)
Roy on India. Her own work, Azadi (2020) and essays on the Aadhaar ID, the National Register of Citizens in Assam, and Kashmir, treats data systems as tools of exclusion. Her claim about removal from voter lists belongs to this line. Fact context: the Election Commission says its roll revisions remove dead, duplicate and moved voters; critics say the process targets minorities. In Assam, a roll revision removed about 10.56 lakh names in the Dec 2025 draft, and the chief minister said some deletions followed complaints from BJP workers (The Statesman, Scroll; reports differ on totals).

Where others push back.

Causation. Many of these harms (voter-roll politics, military targeting, protest policing) predate AI and run on paper, phones and databases. Calling them “AI” may give AI credit it does not deserve (the critihype point, idea 5). A narrower claim, “AI speeds and scales existing control”, is more defensible.
Assam and SIR are contested in both directions. Defenders say that clean rolls are a basic democratic duty and that large deletions reflect real duplicates and deaths. Critics say the process puts the burden on citizens and targets minorities. The talk gives only the second.
Military AI. Defenders argue that data-driven targeting can reduce civilian harm if well governed. The reporting disputes whether this happened in Gaza. The general question (can AI targeting be more discriminate?) is open.
“Surplus humanity” is a forecast, not a fact. No large labour-market collapse has yet been shown from AI. Studies of employment effects are early and mixed (Acemoglu, Benanav; Roots).
Rhetoric and evidence. Moving from “we already have X” to “this is the same thing” asks the listener to accept several links at once. A reader who rejects one link (for example the legal term in Gaza) may dismiss the lot.

What they leave out. A distinction between AI as a cause and AI as an instrument; Indian state capacity for good uses (welfare delivery, language tools); workers’ own views on data-labelling jobs; and any concrete policy for “surplus” people (for instance basic income) other than “democratic socialism”.

Check yourself

What does “surplus humanity” mean in Klein’s mouth? Which earlier thinkers used the same words?
Which of Roy’s examples did I verify, and which did I not?
State the strongest form of the “AI is only an instrument” objection. How could Roy answer it?

5. Guilt, benefits and movement-building (Hao’s two audience questions)

In short. Hao puts two audience questions: guilt at taking part in the system, and the real benefits people get from AI. Klein says she has never seen a movement rise so quickly and so organically. People use AI joylessly, in resignation or overwhelm, so the job is to make life worth living. She warns against purity tests. Roy says there is no anti-AI movement in India.

🎙 In the talk.

Hao poses two questions she says audiences often ask, in their voice. First: “I feel so much guilt suddenly becoming a part of the system that I actually don’t want to partake in. Does that mean I’m a bad person?“. Second: I’ve gotten so much benefit from AI. It’s really helped me with my job interviews… helped me reduce taxes… save a lot of money on home improvement.
Klein (first): “I have never seen a movement rise up this quickly this organically”; it is “barely foundation funded” because “most of the big foundations actually don’t want people taking on AI”. AI Now organised “the first panel of organizers in local data center fights who were meeting each other for the first time”.
Klein: people use it “joylessly… in a spirit of resignation, in a spirit of overwhelm, in a spirit of their bosses forcing them… our job is not just to fight AI, it’s to make life worth living for more people”.
Klein on purity: “one mistake we made in the climate movement was making people feel that they had to be very pure”; “build a really welcoming movement that… holds on to a deep critique… but just not have too many purity tests”.
Roy (second): “truthfully there’s no anti-AI movement in India. India’s gearing up to be one of the greatest suppliers of data of all time”. Describing the attitude of Indian leaders (not her own): “Please bring your toxic waste to us. We’ll cherish it. That’s the attitude over there”; if data centres cannot be built elsewhere, oh please bring them here.
Roy: the pattern is old: “the science of AI is new, but the way it’s been forced on us is very, very old”. I was thinking about Oxycontin… it does cure pain, but what happens next?… these industries are very good at boosting the help and underreporting the consequences.
Roy reads the tsunami passage: the sea draws back, people “marvel at the miles of pristine new beach”, while “a wall of water was building up”; the “magical treasures, jeweled seashells” are “consumer applications”. She lists what the apps do: “They can be your mother, lover, therapist, doctor… What they can do in the field of medicine is nothing short of a miracle. But it won’t stop the poor from continuing to die of cholera, malaria, tuberculosis, or simply hunger”.
Roy: the cost: “a cognitive decline that’s happening and we are eventually going to be reduced to like people with atrophied brains and long tentacles… prompticles”. “You can’t produce morphine and say have this, you’ll feel better… the benefits are nothing compared to the costs that generations are going to pay”.
Roy also names allies who use AI: “some of our best, most beloved anti-fascist YouTubers… devoted proponents of AI”; the Gen Z “cockroach” protests rolled in on “the most delightful irreverent AI news”; but “when the threat assessment spells real danger, that window can be slammed shut in an instant”. The internet in her neighbourhood “has been shut down” during “season two”.

📚 Deeper theory

Purity politics. The “carbon footprint” debate is a known case: in the 2000s fossil-fuel firms promoted personal footprint calculators (BP’s 2004 campaign is the usual example; from memory), which moved attention from industry to individuals. Klein’s warning fits that history: shame on users helps the industry. The same logic runs the other way: individual choices matter when they build norms (boycotts, consumer movements).
How movements grow. Coalitions typically need (a) a shared grievance, (b) a low bar to enter, (c) organisers and (d) a target. Klein’s “welcoming” movement is a choice of (b) low bar. Chenoweth’s data (Roots) is often cited for breadth. The AI Now convenings she cites are evidence for (c).
The pharmakon. The old Greek word for both poison and cure (Plato’s Phaedrus; Derrida; Stiegler). Roy’s Oxycontin and morphine images are pharmakon arguments: a real benefit that comes with a dependency. Her tsunami image makes the same point as timing: you cannot see the cost yet.
Evidence on the “benefits” side. Medical AI shows real gains in narrow tasks (imaging, protein structure prediction). Evidence for consumer uses (legal, tax, home repair) is anecdotal. Roy’s examples (“improve your Nazi salute”, “write novels in a minute”) are a deliberate mix of real and absurd uses.
The cognitive-decline evidence. The MIT Media Lab study [Kosmyna et al., 2025] found lower EEG connectivity and weaker recall in an LLM-assisted essay group. It had 54 participants, was a preprint, and a published comment questions its EEG methods and reporting (arXiv 2601.00856). The authors have also asked journalists not to read it as “AI makes you dumb”. I found no replication. Related, more cautious literature: “cognitive offloading” (Sparrow et al., Science, 2011, on Google and memory; from memory).
India’s AI politics. India has tried to become an AI and data-centre hub (a global AI summit was held in New Delhi in Feb 2026; from memory). Roy’s claim that there is no organised opposition is a claim about civil society, and one I could not test. Her claim sits beside real protests on other issues (the Gen Z “cockroach” movement; NPR).
Internet shutdowns. India has often led the world in shutdowns, per Access Now’s yearly counts (from memory). Roy uses this to show that digital tools of protest are conditional.

📚 Roots: Movements and strategy — and the strongest counter-case

The point. Klein’s Q&A answers draw on the movements Roots in idea 5. Her bubble remarks draw on the political-economy Roots in idea 6 and are challenged by Perez. Roy’s harm claims draw on the Roots in ideas 1, 3, 4 and 6 and are challenged by the benefits case. The “counter-case” box is the yardstick for the pushback sections below.

The problem it addresses (movements). How do broad, slow, divided publics win against concentrated power? And what is the counter-case from people who think AI is mostly good or mostly ordinary?

Movements

The Luddites (1811–16). Textile workers who broke machines not out of ignorance but to bargain over how machines were used. Eric Hobsbawm, “The Machine Breakers” (1952), called it “collective bargaining by riot”. Brian Merchant, Blood in the Machine (2023), argues for a fresh reading. In 2026 some anti-AI groups use the name (the “Luddite Lab” at Circuit Breakers; found in search). Neither speaker uses the word.
The Narmada Bachao Andolan and Roy’s “The Greater Common Good” (1999). Roy’s own history: a decades-long movement against large dams. She brings its lessons: the benefits are advertised; the costs fall on those who cannot speak. The Oxycontin parallel she gives belongs here (see Patrick Radden Keefe, Empire of Pain, 2021).
Climate movement and “purity”. Klein’s warning (“you had to purge your whole life of fossil fuels”) concerns a real, well-known argument in climate politics about individual vs system change.
Erica Chenoweth and Maria Stephan, Why Civil Resistance Works (2011). Nonviolent mass movements have succeeded more often than violent ones; a “3.5% rule” of the population is often quoted (contested and based on few cases). A broad, welcoming movement is a strategic choice in this literature, not just a moral one.
Polanyi’s double movement (again). Societies protect themselves. If the AI boom hurts enough people, a protective movement is predicted. Which form it takes (left, right, fascist) is open. This makes Klein’s urgency intelligible.
Green New Deal and the “Leap Manifesto” (2015). Klein’s own policy lineage. Her talk’s final request, “democratic socialism”, is a clear political commitment and not a neutral policy suggestion.

The counter-case

Carlota Perez, Technological Revolutions and Financial Capital (2002). Each technological revolution has a financial bubble phase (excess), a crash, then a “golden age” where the infrastructure is used well. If right, the bubble Klein predicts is normal, and the lasting infrastructure is the point. (She also warns that the golden age needs political action.)
Narayanan and Kapoor, “AI as Normal Technology” (2025). AI will diffuse slowly; capability is not the same as adoption; the real risks are ordinary (misuse, concentration, labour). See Princeton CITP. A Prospect piece cited in search notes that hospitals using AI heavily have hired more radiologists than when Geoffrey Hinton predicted their replacement (Prospect). This undercuts both the “AGI soon” and the “surplus humanity soon” stories.
Dario Amodei, “Machines of Loving Grace” (2024); Sam Altman, “The Intelligence Age” (2024). The pro-AI case: faster science, cures, abundance. Marc Andreessen, “The Techno-Optimist Manifesto” (2023). Slowing AI is itself a moral harm because of what it delays. (I cite these from memory.)
Erik Brynjolfsson, “The Turing Trap” (2022). A mid-position: the risk is building AI to replace humans rather than to augment them; policy and design choices decide which. Close to Acemoglu.
The sceptic-of-the-sceptics. Some argue that critics like Roy and Klein overstate AI’s power, which is a kind of hype. This is the “critihype” argument (Lee Vinsel, 2021): critics who repeat the industry’s claims about capability end up amplifying them. Klein half-concedes this: “I don’t know whether they can do it or not”.

Where others push back.

The benefits are not trivial. For people with little money, a free tutor, lawyer-like explainer or health advisor can be a large gain. Roy’s reply, “this is not the solution for the lack that exists in society”, is correct as a point about policy and weaker as a point about the person who needs help tonight. The morphine analogy assumes dependency and harm; for many uses, evidence for either is thin.
Analogies are not evidence. Oxycontin and the tsunami show a form of argument. They do not show that AI has the same cost structure. A reader should ask for the measure, not the metaphor.
Cognitive decline is a hypothesis. People also use calculators, maps and search engines. Critics ask whether AI differs in kind. The best tests are still being run.
Purity vs critique. Some activists say a movement that tolerates use dilutes its message. Others say that hypocrisy charges are used to silence critics (the climate case). The tension is real; Klein picks a side.
“Toxic waste to us”. Indian policy-makers argue that data centres bring jobs, investment and sovereignty over data. Roy’s frame (extraction) and theirs (development) disagree on the same facts: who benefits, who pays for power and water.

What they leave out. Concrete tests of harm (what outcome, measured how); disaggregation by use (medical vs chat vs code vs companionship); positive Indian examples (language tools, welfare delivery); and any account of how to use AI in a way they would endorse. Klein hints (“targeted pro-human uses of AI”,) and then asks whether these are possible without climate collapse. The question is left open.

Check yourself

Why does Klein think “purity tests” hurt a movement? What past case does she use?
What kind of argument is Roy making with Oxycontin and the tsunami? What would turn it into an empirical argument?
Name two responses a defender of AI could make to “morphine”, one a policy answer and one an evidence answer.

6. The crash, the window and the alternative (Klein; Roy on privatization)

In short. Klein expects the bubble to pop, which will give a narrow window of more leverage than we will ever have again. She asks what to regulate, what to nationalize or internationalize, and what to abolish. She says our job is to keep the ideas ready for when the politically impossible becomes politically inevitable. Roy describes how privatization works: first undermine public infrastructure, then sell civilization back to us by the meter.

🎙 In the talk.

Klein (reading): “the bubble currently holding up the global economy audibly hisses air. When the inevitable pop happens, we collectively are going to have a narrow window of more leverage than we will ever have again”.
Klein: her own questions: “What parts of this nascent industry do we want to heavily regulate? What parts do we want to nationalize or internationalize? And what parts should be abolished altogether?“. And whether “targeted pro-human uses of AI” are possible “without tipping our already overheating planet into full-blown collapse”.
Klein on the Green New Deal: “we heard again and again that there was no money for this”, but “they’ve somehow managed to find somewhere around eight trillion dollars” for AI.
Roy on how privatization works: “first you undermine public infrastructure and then say oh privatization is the answer… So you want to privatize the history of human endeavor. You want to privatize civilization itself… and sell it back to us by the meter”.
Klein, closing: Friedman said “Only a crisis, real or perceived, produces real change… Our job is to keep the ideas ready for when the politically impossible becomes politically inevitable”. “There’s a big crisis coming because these folks have gambled with the entire economy… We’re going to have another window and we need democratic socialism”.

📚 Deeper theory

The source. The Friedman line is real (1982 preface, Roots). Klein’s gloss is that he saw the New Deal as the error and wanted to reverse it. A libertarian critique says Klein pulls the line out of context and over-reads a modest claim about being prepared (Libertarianism.org). I could not verify the Pinochet letter detail she mentions (not confirmed).
Why a pop would open a window. After 2008, banks were rescued and the system was patched; no deep reform followed (Mirowski, Roots). Klein’s view: the left was not ready. A better-prepared left would have specific plans (public ownership, caps, compute rules). Her three options (regulate / nationalise-internationalise / abolish) are a list of such plans, with no ranking.
What “nationalise or internationalise” could mean. Public compute (a national or intergovernmental facility), public ownership of frontier labs, or a treaty regime. Real analogues: the nationalisation of the Bank of England (1946), CERN, the International Atomic Energy Agency (1957). Whether these fit AI is debated. Coxon’s call for global cooperation on limits [reported, secondary] points to a similar structure from a different starting view (Crypto Briefing).
Public alternatives. Public-interest AI (language models for Indian languages; EU or Swiss open models; “AI commons” proposals) is a live policy area. Some of this exists inside India’s public programmes (from memory). It sits between Klein’s “abolish” and the industry’s “scale”.
“Privatise and sell it back by the meter.” A play on Altman’s own line about selling intelligence “on a meter” (reported at a March 2026 BlackRock event; bioethics.com summary of Gizmodo). Roy’s analysis follows Harvey’s “accumulation by dispossession” (idea 3): train on commons (the web, books, art), then rent access back.
The money comparison. Klein sets “no money for a Green New Deal” against large AI finance. Careful version: private capital flows into AI are not the same as public budgets, so the comparison is rhetorical. But it fits her political point: money exists when elites want it.
Democratic socialism. Her word. It names an aim (democratic control of key economic sectors), and it is a political commitment, not a policy. A hearer who disagrees with it can still agree with the diagnosis.

📚 Roots: Political economy of technology — who owns the machine, and what is it for?

The problem. Is a technology good or bad in itself? Or does the answer depend on who owns it and what it is built to do? This tradition says the second. It treats AI as a capital good and a political project, not only as software.

The point. Klein’s reading of IPO documents, her point on circular financing, and Roy’s line “privatize the history of human endeavor… sell it back to us by the meter” are all political-economy moves.

The chain of thinkers

Karl Marx, Capital vol. 1 (1867), ch. 15 “Machinery and Large-Scale Industry”. Machines do not free workers by themselves. Under capital, machines are used to cheapen labour, speed it up, and reduce workers’ skill and bargaining power. In the Grundrisse (1857–58, the “Fragment on Machines”) he adds the general intellect: a society’s shared knowledge becomes embodied in machines owned by capital. This is the oldest version of Klein’s point that AI is trained on “our labor and our creativity” and “they’ve stolen all of our art and stolen all of our books”.
Karl Polanyi, The Great Transformation (1944). Markets spread by turning things that are not made for sale into commodities. Labour, land and money are his three “fictitious commodities”. When this goes too far, society pushes back (the double movement): protection, regulation, unions, and sometimes fascism. Hold this thought for the Roots in ideas 6 and 2. It gives a clean reading of Roy’s point that art is “not commoditizable”: AI is a new attempt to turn culture and language into a metered product.
Shoshana Zuboff, The Age of Surveillance Capitalism (2019). Platforms extract behavioural data and sell predictions. This is “the labor and creativity” Klein mentions, plus the surveillance function she gives to “surplus humanity”.
Nick Srnicek, Platform Capitalism (2016). Firms become platforms that own the infrastructure everyone else uses. The AI “stack” (chips, cloud, models, apps) is a stronger version. Yanis Varoufakis, Technofeudalism (2023) goes further and argues that capital has been replaced by cloud rent. (Contested; many economists reject the “feudal” label.)
Kate Crawford, Atlas of AI (2021). AI is a physical industry: minerals, energy, water, underpaid labellers, data taken without consent. Crawford co-founded AI Now (2017, with Meredith Whittaker), the institute that hosts this evening (per Wikipedia and general knowledge; from memory).
Emily Bender, Timnit Gebru and others, “On the Dangers of Stochastic Parrots” (2021). Large language models predict text. They do not understand it. The harms are environmental cost, encoded bias, and the illusion of meaning. Gebru and Émile Torres later named the TESCREAL bundle (transhumanism, extropianism, singularitarianism, cosmism, rationalism, effective altruism, longtermism) as the ideology of the AI-risk-and-abundance wing (First Monday, 2024; from memory). The German summary of Roy’s essay names Gebru and Meredith Whittaker as critics who were ignored (zeli.app); I could not check this in the essay.
Karen Hao, Empire of AI (2025). The direct source of the evening’s frame. Hao argues that OpenAI moved from a nonprofit mission to something like an empire that takes land, labour and resources, and justifies this with a story about benefiting humanity (as described in reviews; I read summaries, not the book).

The thread. Marx: machines serve their owners. Polanyi: markets that eat what should not be sold provoke a backlash. Zuboff and Srnicek: data and platforms are the new means. Crawford, Bender, Gebru, Hao: AI is a material, social and ideological project. Klein’s “Cyclops” claim is a claim in this tradition: that the firms are one class interest, with competition among them as a surface.

📚 Roots: Crisis and shock — “keep the ideas ready”

The problem. How do big policy shifts actually happen? Rarely by gradual persuasion. Usually in a crisis, when old arrangements fail and people reach for whatever is on the table.

The point. The “window” appears twice, (“a narrow window of more leverage than we will ever have again”) and. If you remember one structural point about Klein’s talk, it is this: she is making a prediction about timing (a crash is coming), and a recommendation about preparation (have the policy ready).

The chain

Antonio Gramsci, Prison Notebooks (1929–35). “The old world is dying, the new world struggles to be born; now is the time of monsters” is the commonly quoted line (a loose translation of a passage on the interregnum). The idea: in a crisis of authority, the outcome depends on who has organised and prepared. Klein’s whole closing argument is a Gramscian one.
Karl Polanyi (1944). Crises of market society produce both protection movements and authoritarian solutions. The 1930s are his case.
Milton Friedman, preface to Capitalism and Freedom (1982). “Only a crisis—actual or perceived—produces real change. When that crisis occurs, the actions that are taken depend on the ideas that are lying around.” He says the job is to “develop alternatives to existing policies, to keep them alive and available until the politically impossible becomes politically inevitable.” Klein quotes this. (The wording above is the well-known version; I confirmed the passage exists and sits in the 1982 preface through secondary pages.)
Naomi Klein, The Shock Doctrine (2007). She uses Friedman’s line as her key. Chile after the 1973 coup, Iraq, post-Katrina New Orleans, and others: market reforms rushed through while people were disoriented. The critics. Libertarian writers say Klein misreads Friedman and overstates the pattern (Libertarianism.org, Cato). Historians of Chile also debate how far the “Chicago Boys” followed a plan from Friedman.
Philip Mirowski, Never Let a Serious Crisis Go to Waste (2013). After 2008, neoliberalism survived the crash. Why? Not because its ideas were true but because its networks were organised. This is the dark mirror of Klein’s hope: the window can close on the left.
Klein’s own pivot. In The Shock Doctrine the villains had the ideas ready. In this talk, Klein says the opposition must have them ready. Her earlier example is the Green New Deal: “we heard again and again there was no money”. Her 2015 Leap Manifesto is the same move: put the plan on paper before the moment.

Where others push back.

A crash may help incumbents. Past crashes concentrated assets in the hands of those with cash (the 2008 pattern). Strong firms buy weak ones. If so, the window favours the Cyclops, not its critics.
Crashes are not announced. A “bubble” is a judgement. If the boom continues, the “window” never opens. The argument should not depend on timing.
Nationalisation has costs. Public ownership of fast-moving technology has a mixed record: state firms can be slow and politicised. India’s own history (the licence raj; also ISRO’s success) shows both.
Abolition. Banning parts of AI would meet a global competition problem: if one bloc stops, others do not. Enforcement across borders is hard.
The Friedman-as-plan reading. Historians doubt that a single idea-circulation plan explains the policy shifts of the 1970s–80s. Many factors (stagflation, politics, the Cold War) worked together.

What they leave out. What exactly “democratic socialism” would own, regulate or share. Who would run a public AI sector and with what governance. And the international dimension: how a window in one country (the US) helps people in India or Africa.

Check yourself

What does Klein mean by “keep the ideas ready”? How does she reverse the way Friedman used it?
Name two reasons a crash might not produce reform.
Which of her three options (regulate, nationalise/internationalise, abolish) is easiest to defend, and why?

7. Doom talk, the “tie me up” reading, and joy (Roy; the close)

In short. Roy notes that CEOs speak of a fair chance of the end of civilization, then ask for regulation. She calls that a cry for help with something a little kinky about it: “tie me up, tie me down”. Asked for a joyful nugget, she answers “Look at my shoes” and gives a manifesto. It runs from loving and being loved to watching, trying to understand and never looking away.

🎙 In the talk.

Roy: “Perhaps we should listen to the people who are building this glorious new technology. Last month, a young employee at Anthropic, Jacob Coxon, quit his job, saying that his conscience would not allow him to work on a project that could potentially annihilate the human race”. She calls the picture “the Manhattan Project on steroids”, but now “the scientists are motivated in large part by profit and a race to get there first”.
Roy: the CEOs “fell over each other in their rush to discuss, disagree, or concur with the whistleblower” and “informed us in somber… tones” of “a fair chance” of the end of civilization. Then “they reassured us that they were doing their best to mitigate the risk. They made a plea for government regulation. Their cry for help hinted at some sort of pathology. There was something a little kinky about it… tie me up, tie me down”.
Roy: “they just casually tell us… there’s a 10% chance that you’ll be annihilated”.
Hao asks for “one little joyful nugget”. Roy answers: “Look at my shoes”. Then her manifesto: “To love, to be loved, to never forget your own insignificance… To seek joy in the saddest places. To pursue beauty to its lair… To respect strength never power. Above all to watch. To try and understand. To never look away. And never never to forget”.
Hao: “we are going to spend some of this panel talking about the grotesque, but I want to begin and end with the beautiful”. At the end Hao and the room chant When people rise, empires fall.

📚 Deeper theory

Two readings of doom talk. (a) Sincere risk: some researchers believe powerful AI could cause catastrophe. Hubinger’s reported “>10%” figure and Coxon’s resignation are of this kind (secondary reports: AI Weekly). Also Geoffrey Hinton, Yoshua Bengio and Stuart Russell (from memory). (b) Strategic ideology: doom talk makes the technology sound all-powerful, steers the debate to far-future risk, and invites rules that large firms help write (Gebru, Whittaker, Bender; idea 6). Roy’s “tie me up” joke reads the behaviour of the CEOs through (b); her use of Coxon treats his act through (a). The two are not in conflict: one person can be sincere while the firm uses the frame.
The “Manhattan Project” frame. It borrows the moral drama of Oppenheimer. Historians note the analogy is imperfect: the Manhattan Project was a state project with a clear end; AI is a diffuse commercial race.
Pathology vs structure. Roy’s alternative explanation is capitalism: “helplessness in the face of untethered greed”. This matches the structural account in game theory (a “race to the bottom” or Moloch): each firm would prefer a slowdown if the others did too, and none can slow alone. Coxon’s plea for coordinated slowdown is, in this reading, a request to solve a coordination problem. Interestingly, it is close to Roy’s diagnosis, even if the remedy differs.
Joy as politics. Roy’s manifesto belongs to a tradition of joyful resistance: Audre Lorde (“caring for myself is… an act of political warfare”, 1988), Emma Goldman’s (apocryphal) “If I can’t dance…”, bell hooks on love. Klein’s closing (welcoming, not pure) pairs with it. “To respect strength never power” is a distinction between capacity and domination (Arendt’s The Human Condition draws a similar line between power and violence).
“Never look away.” This is the witness tradition (Primo Levi; Susan Sontag, Regarding the Pain of Others, 2003). Its worry: attention without action can turn into spectacle. Roy’s list ends with “never forget”, which is a memory duty.
The manifesto’s source. Roy says it is from “one of my first political essays”. I did not verify which essay.

Where others push back.

If the risk is real, “pathology” is the wrong frame. People who think extinction risk is serious argue that reading it as marketing leads to under-response. They ask critics to say what evidence would convince them.
If the risk is hype, quoting it spreads it. The “critihype” point (idea 5). Roy and Klein both say they cannot tell whether the systems will work as claimed (Klein,), yet both repeat the warnings. This is an internal tension in the talk.
“Low IQ” and “lizard blood.” These are insults, not arguments. Some listeners will take them as proof that the critique is about contempt. Others will take them as a writer’s licence.
Joy does not equal strategy. Hao asks for “one little joyful nugget”; the answers are a joke, a manifesto and a chant. They fit the evening but do not tell a listener what to do on Monday.

What they leave out. A shared test for sorting sincere from strategic warnings; any account of what safe enough would be; and what a person within an AI firm should do (Coxon’s case). Also, the possibility that both “AI is overhyped” and “AI is dangerous” are partly true.

Check yourself

Give the two readings of doom talk. Which does Roy apply to the CEOs, and which to Coxon?
What is the internal tension created by Klein’s “I don’t know whether they can do it or not” and the use of the same firms’ warnings?
Why might Roy’s “strength never power” matter for how a movement acts?

Q&A

(Add questions and answers here only if worth keeping.)

About the source and how this page was made

Speaker: Arundhati Roy, Naomi Klein (moderator Karen Hao)

Event: AI Now event, final panel of the evening (moderator says AI Now; not independently confirmed)

Date: ~late Oct 2026 (inferred: Roy’s essay is dated 16 Oct 2026 by one aggregator; Klein-Taylor book out 15 Sep 2026; Roy says Coxon quit “last month” = Sep 2026). Exact date not confirmed.

Length: 50:52

A note on balance. The two main speakers are strong critics of the AI industry. They speak to a friendly audience at an event run by an advocacy research institute. So the talk is a position, not a debate. Each deep dive has a “Where others push back” section. It gives the best counter-case, so you can test the position. The pushback is not the page’s own view either.

Speakers

Label Who Role
Hao Karen Hao Moderator. Journalist, author of Empire of AI (Penguin Press, 2025). Introduces herself as “a very proud board member of AI Now”. (Her board seat is not confirmed by my search.)
Roy Arundhati Roy Indian novelist and essayist. The God of Small Things (1997), The Ministry of Utmost Happiness (2017), memoir Mother Mary Comes to Me (2025). Long-time writer on dams, mining, Kashmir, Hindu nationalism, corporate power. Reads from a new essay on AI.
Klein Naomi Klein Canadian author and activist. No Logo, The Shock Doctrine (2007), This Changes Everything, Doppelganger (2023), and, with Astra Taylor, End Times Fascism (out 15 Sep 2026, per Democracy Now).
Audience Not heard except applause and cheers A room of organizers and supporters. One earlier speaker is called “Kashan”; not identified.

Identity check. The panel is introduced by Hao with all three names. A web search confirmed the books and roles above, and that Hao and Klein have appeared together before (UBC, 12 Mar 2026, per UBC listing). I could not find a public listing for this AI Now event, its date, or Roy’s presence on it (not confirmed). The identification rests on the introduction in the transcript.

Caption warnings. These are auto-captions of live speech. Corrections used on this page:

“Karen How” = Karen Hao. “Naomi Klene” = Naomi Klein. “Arunati Royve / Arendati / Arjetti / Aronati” = Arundhati Roy. “Ashra Taylor” = Astra Taylor. “Mine and Astra” = Naomi and Astra.

“Jacob Coxen / Coxson” = Jacob Coxon (see slip box below).

“Grock” = Grok (xAI). “Palunteer” = Palantir. “Nvidia” is right.

“Pinocha” = Pinochet. “Milton Freriedman” = Milton Friedman. “Hannah Arent” = Hannah Arendt.

“John Burgger” = John Berger (British writer, d. 2017).

“Chad GBT” = ChatGPT. “inab zindabad” = Inquilab Zindabad (“long live revolution”).

Some words in Roy’s opening are garbled: “Audies” (unclear; do not quote), “mant” (probably “manic” or similar), “Sephiler” (probably “sepulchral”). “Gigification” = gig-ification. “Mdash” = em dash.

“cyclopes” = Cyclops (Hao’s word for Roy’s image of a single-headed, many-bodied monster).

“Kinky… tie me up, tie me down” is Roy’s joke, not a caption error.

The captions merge speakers inside long turns. Where one block holds two voices, this page splits them by content (for example, where Klein stops and Roy begins at “Yeah. This is the same thing that’s happening in India now”).

Factual slips and claims to treat with care (checked or flagged):

Coxon. Roy says a young employee “Jacob Coxon” quit Anthropic “last month” because he feared the technology “could potentially annihilate the human race”. Reports confirm that Jacob Coxon, a British pretraining researcher, resigned from Anthropic on 8–9 Sep 2026 and warned the labs were “gambling with our lives” (WCAX, AI Weekly). Reports say he asked for a slowdown and cooperation with China. “Annihilate” is Roy’s wording of his warning, not a quote of his words. Secondary reports only.

“3,000 languages”. Roy says India has 3,000 languages. This is a rough figure. The 2011 census recorded about 19,500 “mother tongues” grouped into 121 languages; 22 are scheduled. (†) It is a speech figure, not a claim she builds on.

SpaceX “$28 trillion”. Klein says Musk said SpaceX has a future market of “$28 trillion”. Reports of the S-1 filing give $28.5 trillion, most of it labelled AI-related (Yahoo Finance). Close enough. This is a total addressable market claim, not a forecast of revenue.

Anthropic “$30 trillion”. Klein says Anthropic tells investors it has a $30 trillion future market. Press reports (citing the Wall Street Journal’s anonymous sources, Aug 2026) say Anthropic planned to tell IPO investors its total addressable market exceeds $30 trillion. Anthropic has not published it (not confirmed). See The Next Web. TAM is a ceiling, not a forecast. Reports note this.

Trump and the “constitution”. Klein says the AI companies signed a document that Trump said was “maybe more important than the US Constitution”. News reports of the 29–30 Sep 2026 signing say Trump called it “almost like a Constitution” and “morally binding”; it is voluntary and has no enforcement (Axios, Computing). I did not find the “more important than” wording (not confirmed).

“$8 trillion”. Klein says the industry found “around eight trillion dollars” to finance the build-out. I did not find this number. Other estimates I found are lower: about $725 billion of 2026 capex by the five biggest hyperscalers, and about $5 trillion over five years for data-centre and AI infrastructure worldwide (JP Morgan base case, per a search summary). Treat $8T as unverified; it may count commitments or announcements, not spending.

Friedman and Pinochet. Klein says Friedman wrote to Pinochet that “everything went wrong” because of the New Deal. Friedman did write a letter to Pinochet in 1975 (from memory). I could not verify the New Deal passage (not confirmed). The “only a crisis” quote is real and comes from the 1982 preface to Capitalism and Freedom (see Klein Doctrine critique, which disputes how Klein uses it).

“Studies show cognitive decline”. Roy says studies show a direct cognitive decline. The best-known study (MIT Media Lab, 2025) had 54 people, was a preprint, and has a published methods critique (arXiv comment). See the deep dive.

“Cameras on foreheads” and voters building their own detention centres. I did not verify these claims (not confirmed). The voter-roll deletions themselves are real and disputed (see idea 4).

The essay. The reading is called “Apocalypse Now and Then”. A web aggregator dated 16 Oct 2026 carries an Arundhati Roy essay under the headline “Artificial intelligence is capitalism’s final destination” (zeli.app summary). It may be the same essay under another headline (not confirmed). I could not read the original or confirm where it was published.

A quote from Roy that people will want to cite. “They can improve your Nazi salute and arm you with a mental chainsaw” is sharp. Read it in its sentence: she is listing what the apps promise.

🎙 marks what the speaker says. 📚 marks added background and the wider debate, not the speaker’s claim. † marks background from general knowledge that was not checked against a source.

Talks

AI, Empire and the Living World — Arundhati Roy and Naomi Klein, with Karen Hao

Summary

Map — talking points

Putting it together

Part 1 — Before you read

Key terms

People named (a reading list)

The speakers’ own frame

The current debate (2025–26)

The ideas

1. Language, writing and what is lost (Roy’s essay; the writers’ exchange)

📚 Roots: Language, art and mystery — what is lost when words become supply?

2. Why people use AI: “the bots are the closers in a bad deal” (Klein)

📚 Roots: Work, burnout and care — why people reach for the bot

3. One Cyclops: coordination, “the prize”, and the bubble (Klein, Roy, Hao)

📚 Roots: Empire, colonialism and dispossession — “this is East India”

4. Surplus humanity, the state and the front line (Klein, Roy)

📚 Roots: Surplus people — from “reserve army” to “permanent underclass”

📚 Roots: Fascism and technology — three different questions

5. Guilt, benefits and movement-building (Hao’s two audience questions)

📚 Roots: Movements and strategy — and the strongest counter-case

6. The crash, the window and the alternative (Klein; Roy on privatization)

📚 Roots: Political economy of technology — who owns the machine, and what is it for?

📚 Roots: Crisis and shock — “keep the ideas ready”

7. Doom talk, the “tie me up” reading, and joy (Roy; the close)

Q&A

Table of Contents

Label	Who	Role
Hao	Karen Hao	Moderator. Journalist, author of Empire of AI (Penguin Press, 2025). Introduces herself as “a very proud board member of AI Now”. (Her board seat is not confirmed by my search.)
Roy	Arundhati Roy	Indian novelist and essayist. The God of Small Things (1997), The Ministry of Utmost Happiness (2017), memoir Mother Mary Comes to Me (2025). Long-time writer on dams, mining, Kashmir, Hindu nationalism, corporate power. Reads from a new essay on AI.
Klein	Naomi Klein	Canadian author and activist. No Logo, The Shock Doctrine (2007), This Changes Everything, Doppelganger (2023), and, with Astra Taylor, End Times Fascism (out 15 Sep 2026, per Democracy Now).
Audience	Not heard except applause and cheers	A room of organizers and supporters. One earlier speaker is called “Kashan”; not identified.