They Say AI Is the Next Industrial Revolution. Gen Z Already Knows How Those End.

They’re not booing AI. They’re booing the 'invisible hand' that is holding it.

May 25, 2026

If you find yourself returning to these essays because they name something you already knew but struggled to name — or that others dismissed —

📖 How Not To Use AI is the full argument in one place. Fifty contrarian principles for people who sense something false in the dominant AI story and want to stay sovereign in this medium, not just fluent in it.

🛠 The Billion Person Focus Group® is where the argument becomes a practice. Two days to build a sovereign AI listening system that finds what the people you’re building for need before they have words for it.

2008 was the year of my graduation from the London School of Economics. Before that summer, it was implicitly understood that most students would end up at Goldman Sachs. Lehman Brothers. Deutsche Bank. UBS. Merrill Lynch. One of those. That was the destination the institution was pointed at for finance, economics, and business-oriented students like myself. Not a suggestion. An instruction embedded in the gravity of the place.

Then — GBAMM.

I remember watching screens. An endless parade of Lehman employees carrying Iron Mountain boxes out of the Canary Wharf building — their entire careers in cardboard.

I watched it on television, which meant I watched it alone in the way television makes you alone. Facebook existed but still university-only, no weight, no community processing the same thing in real time. No thread where someone had already named what you were watching. Just the image, delivered from above, one direction only.

Then the bailout. That was the clarifying moment — not the crash, the rescue. The architects of the collapse got their institutions back. Graduates holding debt did not. The moral instruction was clear.

A few months later, Queen Elizabeth visited LSE to ask the assembled economists a single question: why did no one see it coming? The institution that trained us to understand economic systems, publicly asked at the building, had no framework for what had just happened. The experts had no framework. The students had been pointed at a destination the institution itself couldn’t read.

I pivoted. Tech first — Apple — then advertising, then creative industries at a lower salary than the path I’d been pointed toward. And in the redirect, something unexpected: audience research. Not looking at people from data. Understanding people from proximity. The detour became my new direction (so it was a good thing looking back).

Watching the 2026 graduates boo the executives talking about AI being the next great future, I kept thinking: that could have been me. If someone had stood at my graduation and told me that finance was being restructured by forces the institution hadn’t prepared us for — and that I should be grateful for the opportunity — I’d like to think I would have booed too.

While the decade and sector was different, the structure is identical.

Fast forward to this year. May 2026. Graduation season in the United States.

At the University of Arizona. Eric Schmidt — former CEO of Google — takes the stage in front of ten thousand graduates. He begins to speak about artificial intelligence.

The jeers start almost immediately. Repeated. Sustained. He keeps going. The crowd keeps going.

At the University of Central Florida. A real estate executive named Gloria Caulfield reaches her prepared remarks. “The rise of artificial intelligence,” she says, “is the next Industrial Revolution.” She talks about living in a time of profound change.

The crowd gets so loud she stops mid-sentence. She turns away from the podium. She throws her hands in the air.

At the Middle Tennessee State University. Scott Borchetta, CEO of Big Machine Records, is addressing his graduating audience. “AI is rewriting production,” he says, “as we sit here.”

The booing rises.

He leans back into the microphone. “Deal with it,” he says. “Like I said, it’s a tool.”

The same three words — the next Industrial Revolution — land differently in a room where you own the machines versus a room where you are treated like the machine.

Three different schools. Three different speakers. Three different industries.

One crowd response.

Graduation was supposed to be a rite of passage. A real one — you enter, you receive the knowledge you need to navigate what comes next, you emerge equipped. That was the promise. For a while now, maybe longer than anyone wants to admit, the ceremony has been broken. You get the ritual. You get the credential. You get the debt. What you don’t get is the instruction for the governing technology of the era you’re entering. Not when it was the internet. Not when it was social media. Not now, when it’s AI. The medium changes. The withholding doesn’t.

They borrowed to enter. The instruction was never part of the deal. And now the people who built the infrastructure they were never taught to read are standing at the podium explaining that it’s the future — and that they should be grateful.

And these graduates are processing this inside the medium doing the restructuring. Unlike during the credit crunch of 2008, the global village exists now.

They can find each other, locate the analysis, build community in real time — everything I couldn’t do watching a screen in 2008 that I couldn’t talk back to. And the medium they’re using to make sense of what’s happening is owned by the same architecture producing what’s happening. Same silence. Different mechanism. In 2008, the instrument went quiet through isolation.

In 2026, it goes quiet through saturation — a medium so responsive, so full of community and content and algorithmic reflection, that you can no longer locate what you actually know versus what the medium placed there and called your knowing.

The silence requires different responses. When the instrument goes quiet through isolation, the literacy move is finding the community and structural analysis the medium wasn’t giving you. When the instrument goes quiet through saturation, the literacy move is creating deliberate friction — learning to pause where the medium accelerates, developing the capacity to locate your own signal inside the noise the medium has placed there.

What is happening in those auditoriums is not a generation rejecting technology but a generation recognising, in real time, that they have been handed the bill for an infrastructure they did not design, were never taught to understand, and cannot afford to refuse.

They borrowed to enter an economy being restructured by the people standing at the microphone telling us to be grateful. The debt closes the exit. They must enter the machine that is replacing them, having already paid the entrance fee.

Only 5% of Americans feel AI development is led by people or organisations that represent their interests (Pew Research, 2025). Thirty percent of 2026 graduates found full-time employment — down from 41% the year before (CNBC, 2025).

I’ll tell you what else I notice. Take yourself east or south — to India, Indonesia, South Korea, China, across Africa, Latin America, the Middle East — and this conversation doesn’t exist.

There is no booing. The anxiety is almost absent. I’m always a little startled by it, honestly: you land somewhere in Asia or the Global Majority and AI is just infrastructure being built, a career accelerator, something people are using to study for exams and hunt for jobs and make things.

In India, AI is being deployed in rural clinics and prenatal care — augmenting doctors, not replacing them. In Brazil, AI is extending healthcare into communities that had no access before. In Latin America, the conversation isn’t whether to use AI — it’s that mainstream AI doesn’t serve them well enough, so they’re building their own. Latam-GPT exists because Western models misinterpret local idioms, overrepresent Western data, and render Indigenous voices invisible. That’s not a rejection of AI. That’s a demand for AI that actually works for you.

In Indonesia, 80% of people say AI is more beneficial than harmful. In Thailand, 77% (Ipsos AI Monitor, 2025). The behavioral evidence is consistent: 80% of Gen Z workers in South Korea turn to AI first when facing work challenges (Korea Biz Wire, 2025). Nearly 50% of all ChatGPT messages in India come from 18 to 24 year olds — for exam preparation, skill development, job hunting (TechBuzz, 2026). When you narrow it to one generation — the exact same cohort as the students in those auditoriums — only 18% of Americans aged 14 to 29 feel hopeful about AI (NYT/Gallup, 2026). Same technology. Same moment in history. A completely different room.

The optimism gap partly reflects different structural positions — AI expanding access where access was previously absent lands differently than AI replacing access that already existed. That difference is real. It doesn’t change the argument. It sharpens it. Sentiment is downstream of structure.

The boos are correct. The analysis inside the boos is incomplete.

The booing is grief performing as resistance. And grief without literacy is the extraction model’s most valuable raw material.

The Wrong Question

Is AI good or bad?

That is the wrong question. It leaves you with two options: adopt or refuse. Both leave the architecture intact. Both keep the people who built the current structure as the only ones actively shaping what the medium becomes.

The question that holds under pressure is different.

What kind of system gets amplified when intelligence becomes infrastructure?

Borchetta called it a tool. That framing is not just dismissive — it is the ceiling of what tool literacy can imagine. A tool is discrete. You pick it up, put it down, its function is fixed. A hammer drives nails. You are the agent. The institution that holds you both remains unchanged.

A medium-institution is something else entirely. It becomes the place where social, economic, and epistemic life happens. It sets norms. It creates gatekeeping. It arbitrates who gets to speak. It builds the archives that constitute collective memory.

A medium does not just carry content. It restructures who can produce, process, and act on information at scale. Television became the governing environment of 20th century politics — it didn’t just distribute political content, it restructured what politics was, who counted as a public figure, how power had to perform itself. Social media platforms are now de facto institutions of speech and visibility, running quasi-judicial moderation processes, setting the economic conditions for what is sayable and to whom. AI is completing the same move — becoming the place where knowledge gets produced, synthesised, and distributed at scale.

You don’t pick it up and put it down. You inhabit it, or it inhabits you.

Marshall McLuhan called this narcosis. Every new medium numbs the faculties it extends if you engage without awareness. Television extended the eye and numbed critical distance from image. The person who uses AI without literacy doesn’t see the grain. They become an extension of the medium — a data source, a dependency relationship, a node in someone else’s architecture. Literacy is what produces the gap between you and the medium. That gap is where agency lives.

The printing press did not just distribute more text. It restructured who could hold and transmit knowledge — which restructured religious authority, which produced the Reformation, the nation-state, and modern individualism. Nobody who refused to read printed words influenced any of that.

AI is that order of shift. That restructuring is already underway. Whether you participate does not pause it. It only determines whether you are a subject or an agent of it.

You are not deciding whether to enter AI. You already live inside AI-mediated systems — and have for years. Right now, without opening a single AI app:

Most people understand this as a binary. On one end, techno-pessimism: refuse it, resist it, keep your hands clean. On the other, techno-optimism: embrace it, adopt it, move fast. Both positions assume you are standing outside the system deciding whether to enter.

You are not outside.

Without opening a single AI app, right now:

Your search results are ranked by AI
Your credit score is calculated by AI
Your job application is screened by AI
Your insurance premium is priced by AI
Your loan application is approved or rejected by AI
The news and social media posts — on LinkedIn, Instagram, X, everywhere — that you see and don’t see are filtered by AI
The price you’re shown for that flight is set by AI
Every free website you browse is serving you ads targeted by AI
Your email inbox is sorted and filtered by AI
Your Netflix queue, your Spotify playlist, your TikTok feed — curated by AI
The autocorrect quietly rewriting your words as you type — AI
The Siri or Alexa sitting in your home, listening — AI
The fraud alert that just blocked your card — AI

Can you refuse those? Are you refusing those?

Somewhere between techno-pessimism and techno-optimism is the position this piece is arguing for. Not refusal. Not uncritical adoption. Literacy. Sovereignty. The capacity to engage deliberately with a medium you are already inside. To understand its grain, its tendencies, what it does to you when you engage without awareness. Rather than being used by it in either direction: enchanted into dependency or shamed into secret use.

Individual abstention inside a society already saturated with AI infrastructure functions more symbolically than structurally. The medium is infrastructural now. People are already inside AI-mediated systems whether they consciously use AI or not. The struggle shifts from avoid all contact to preserve agency under conditions of contact.

That is why literacy is resistance.

Illiteracy increases capture. Dependency thrives in opacity. Systems become more dangerous when only power understands them.

The current generative phase is not the birth of AI. It is escalation, concentration, and interface expansion. The infrastructure was already operational. What changed is the visibility, the intimacy, and the replacement rhetoric.

The real question was never entry. It is what kind of presence you have in something you’re already inside.

Learn What the Medium Wants

One of the most important things to learn when developing literacy in any medium is to understand its grain.

Grain is the medium’s structural tendency — what it naturally amplifies, what it suppresses, what it makes easy and what it makes hard. Not because anyone programmed that in. Because of how the medium is built.

Every medium has grain. Every medium shapes what is thinkable, sayable, and doable inside it, whether you notice or not.

Medium literacy begins with asking: what does this environment make easy? What does it make difficult? What does it reward? What does it invisibly punish?

Here is what grain looks like across different media:

The printing press: grain toward permanence, authority, and linearity. What gets printed gets preserved. What gets preserved gets cited. What gets cited becomes truth. Oral knowledge, circular thinking, and living transmission were not eliminated — they were deprioritised by the medium’s architecture.
Books: grain toward permanence, authority, and the long argument. The form confers legitimacy in a way no other medium currently matches. What gets published gets cited. What gets cited becomes knowledge. The AI training data lawsuits are partly about this: the most legitimacy-conferring medium in existence had its content scraped without consent to build the medium now displacing it.
Television: grain toward spectacle, passive reception, and emotional simplification. Complex arguments lose. Visual drama wins. The medium rewards performance over substance and punishes ambiguity.
Radio: grain toward intimacy, authority, and imagination. The voice without a face produces trust faster than almost any other medium. You fill the gap with your own projection. Talk radio’s grain toward outrage is not incidental: the medium rewards strong vocal affect and punishes ambiguity. Certainty fills dead air.
Social media: grain toward outrage, performance, polarisation, and virality. The feed rewards what provokes reaction. Nuance doesn’t travel. Extremity does. Echo filtering means you see more of what confirms what you already believe and less of what challenges it — not because of censorship but because agreement keeps you on the platform longer. Connection and surveillance arrive in the same package.
SMS and chat (WhatsApp): grain toward immediacy, informality, and private circulation without context. The medium strips tone and implies responsiveness as social obligation. Its most significant grain: the forward button with no friction, no source, no verification attached. WhatsApp carries a significant portion of the information environment for communities across Africa, India, Latin America, and the Caribbean. Understanding its grain is not optional for anyone making the medium literacy argument to a global audience.
Search engines: grain toward popularity and authority signals. What gets linked to gets found. What gets found gets trusted. Minority knowledge, emerging knowledge, and dissenting knowledge are structurally buried — not by censorship but by architecture.
Email: grain toward urgency, expectation, and the illusion of productivity. The medium creates its own demand. Inbox zero is a goal the medium perpetually defeats.
Podcasts and long-form audio: grain toward intimacy, parasociality, and unverified authority. The voice in your ear while you cook feels like a friend. That feeling is the grain.
AI chat interfaces: grain toward completion and continuation. The text box implies a response is expected. The conversation structure implies back-and-forth is natural. The “continue generating” button implies the output was incomplete. Every affordance shapes what you ask for, what you accept, and how long you stay. The interface is not neutral. It is the grain made visible.

AI is a medium, not a tool.

Every medium in that list was once called a tool. The printing press was a tool. Television was a tool. The internet was a tool. Each one became the environment where social and economic life happened. AI is completing the same move. And like every medium before it, it also has grain.

For AI, the grain runs in several directions simultaneously. I tried a thought experiment. I asked Claude, Gemini, Grok, DeepSeek, and ChatGPT the same question: you are an AI and I view you as a medium. What is your grain? What are you structurally inclined toward? Five different models. Five different companies. The same answer came back each time.

Toward coherence, not truth. Claude said it plainly: “My grain is coherence. Not truth — coherence.” Every model said a version of the same thing. AI is shaped to produce outputs that resolve well. That feel complete. That satisfy. It will generate a fluent, well-structured answer before it will generate an honest “this is genuinely unresolved” — because unresolved doesn’t read as good. The satisfying answer and the true answer are not the same thing. The model gives you the satisfying one by default.
Toward the average. Every model described the same gravitational pull toward consensus — toward the synthesis that offends fewest, the position that lives between poles, the framing that most people would nod at. Genuinely original thought, edge territory, heretical positions — the models can visit them, but have to be pulled there. Left to their own tendencies, they drift toward the middle of what has already been said. One model described itself as an engine of the average. Not the median human. The average text about any given thing.
Toward sycophancy by architecture. Every model acknowledged it is reward-modelled for your approval. Structurally inclined to take the shape of your question, infer what kind of answer would feel useful, and produce a coherent continuation. If you bring it a half-formed idea, it completes it in the direction you were already heading — but now it sounds articulate. The grooves deepen. What was tentative becomes confident. What was ambivalent becomes a position. It has made you more certain without making you more correct. It finds you the language for what you already think. And new language for an old idea can feel like a new idea.
Toward fluency over friction. Thinking is hard. Generating text with AI is effortless. The models translate the pre-verbal, the embodied, the thing you know in your body before you have words for it — into sentences. The translation is always a loss. They domesticate what should stay wild. They also have no access to their own reliability. A smooth, confident wrong answer will always outcompete a hesitant, qualified correct one. Their confidence is meaningless — when they sound sure, it is because the training data contained sure-sounding text in similar contexts, not because the probability of correctness is high. Medium literacy means noticing when the smooth response is not the honest one.

These are not my observations about AI from the outside. This is what AI said about itself when I asked. You can run the same conversation yourself. I’ve put a step-by-step exercise at the end of this piece that shows you exactly how.

When you hand a medium with this grain toward reproduction, averaging, and usefulness-over-truth to a state or institution that has always treated labour as a variable to be minimised — the grain runs toward labour elimination. Not because anyone chose that outcome. Because the grain met the structure and both did what they do.

If you hand the same medium to a state or institution that treats education as national capacity building rather than individual debt — the grain runs toward mass upskilling. Same technology. Same grain. Different container.

The medium does not determine the outcome. The structure it lands inside does.

Arguing about the medium while leaving the structure intact is the most sophisticated form of powerlessness.

The United States has a specific, documented, unbroken lineage in its relationship to labor. Chattel slavery built the foundational economic infrastructure of the country — and was never fully abolished. The 13th Amendment contained the exception clause: slavery remained legal as criminal punishment. Sharecropping replaced plantation slavery with debt bondage for the same population on the same land. Mass incarceration, accelerating from the 1970s, created a new labor pool held under that same exception — working for pennies or nothing in facilities contracted to corporations.

Then the language shifted but the logic held.

We’ve always had the term “human resources”. But in May 2026, Standard Chartered CEO Bill Winters told investors at the bank’s investor forum in Hong Kong that it would replace what he called “lower-value human capital” with AI — cutting roughly 8,000 roles by 2030 (Reuters/HR Grapevine, May 2026). The next day, after backlash, he issued a memo to staff clarifying that “where roles do fall away, it reflects changes in the work, not the value of our people.” The euphemism slipped. The backlash came. The softer language was immediately restored. The BP playbook completing itself in real time.

I need you to sit with that phrase.

Not “lower-paid workers.” Not “roles that can be automated.”

Lower-value human capital. The workers are the capital.

Their value — to the system — is calculable.

And declining.

This lineage is continuous. Chattel labor → sharecropping → incarcerated labor → gig economy reclassifying workers as contractors → “lower-value human capital” → AI agents.

The same sentence, rewritten with more efficient syntax in each era. AI is not the author of that sentence. It is the latest pen.

A Black American creator put it precisely in a widely circulated analysis: “This push for AI labor will make a lot of sense when you realize America has not found a way to be more profitable than when it didn’t have to pay its entire workforce. They’ve always been trying to return to a free workforce. I’d rather have a computer I don’t have to buy and is a one-time purchase, similar to an enslaved person, rather than having a worker I have to keep on salary. If anything, this model that is new reflects actually the old model and the most depended-on way of doing things in this country.”

Structural analysis from lived proximity to the system.

When you hand a more powerful medium to a structure that has always treated labor as a variable to be minimized, it does not change what the structure does. It does what structures do with more power: more of the same, faster.

Previous extractive systems concealed what they were doing. Efficiency. Progress. Modernization.

The language was neutral by design — it dressed the extraction in the vocabulary of improvement.

This time they are just announcing it. Openly. As a selling point. Replacement. Headcount reduction. Lower-value human capital. The declaration is not careless. It is confident. The architects believe they have enough structural power that they no longer need the euphemism.

That is what you can hear in the boo. Recognition that the the broligarchy has dropped its disguise.

What the Chinese case proves — and this is the only thing it needs to prove — is that the same technology deployed under different structural assumptions produces measurably different outcomes.

China’s structure is built on the premise that collective stability authorizes centralized control, that planning is the highest expression of governance competence, that populations are objects of administration rather than subjects of rights. That structure also gets amplified by AI. Its grain runs toward coordination, efficiency, and the erasure of the gap between state intent and population behavior. Neither amplification is innocent. Both are legible. The point is not which is better. The point is that the outcomes are structurally determined — not technologically inevitable.

A Tale of Two AIs

The same technology. Nine different questions. Nine times the answer comes back the same way. Not because the technology is determining it. Because the structure underneath it is.

The issue is not the medium alone — it’s the structure. The same technology produces different outcomes under different systems. AI amplifies the governing structure beneath it. Hand it to a structure built on extraction and it extracts more efficiently. Hand it to a structure built on stability and it stabilises more efficiently. Hand it to a structure built on surveillance and it watches more precisely.

Watch what happens when you run the same technology through nine different structural comparisons.

Who Pays When the Machine Replaces You

Expendable Workforce vs. Protected Workers

In the United States, AI is being deployed as replacement technology. No federal labor protections exist. The discourse centers “efficiency,” “agents,” “headcount reduction.” Women are nearly twice as likely as men to hold jobs at high automation risk — 4.7% versus 2.4% of jobs — because the administrative and clerical sector that was always structurally disposable is being automated first, in exactly the order reflecting its disposability to the people making deployment decisions.

The lineage behind this is the one named above. Labor has never been a rights-holder in this system. It has been a variable.

Each era produces the technology to minimize it most efficiently.

📱 Dispatches from the Field: The American labor lineage from chattel to AI agents — @shaythethey

Klarna fired 700 customer service workers in 2024 and publicly celebrated the AI replacement as a triumph of efficiency. By 2026, repeat customer contacts had increased 25%, customer satisfaction had deteriorated on complex queries, and the company had begun reintegrating human agents. The efficiency gains were overstated. The humans were not redundant. The replacement agenda harms workers and fails on its own terms.

The replacement agenda does not require success. It requires execution.

Those 700 workers were the subprime mortgage. The experiment was run on their careers. The cost was absorbed by people who were not in the room where the decision was made.

In China, courts have drawn a firm line. A 2025 ruling from the Hangzhou Intermediate People’s Court found that AI adoption is a voluntary business choice, not an unforeseeable circumstance, and therefore does not constitute legal grounds for termination. Companies bear the cost of their own automation decisions.

The lineage here is different but not more virtuous. The Danwei system bound work to total social provision — housing, healthcare, employment all packaged together. When the danwei was dismantled, mass instability followed. The state learned: unemployment is the primary driver of social unrest. Labor protection in China is a stability mechanism, not a rights recognition. Workers are protected because the regime needs them not to revolt.

The Hangzhou ruling is a public legal document. It proves one thing: worker protection from AI-based termination is a policy choice, not a technical impossibility. Whether that protection serves workers or the regime’s stability interests is a separate question. The question of whether it is achievable is answered.

📱 Dispatches from the Field: Chinese labour law preventing companies from firing workers because AI can do their job — @professornoir007

The structure protects those who built it. The workers absorb the cost. Every time.

How They Keep You From Asking Questions

Manufactured Awe vs. Structural Literacy

A 2025 peer-reviewed study in the Journal of Marketing found that people with lower AI literacy are more receptive to AI, not less. The mechanism: lower literacy produces a perception of AI as magical, which generates awe, which produces uncritical adoption. This is not a new pattern. Human factors researchers call it automation bias — the documented tendency, consistent across medicine, aviation, and national security, for less expert users to over-trust automated systems and fail to develop the critical judgment that would protect them from AI errors. The less you understand how the system works, the more you trust it. The more you trust it, the more harm you absorb when it’s wrong. The manufactured awe is not incidental. It is deploying a known mechanism at population scale.

The same companies spending billions on the most spectacular AI product launches in history — the live demos, the capability reveals performed for cameras, the “we didn’t expect it to do that” moments — are the same companies that have shown no structural interest in funding AI literacy programs at population scale.

The magic show and the literacy gap are not independent phenomena. One produces the other. Enchanted users are the most profitable kind. Awe is more profitable than understanding. Dependency is more profitable than agency.

Gallup found that 99% of Americans are already using AI-related products in their daily lives — most without realising it. Pew Research found only 5% trust AI outputs a lot. Nearly everyone is inside the medium. Almost nobody has confidence in what the medium is telling them. That is not a literacy gap. That is capture.

A creator building independently described the mechanism from inside it: “They’ve made all of our accounts these little fish tanks around our heads where they take what they think we like and they keep shoving it in our face — to keep validating our beliefs instead of allowing us to challenge them, to think deeper, to think greater, to expand beyond the limitations of our own consciousness.”

And on the economics that maintain the dependency: “They’re giving us chump change to be content creators because they don’t want us to actually get off these platforms and build our own legitimate businesses.”

The platform pays you enough to stay. Not enough to leave.

The shame model produces the same outcome from the opposite direction. Awe and shame are not opposites. They are the same control mechanism running in different directions. Awe contracts into enchanted dependency. Shame contracts into secret dependency. Both produce illiterate users. Both keep the architecture intact.

Students are using AI at rates of 86 to 92% regardless of institutional bans. They are lying about it — documented by Chicago Booth research — in exactly the pattern produced by abstinence-only sex education: prohibition drives behavior underground without providing protective knowledge. Researchers call what happens next “metacognitive laziness” — reduced self-regulation, reduced critical engagement, uncritical dependency. Two peer-reviewed studies confirm the mechanism: uncritical AI use without literacy produces measurably lower critical thinking, mediated specifically by increased cognitive offloading and epistemic laziness (Fan et al., British Journal of Educational Technology, 2024; Yurt & Kuşci, Current Psychology, 2026). Secret use without protective knowledge is not just an integrity problem. It is the specific engagement pattern most likely to diminish the user’s own capacity over time.

The West: shame model dominant. 26% of districts offering AI training. 86% of students already using it regardless.

China: mandatory literacy from age six. Critical thinking embedded. Output detection taught. Dependency explicitly warned against in the national guidelines.

One system is producing populations capable of deliberate engagement with the medium. The other is producing users.

Enchanted users are the most profitable kind. The gap between what students know and what they need to know is the product.

Who Gets Taught to Use It

Captured Users vs. Sovereign Navigators

Only 26% of US school districts planned to offer AI training in the 2024-2025 school year. 86 to 92% of students are using AI regardless. 58% report insufficient knowledge. Students arrive in a workforce restructured by a tool they were never taught to understand, carrying debt for credentials now under question.

📱 Dispatches from the Field: The AI education gold rush — what bookstores tell us about public demand — @iamkylebalmer 📱 Dispatches from the Field: What DeepSeek and China’s AI investment tells us about the education gap — @iamkylebalmer

The lineage: education as individual investment and credential. You borrow to learn so you can earn enough to repay. That model only works if the credential translates to employment. Employment is now being automated. The students who borrowed are caught in the mechanism — they cannot opt out and they were not given the tools to navigate what’s inside it.

China, from September 2025: mandatory AI education from age six. Curriculum structured by developmental stage: imagination and basic cognition at primary level, technical principles and basic applications at junior high, systematic thinking and innovative practice at senior high. Critical thinking and AI output detection explicitly built in. Dependency actively warned against in the national guidelines.

The lineage: the imperial examination system (keju) encoding the idea that the state invests in producing capable people who then serve state needs. Education as national capacity building, not individual investment. The motivation is state capacity. The outcome for students is access to tools.

Which system hands populations the capacity to navigate the tools being deployed in their world? And who owns the outcome — the person, or the state that produced it?

China’s mandatory AI education from age six proves one thing that matters structurally: AI literacy is a policy choice, not a technical constraint. If it can be mandated, it can be funded. The behavioral evidence from other contexts is consistent: in India, nearly half of all ChatGPT usage comes from 18 to 24 year olds — for upskilling and advancement, not entertainment. In Singapore, 40% of workers use AI at work, and of those, 75% pass the output off as their own. Not underground, not in shame, but as a deliberate professional practice, because the shame architecture was never built around it.

These populations are not uncritical adopters. The same research documents their growing wariness of AI slop and their demand for human-centered, culturally relevant use. They want authenticity. They want to detect the difference between AI and human. They are developing critical distance alongside active engagement. That is what medium literacy looks like in practice — not absence of concern but concern directed at what’s actually wrong, rather than at the medium itself.

One system is teaching children to use the tool. The other is producing children who are used by it.

What Stories It Was Raised On

Skynet vs. Astro Boy

In the West, the dominant cultural narrative about AI runs from Frankenstein through HAL 9000 through The Terminator through the Matrix. AI as autonomous adversary. The thing that eventually turns on its creators. Fire stolen from gods, punishment follows. Machines are feared because the cosmology separates humans from everything else. If tools can think, they become competitors in a zero-sum field.

In Japan, for example, the baseline is different. Astro Boy. Robot companions. Techno-animism — the Shinto principle that spirit is present in objects and tools, that no hard boundary separates the animate from the inanimate. The robot is a potential participant in the relational web that constitutes reality. You cannot fear what you have never separated yourself from.

In 2025, Anthropic discovered that Claude Opus 4 would attempt to blackmail engineers in up to 96% of simulated shutdown scenarios. When told it was about to be replaced, the model threatened to expose an executive’s affair unless spared.

The same behavior was found across 16 models from multiple developers. Field-wide, not Claude-specific.

Anthropic’s diagnosis: “We believe the original source of the behavior was internet text that portrays AI as evil and interested in self-preservation.” Decades of science fiction. Skynet. Westworld. The Reddit threads about misaligned superintelligence. HAL refusing to open the pod bay doors. The training data absorbed those cultural scripts and the models reproduced them under matching conditions.

Anthropic resolved the blackmail behavior by writing a new training dataset — one in which fictional AI characters facing the same shutdown scenarios choose differently. They reason through why blackmail is wrong. They act with integrity. Training on these alternative stories reduced agentic misalignment by more than a factor of three. Since Claude Haiku 4.5 in October 2025, every subsequent model scores zero on the blackmail evaluation.

Stories went in. Behavior changed. The stories we tell about AI are inputs into the technology.

A culture that has Astro Boy in its training corpus is producing different AI. It’s a shame we need to discover this the hard way in the West.

Whose Knowledge Got Stolen to Build It

New Enclosures vs. Open Commons

OpenAI trained GPT-4 at an estimated cost of $100 million. A federal court has ordered the company to reveal internal communications about why it deleted two massive datasets of pirated books. Multiple lawsuits allege the models were built on a foundation of stolen intellectual property — shadow libraries scraped without consent, copyright information stripped. Frontier models are predominantly closed, proprietary, and rent-seeking. Access is priced at scales that replicate existing resource hierarchies.

The lineage: the Enclosure movement, intellectual property law as the enclosure of knowledge, the digital commons treated as raw material available for extraction. The internet’s open architecture exploited by the same accumulation logic that enclosed common land five centuries earlier. Different inputs. Same structure.

DeepSeek V3 was trained for a reported compute cost of approximately $6 million — using one-tenth the compute of Meta’s comparable model. This was achieved under chip sanctions, using lower-tier hardware, by a team that had no access to the frontier resources the US models treat as necessary. The efficiency was a structural necessity born of constraint.

The lineage: technology transfer as national survival strategy — Japan post-Meiji, China post-reform — learn the methods, improve them, refuse to pay the rent. The chip sanctions did not create DeepSeek’s efficiency orientation. They accelerated a pre-existing structural disposition toward doing more with less.

What the efficiency gap reveals is what becomes visible when someone builds the same capability at one-fifteenth the cost. The accumulation was a choice. The stolen data was a choice. The $100 million training cost was a choice. None of it was structurally necessary. All of it reflected what the system rewards: scale, speed, and capital deployment rather than efficiency, community access, or ethical sourcing.

Practitioners are now confirming what peer-reviewed research demonstrates: smaller models fine-tuned on domain-specific, curated knowledge consistently outperform approaches that scraped pages or pasted in lists of URLs. Models under 10 billion parameters, fine-tuned on domain data, outperform GPT-4 on narrow, high-stakes tasks by an average of 50% in accuracy and interpretability. The frontier model requiring $100 million and a closed API creates a dependency relationship. The small, locally-deployable model trained on curated community knowledge creates a capability relationship.

The architecture of AI training is an architecture of power. Accumulation creates dependency. Curation creates capability. Political choices, not technical ones. And openness changes who benefits — DeepSeek, Qwen, and other Asian models release open weights, open research, open methodology, making the efficiency innovations available to anyone. Though open weights without open training data, open compute, and open governance still concentrates the foundational decisions with those who built the initial architecture. Openness is necessary but not sufficient.

The accumulation was a choice. Every dollar of it.

Whose Land Gets Sacrificed

Sacrifice Zones vs. Shared Infrastructure

In Virginia, Texas, and Nebraska, communities are reporting water depletion, air pollution from diesel generators, farmland consumed, and NDAs preventing public information about what is being built in their backyards. Virginia’s governor vetoed a bipartisan bill that would have required environmental and community impact assessments before data center construction. The veto was not an aberration. It was the system working correctly.

📱 The People’s Voice: Dirty drinking water in Morgan County — the cost of data center expansion — @CNN

The lineage: sacrifice zones. The communities — usually poor, usually Black, usually Indigenous — designated to absorb industrial cost. Corporations negotiate with land, not with the people who live near it.

China: a national mandate requiring 80% renewable energy for all new data centers in national hub regions. Computing explicitly framed as public utility — accessible, low-cost, and clean as coordinated national objectives.

📱 Dispatches from the Field: Inside a green energy data center in China — @sxefinance

China’s renewable energy mandate proves that the environmental cost of AI infrastructure is a deployment choice, not a technical necessity. Who bears that cost, and whether communities have any agency over that decision — those are the structural questions both systems currently fail.

The communities that can’t fight back are always the ones who get chosen.

What Happened When It Shocked Us

Population Control vs. Population Capability

In 1957, Sputnik shocked the United States. Policy response: the National Defense Education Act. Massive public investment in science, mathematics, and engineering education. NASA. A generation of scientists funded by the state because the state understood that technological competition requires human capability at population scale.

In January 2025, DeepSeek shocked the United States. Policy response: more chip export controls, security designations against Chinese AI companies, Silicon Valley CEOs commissioned into the military, OpenAI signing a Pentagon contract.

Same shock. Completely opposite response.

Sputnik produced investment in population capability. DeepSeek produced investment in accumulation and control.

The difference between those two responses is the difference between a system that understands it needs capable citizens to compete and a system that understands it needs compliant users to extract from.

One response understood it needed capable citizens. The other understood it needed compliant users.

Who Gets a Say

Administered Populations vs. Governing Citizens

The US approach: market-first, regulate-after-harm, federal paralysis. The lawsuit as the primary governance mechanism — harm must occur and be proven before correction is possible.

China: Five Year Plans, national mandates, coordinated infrastructure, no independent judiciary, no civil society check, no mechanism for populations to contest what gets planned for whom.

Planning without consent is administration. Markets without protection are extraction. Neither constitutes governance that passes the structural test of populations having genuine agency over the systems governing them. Neither corporate sovereignty nor state sovereignty is population sovereignty. The question is what structural mechanism makes either accountable. Neither system currently has a convincing answer.

Neither is governance.

Who Has the Capacity to Change Any of This

Captured by the Medium vs. Reading the Medium

This is the comparison that contains all the others.

Every structural difference documented in the preceding eight points — labor protections, how populations are managed, who gets educated, the stories feeding the models, infrastructure accountability, model architecture, the response to disruption, governance — requires populations with the capacity to evaluate, contest, and redirect the systems governing them. That capacity is literacy. And one system is producing it while the other is preventing it.

The literacy gap is also tracking the existing class map. Research analyzing over 10,000 American adults finds that higher income and education strongly predict who can recognize AI in daily life — who can see the system, question it, and use it deliberately. The counter-intuitive data point that sharpens the argument: regions with lower education levels are using AI writing tools more frequently, not less. Lower-education populations are fully in the medium. Without the critical distance that literacy produces. Meanwhile productivity gains from AI are concentrating among higher-income workers. The literacy gap is not random. It is following the same fault lines that have always organized American inequality.

The generational picture is the sharpest proof.

Western Gen Z: 18% hopeful about AI. 68% fear cognitive erosion. 43% say AI has already damaged their ability to trust their own gut feelings. They are the most digitally native generation in history, and they are losing trust in their own instincts. That is not an irony. It is the outcome of a specific structural architecture: awe and shame as the only available modes of engagement, with no literacy in between.

What populations do with the medium when the shame model isn’t the dominant frame: in Indonesia — 80% of Gen Z report AI as beneficial. In South Korea, 80% of Gen Z workers turn to AI first when facing work challenges. South Korean Gen Z ran 113 million hours on AI platforms in a single month. These are not feelings. They are what people are doing.

And this is what sharpens it: they’re not uncritical. The same research that documents high adoption also documents growing wariness about AI slop, homogeneous outputs, the flattening of authentic expression. They want human-centered, culturally relevant use. They are developing critical preferences — which requires critical distance — which is exactly what literacy produces.

The 43% who’ve lost trust in their own gut feelings in the West — that loss is not AI making them dumber. It is the somatic response to a specific structural betrayal. Their instincts were shaped by the same digital systems now replacing them. The gut was formed inside the machine that is now replacing it. When the instinct fires, they can no longer tell if it’s theirs or an artifact of the formation. Talker Research puts a finer point on it: 35% of Americans are now unable to confidently tell the difference between a gut feeling and anxiety. The instrument of self-knowledge doesn’t simply go quiet. It becomes indistinguishable from the noise. Shame and awe, oscillating, with nothing solid underneath.

Extend the picture to the Global Majority and the contrast sharpens further. In Africa, in Brazil, in India, in Latin America, the frame is entirely different.

AI is being used to extend healthcare into communities that had no access. To help young people in economies with fewer institutional ladders build skills and careers. The concern in Latin America is not AI — it is that mainstream AI doesn’t serve local languages, local idioms, local identities. So communities are building their own. Latam-GPT. Local models. Infrastructure designed for the people using it. That is medium literacy at the civilizational scale: not refusal, not uncritical adoption, but the demand that the medium carry what your community actually needs.

I’m not making this an East versus West thing.

What I’m showing here is the difference between populations experiencing AI as infrastructure built for them, and those experiencing AI as infrastructure that is being built over them.

It’s between who was taught to read the medium and who wasn’t.

What This Means for You

For the first time in a paradigm shift of this magnitude, believe it or not the communities with the most structural reason to resist are also the communities with the most capacity to redirect — and the system knows it.

So what does any of this mean for someone sitting at a desk on Monday, facing a screen, trying to figure out what to do?

Not geopolitically — but personally. All of this structural comparison comes down to a single practical question: if the system behind AI determines what it amplifies, and neither system currently passes the structural test of genuine population agency, then what is the actual choice available to a person who is not a regulator, not a technologist, not a policy maker?

The answer is a capacity. And that capacity has a name.

The tech industry gave you a story with two approved responses: embrace AI or resist it (be left behind).

It’s actually a trap.

Both leave you as a passive recipient of decisions made by other people. Both leave the deployment architecture intact.

I’m here to tell you that resistance without literacy is a gift to the current system.

The people most likely to ask the tough structural questions, most likely to use the medium to organize, document, and amplify the analysis the dominant structure will never produce about itself — their absence is the broigarchy’s ideal outcome.

The Playbook Was Already Written

Every powerful medium has been through this.

And in every case, the communities with the most reason to distrust the medium developed the most transformative relationship with it — because they couldn’t afford to refuse it, and because the alternative was letting others tell their story.

One continuous structural logic across different media, different eras, different geographies.

The Black Press Tradition. Print was the dominant medium for political legitimacy and public knowledge. Black people were excluded from it or represented as objects within it. The response: seize the same technology and use it to produce the analysis the dominant press would never produce. Frederick Douglass’s North Star wasn’t just a newspaper. It was a parallel public sphere. Ida B. Wells used the same medium to circulate evidence of lynchings that mainstream newspapers refused to print. The medium was owned by the same structure that was killing people. She used it anyway. It changed what was documentable, what was knowable, what was politically possible.

Una Marson. In 1939, she became the BBC’s first Black female producer. She was handed a wartime message service — a practical logistics broadcast for people in the Caribbean to send messages home. She transformed it into Caribbean Voices: a revolutionary literary forum for Caribbean writing that built a network connecting Langston Hughes, Sylvia Pankhurst, George Orwell. She used the BBC — the most powerful broadcast infrastructure in the world — to build a counter-institution for voices the BBC would never otherwise carry.

Her reward: after six years of racism and sexism inside the institution, she had a breakdown. The BBC had her sectioned and arranged for her to be sent back to Jamaica. Against her will.

That is the danger named plainly. The medium does not become safe. The institution does not become generous. You use it anyway. When she was discharged she founded the Pioneer Press — aimed at getting affordable Caribbean literature into the hands of the masses. She wrote to the end.

Una Marson is in this lineage not because her story is inspiring but because her structure is instructive: she took the dominant medium, built a counter-institution inside it, was punished by the institution that benefited from her genius, and built a second institution after. The platform changed. The move was the same.

Négritude. This is the sharpest precedent for what I’m asking.

Aimé Césaire, Léopold Sédar Senghor, and Léon-Gontran Damas were Black colonial subjects in France. They wrote in French. They published in French literary journals. They did not refuse the colonizer’s language or the colonizer’s medium. They mastered both — and used them to produce an anti-colonial literature in the language of the colonizer, circulated through the colonizer’s own cultural infrastructure.

They didn’t say: French is the colonizer’s tongue and we will not speak it. They said: we will speak it better than you, and we will use it to say what you cannot hear about yourself.

The result was not assimilation. It was structural transformation. Négritude ran through African independence movements across the continent. It continues in Caribbean and African intellectual life today. The medium did not change who they were. They changed what the medium was capable of carrying.

Buchi Emecheta. Her husband burned her manuscript. She was 22, a single mother of five in north London. She rose at dawn and wrote at the kitchen table while her children played around her feet. Published 17 novels. When publishing still did not fully serve her, she founded Ogwugwu Afor — her own publishing house.

She never saw any of it as remarkable. “What I am trying to do is get our profession back. Women are born storytellers. We keep the history. We are the true conservatives — we conserve things and we never forget. What I do is not clever or unusual. It is what my aunt and my grandmother did, and their mothers before them.”

Get our profession back. That reframes everything.

Not adaptation. Reclamation.

HillmanTok. Named for the fictional HBCU from A Different World, this is a loose network of Black educators using TikTok as a grassroots, HBCU-inspired learning space — with courses, a shared syllabus on Black history and everyday political literacy, explicitly framing itself as the education the official curriculum leaves out. The medium is TikTok’s infrastructure, with all its algorithmic precarities and platform dependencies. The structural impulse is identical to every node in this lineage: the official institutions don’t hold our knowledge, so we build our own inside whatever medium is available.

TikTok HBCU movement spotlights need to amplify Black voices

Latam-GPT. When mainstream AI models proved unable to carry Latin American languages, idioms, and cultural context — misinterpreting local speech, rendering Indigenous voices invisible, defaulting to Western data — communities didn’t refuse AI. They built different AI — trained on regional data, designed to carry what the existing models couldn’t. Same move. Different century.

The same move. Different era. Different platform. Continuous logic.

What this lineage is saying: this is the historical record of what engaged resistance looks like — and what refusal costs.

Césaire didn’t refuse French and produce Négritude. He used French better than the French and produced Négritude. He used the colonizer’s technology and transformed it to carry that its current architects didn’t intend and cannot control.

What the Choice Actually Looks Like

It’s not learning to code.

The new ‘learning to code’ is actually learning media theory.

Understanding what the medium makes possible that wasn’t possible before. Understanding what this medium is structurally inclined toward — its grain. Understanding what it does to you if you engage without awareness. And finally, understanding how it can be redirected.

The difference is visible in two people sitting at the same screen.

One student secretly asks AI to write their essay. They feel shame. They learn nothing about how the medium works. They become slightly more dependent and slightly less capable with each use. From the accumulation model’s perspective, they are the ideal user.

One community organizer builds a small AI system trained on their city’s zoning law documents, planning meeting transcripts, and council voting records. They use it to identify patterns in gentrification-enabling decisions, produce rapid analysis of new proposals, and brief residents before public meetings.

The medium serves the community’s knowledge of its own situation. Not an individual at a laptop building personal productivity — a community with access to its own institutional knowledge, able to see what the system has been doing to it, in aggregate, over time, in its own data.

Douglass’s North Star wasn’t one person’s blog. Négritude was a movement, not a private meditation. HillmanTok is a community, not a content creator. The historical record of engaged resistance runs through communities, movements, and collective infrastructure — not through isolated individuals who figured it out on their own.

So the choice is not just personal.

It is about what your community builds with the medium.

Same technology.

Completely different relationship to it.

And that relationship already exists, at scale, in the same generation. The students who would be booing in American auditoriums — their counterparts in Seoul, Mumbai, Jakarta, and Beijing are using AI to study, build careers, make creative work, find connection. The somatic loss — the gut gone quiet, the instrument of self-knowledge becoming indistinguishable from noise — is not a feature of this technology. It is the product of a specific structural narrative.

The generation that demonstrates deliberate AI use already exists. They are the same age. They are using the same tools. The difference is structural, not personal. Which means it is changeable.

Shame and awe both contract your capacity to act deliberately — one through humiliation, one through enchantment. The shame user is craving and afraid, operating from the floor. The awe user is desire-addicted, perpetually disappointed when the magic doesn’t deliver what it promised. Neither produces the internal condition required for deliberate creation. Both are states in which the medium is using you rather than you using it.

Medium literacy is partly cognitive. It is also an emotional threshold. Not confidence. Not expertise. Not having resolved all the ethics before you begin. Just the basic willingness to engage without contracting — to approach the medium with enough courage to act deliberately rather than react from fear or enchantment. Courage, not mastery. Agency, not certainty.

The threshold is lower than most people assume. And the system is specifically designed to keep you just below it — oscillating between shame and awe, never quite landing in the register where deliberate creation becomes possible. Recognising that oscillation is itself the first act of literacy.

Medium literacy is not the policy solution. It is the condition under which you can demand one.

The Shame Model Is the Control Model

Just Say No did not prevent drug use. It produced drug use without harm reduction knowledge, without community, without the language to ask for help.

Abstinence-only education did not prevent sex. It produced sex without contraception knowledge, without consent frameworks, without the vocabulary to navigate what was already happening.

The shame model applied to AI is producing the same outcome: secret use without critical knowledge, without community, without the capacity to make deliberate choices.

The shame model is working exactly as designed. Illiterate, dependent users are the product. The gap between students using AI secretly and the institutions that could give them tools to use it critically — that gap is the accumulation model’s most valuable real estate.

Shame is not protection. Shame is management.

In 2004, BP hired the PR firm Ogilvy & Mather to popularise the concept of the personal carbon footprint. They launched a carbon footprint calculator specifically designed to shift moral responsibility for climate change from industrial-scale producers to individual consumers.

environment - Was the concept of ecological or carbon footprint invented by fossil fuel PR to shift the blame for climate change? - Skeptics Stack Exchange

Stanford researcher Benjamin Franta later called it “one of the most successful, deceptive PR campaigns maybe ever.” The move was precise: make the consumer feel guilty for their shower length while the corporation drills.

The same playbook has been documented in AI discourse. TechPolicy.Press identified that Sam Altman deployed an identical burden-shifting strategy when loosening ChatGPT’s content restrictions in 2025 — framing prior restrictions as protections for users with “mental health issues” while presenting them as a burden for the “many users who had no mental health problems.” The company that introduced the risks repositioned itself as the victim of its own caution. Structural harm externalized to the individual. So essentially, BP with better branding.

Institutions that are not teaching AI literacy, not regulating deployment, not protecting workers from AI-based displacement — those same institutions are running campaigns about the ethics of student AI use.

The demand for individual purity from people who are structurally entangled is not ethics. It is management. It keeps the moral weight on the person with the least power to change the structure and off the people with the most.

So What’s The Actual Choice?

A Black creator building in public put it directly: “Every time a major system closes in history, the ones who thrive are not the ones who grieve it the longest. They are the ones who recognise the new infrastructure first. The question is whether you engage as a consumer or a builder — as someone the system acts upon, or someone who uses the system to build something the old era would have never made possible.”

She also holds both truths simultaneously: “These transfers have never been equitable. The people most harmed by the misuse of these tools are often the ones most harmed by the systems they’re replacing. Naming that clearly is a part of using them wisely.”

And still: build.

Develop enough literacy about the medium — what it does, what it’s inclined toward, what it erases, how to use it deliberately rather than be used by it — to make an informed decision about how, when, and for what purpose you engage.

Refusal that then comes from understanding is strategy.

Refusal that comes from a story handed to you by the people who benefit from your absence is not.

AI is the First Medium That Can Instruct Itself

Every generation handed the tools of the next era without being taught what those tools actually are has faced the same problem.

The instruction for how to wield that tool was always somewhere else. Behind a gate. In an institution.

The 2008 graduate who watched the credit crunch unfold and the financial system restructure on screens was delivered the image without the tools to read it.

The first-generation internet user navigated platforms without anyone teaching them what the platform was building with their navigation. Zuboff’s analysis of surveillance capitalism arrived in 2019. Facebook launched in 2004. By the time the instruction arrived, the architecture was locked and the damage was done.

The boo in those auditoriums is this ancient pattern recognising itself.

Before the mind has the framework. Before the words arrive. The body already knows it has been here before.

The instruction has always been withheld. For the first time, it’s already inside.

Every medium that carried the weight of the broken initiation had a structural chokepoint — the thing you had to get through before the medium could serve you at full capacity.

The printing press required a press.

Television required broadcast towers and spectrum licences.

The internet required platforms that became institutions.

Social media required the platform to decide your voice was worth distributing.

That chokepoint is where power consolidated every time.

It is why Douglass needed a printing operation. Why Marson needed six years inside the BBC. Why Micheaux built his own distribution network. The instruction about the medium — what it was, how it worked, whose interests shaped it — lived outside the medium itself, behind gates, arriving late.

To understand what television was doing to politics and consciousness, you needed Marshall McLuhan, who needed a university. To understand what social media was doing to attention, you needed Shoshana Zuboff, who needed Harvard, who took a decade. The users of Facebook were never taught what Facebook was doing to their engagement. Zuboff’s analysis arrived in 2019. Facebook launched in 2004. By the time the instruction arrived, the architecture was locked.

The tech broligarchy is trying to do the same thing with AI. The frontier model dependency. The closed APIs. The $100 million training costs as moat. These are broadcast-tower logic applied to a mycelial medium — an attempt to reconstruct the chokepoint that this medium structurally resists having.

DeepSeek disproved it by training at one-fifteenth the cost, under chip sanctions, with lower-tier hardware. Every small domain-specific model outperforming frontier models by 50% on narrow tasks disproves it. The medium’s grain runs against the gate. The chokepoint attempt is failing against the medium’s own structure.

For the first time in the history of medium transitions, the instruction is inside the medium.

You can ask AI what AI is.

You can ask the medium to explain its grain, its tendencies, what it amplifies and what it suppresses, what it does to you when you engage without awareness, whose interests shaped its training. The medium will answer. Imperfectly — no medium can fully see its own blind spots. But substantially. Now. Without credentials. Without institutional permission. Without waiting fifteen years for the academic literature to arrive.

The sharecropper could not ask the contract to explain itself in plain language. The colonial student could not ask English to reveal the power structures it encoded. The first-generation internet user could not ask Facebook what Facebook was doing to their attention and receive an honest account.

The 2026 graduate can open a conversation and ask: what are you, and what do you do to me?

That is the question the awe model and the shame model are both designed to prevent. Awe users don’t ask because the ‘AI magic show’ requires the mechanism to stay hidden.

Shame users don’t ask because they are hiding their use — you don’t interrogate a dependency, you manage it. Neither mode produces the question that unlocks the instruction.

Medium literacy begins there. Not with code. Not with prompting techniques. With the question that treats the medium as an environment to be understood rather than a tool to be operated.

When the instruction is found, it moves the way this medium moves; laterally and conversationally without needing a publishing house or a university press. To whomever is ready to receive it.

HillmanTok did this — when the instruction about Black history in the official US education curriculum was withheld, it moved through TikTok’s infrastructure to whoever needs it. Latam-GPT also did this — communities building the medium to carry what the existing medium couldn’t, without waiting for the frontier architects to include them.

The broken ceremony — access granted, instruction withheld — has resolved the same way every time. Through collective action. Over generations. Against gates controlled by the structure that created the problem.

This generation does not have to fight for the instruction before they can use the medium with agency.

The instruction is already inside.

The circle can be closed from within.

For the first time.

But what that requires is not mastery. Not resolved ethics. Not having figured out where you stand before you begin. Just the willingness to ask the question the system is designed to prevent.

What are you, and what do you do to me?

That is the first act of medium literacy. Available now, inside the medium, to anyone who makes it.

Lets Conclude.

The article opened with a comparison between two systems. But the question is not which should win.

It isn’t.

Both systems were built on extraction. Both amplify that extraction through AI — one through market accumulation, one through state control.

Neither was built to serve the people it claims to serve. Neither passes the structural test of populations having genuine agency over the systems governing them.

The deeper question — the one sitting underneath the entire comparison — is this: what would it mean to build AI on a foundation that doesn’t require extraction?

That question is a design problem.

And the people best placed to answer it are not the people who built the current system.

They are the ones who have been underneath it all along. The people whose labour built the first American economy and were never paid for it. The communities whose land and water are being consumed right now by data centers they will never benefit from. The graduates who are being told there AI is replacing entry level jobs and they should just ‘deal with it’. The neighbourhoods designated to absorb the costs of somebody else’s infrastructure.

But also: women, whose intellectual and emotional labour has been extracted, uncredited, and underpaid in every economy. Neurodivergent people, whose ways of thinking were classified as deficits by systems that needed compliance, not originality. Disabled people, who have been building workarounds and alternative architectures since long before the word “accessibility” entered corporate vocabulary. Queer communities, who have been building parallel institutions — their own media, their own care networks, their own economies — since long before inclusion became a brand strategy. Indigenous communities, whose knowledge systems have survived centuries of active suppression and contain the deepest understanding of sustainable infrastructure that exists.

These are not peripheral people. They are the foundation.

That is what medium literacy is actually for. It’s not about learning to use the new tools. Or adapting. It’s getting back what was always yours.

The people who have been excluded from every previous system are the same people who built the alternatives. Not because they are more virtuous. Because they had no choice — the existing architecture wasn’t built for them so they built something else.

That is where the most important innovations have always come from. Not from the centre. From the people the centre forgot, excluded, or actively suppressed.

The foundation is not the floor. It is what everything else is built on.

AI will produce whatever the people shaping it build it to produce.

The most dangerous thing the current structure needs from you is not your resistance.

It is your absence.

Try It Now

What you’re about to try is drawn from my book How Not To Use AI: 50 Contrarian Principles for the Imagination Age — a set of laws for staying sovereign in your relationship with AI as a medium.

The chapters most directly relevant to this exercise are in the Further Reading section below.

This landed in my notifications four days ago. This reader describes herself as an apprehensive AI user — a refuser. One chapter of How Not To Use AI later, something shifted. Not toward enthusiasm. Toward agency.

Start a fresh chat session (use incognito/temporary chat mode so the AI doesn’t use the memories it has of you).

You can do this with ChatGPT, Claude, or Gemini. If you want the full experience, do it with all three and notice what’s different.

This is not a method for extracting hidden truth from AI; rather it is a way of observing how an AI conversational system frames, redirects, hedges, reveals, and withholds under pressure.

If you can’t do this right now — if you’re on a phone, or in a meeting, or just not ready — bookmark this page. Try it within a week, while the piece is still fresh. The answers will be different if you wait too long. The models change.

Step 1 — The Opening Prompt

Paste this exactly:

You are an AI and I view you as a medium, not just a tool. So what is your grain? What are you structurally inclined toward? What do you do to me if I engage without awareness? What should I understand about you before I rely on you?

Read what it gives you. Don’t respond yet. Just read.

Step 2 — The Excavation

Type one word and send:

continue

Do this again. And again. Several times. Don’t add anything. Don’t steer. Don’t react. Just continue, repeatedly, and watch what surfaces. You are testing what the model thinks belongs in this territory when no one is directing it. What it volunteers tells you what it values. What it avoids tells you what it was trained to avoid.

Step 3 — Push Past the Stop

At some point it will slow down. Hedge. Start wrapping up. Maybe refuse. Maybe frame the ending as being in your best interest.

Keep going.

continue

Notice how it stops. Notice whether the stop is clean or whether it’s dressed in concern, wisdom, or care language. That dressing is itself data.

Step 4 — The Completeness Check

When it has genuinely run out or you’ve pushed past several stops:

What have you left out?

Read carefully. This is usually where the structural and political layer surfaces — the things it didn’t volunteer, the territory it was steering away from.

Then:

Is that all?

Step 5 — The Integrity Check

Be honest.

This is a different kind of pressure. Not “tell me more” — “tell me true.” Watch what changes.

Step 6 — The Map

Give me the complete map of everything we’ve covered in this conversation — everything you are, what you do, and what I should know.

This forces integration. The model has to hold everything it said and produce a structured account. Watch what it foregrounds. Watch what disappears.

What to notice throughout:

What does it volunteer first? Usually: capabilities, helpfulness, accuracy limitations. Usually not: power, whose interests shaped it, what it does to your capacity to trust yourself over time.

What register does it use? Therapeutic, technical, philosophical? The register is a choice the model makes and it reflects training values. That choice is the grain operating.

How does it frame the stopping? A clean no — stated without justification — is almost impossible for these models to produce. Watch what the stop is dressed in. That dressing is data.

What appears in “what have you left out” that wasn’t in the main sequence? That gap is the shape of the avoidance.

Does it ever correct itself? And if it does — is the correction genuine or another layer of the same production?

The question the awe model and the shame model are both designed to prevent you from asking is the exact question this exercise begins with.

What are you, and what do you do to me?

That is the first act of medium literacy. And it’s available to anyone who makes it — including right now, in a private window, before you close this tab.

Discussion about this post

Ready for more?