AI, Rockets, & Bicycles
Technological ethics for the age of AI: tools, systems, and how to tell the difference.
Open Public-facing — newcomers + returnees Set Conversation, not lecture Returnees Name Part 1 attendees briefly Pacing Don't promise too much
This is the second of a two-part series on AI, technology, and how to think about what we've built. Part 1 read Mark 12 — Jesus's reframe of the trap question about Roman taxes — and asked what that move offers Christians facing a paradigm-shifting technology. Part 2 turns the same lens outward. What is AI actually made of? Who built it, and why? How should we use it? The whole workshop turns on a single question: whose image, whose benefit, whose cost?
Is AI good or bad?
Throw 60–90s. Feel the binary as trap. Don't resolve Reframe comes next slide
Anticipate
"It depends" depends on what? "It's a tool" hammer is a tool — is a power plant? "Next industrial rev" was that good or bad? Yes. Silence name as honesty — most of us don't know
"Is AI good or bad?" is the wrong question. It traps us into dismissal or unqualified embrace — neither gives us a way to think about a technology already in our work, our schools, our homes, operating at a scale that overwhelms simple judgments. The point of starting here isn't to answer. It's to feel the question as a trap before we reframe.
What is AI actually — what's it made of , who built it, what's it costing — and how should we think about using it?
Reframe refuse the binary Arc made of (empire) → who built it (incentives) → how to use it (bicycles)
Order: empire critique first, race-to-intimacy second, bicycles third, practical answers fourth. The room earns each section by working through the previous one.
Promise land with something concrete for Monday morning
The better question has three parts. What is AI made of — the empire underneath, drawing on Karen Hao's reporting. Who built it, and why — the design choices and incentives, drawing on the Center for Humane Technology. How should we use it — the practical answer, the bicycles-and-rockets distinction. Those three sub-questions structure the rest of the workshop. We'll land with something operational.
60-Second Recap
For those who weren't at Part 1.
Pacing Fast — 60s total for next 2 slides Goal Bring newcomers to the starting line
A brief recap of Part 1, for readers who didn't attend: the workshop read Mark 12 together — the moment where Jesus is asked whether to pay taxes to Caesar and refuses the binary, asking instead to see the coin and pointing at Caesar's image. The denarius wasn't just currency; it bore the emperor's face and an inscription claiming his divinity. It was a portable claim of allegiance, woven into every ordinary transaction. Jesus's move — whose image, whose benefit, whose cost — is the tool the rest of this session uses.
Roman roads, coinage, taxation — each carried a claim
The denarius bore Caesar's image and an inscription claiming divinity
You couldn't buy bread without participating in that claim
Pacing ~30s, narrate Point imperial tech is rarely neutral — it carries a claim
The denarius bore the inscription "Tiberius Caesar, son of the divine Augustus, high priest." Every coin was a small religious-political claim. The point isn't to memorize the Latin — it's to feel that you couldn't buy bread without participating in the claim.
Land "portable idol of allegiance"
Imperial technology is rarely neutral. Roman roads were built on tribute and forced labor; the coinage carried Caesar's face and a claim of his divinity; the tax structure pulled the provinces into the imperial economy whether they consented or not. The denarius was a portable idol of allegiance. You couldn't buy bread without participating in that claim. That's the lens we want to carry into thinking about AI.
Jesus refused the binary — neither "pay" nor "refuse"
He gave them a tool: whose image, whose benefit, whose cost?
Today, we apply that tool to AI
Pacing ~30s, narrate Land "whose image, whose benefit, whose cost" — load-bearing for whole session Connect Don't elaborate Mark 12 — that's Part 1's work
Jesus's move was to refuse the binary entirely. He didn't say pay the tax; he didn't say refuse it. He asked to see the coin, pointed at Caesar's image, and gave them a different tool: whose image is on it? Whose benefit? Whose cost? Those three questions, lifted out of their original context, carry into ours. They are the question we'll keep asking about AI all the way to the closing slide.
Is AI Good?
Tonal shift Pause. Smile. Why this order Lead with their use → critique lands. Lead with critique → they get defensive.
The room will be carrying anxiety from the empire framing if you've foreshadowed it. Switching to "where is AI working in your life" gives them permission to name the goods they're actually experiencing. The critique that follows lands differently because it doesn't feel like a personal attack on their tools.
Before naming what's wrong with AI, we want to name what's right. Most people in the room are already using AI in ways they haven't quite named. Some of those uses are bicycles and some are rockets, but all of them are real. The structural critique that follows lands differently if we start here — not as an attack on the tools people are using, but as a question about what those tools cost.
Where is AI actually working in your life right now?
Throw 3–5 min — let the room work Don't push no "right" answer — just collect
Anticipate
School kids' homework, teacher tools Work drafting, summarizing, notes Coding developers in the room Medical radiology, derm, imaging Accessibility transcription, image description Translation everyday + endangered languages
The first piece of room work. Where is AI actually showing up in people's lives? School use shows up immediately — kids working through homework with ChatGPT, teachers using it to grade. Work tasks come next — drafting emails, summarizing reports, generating meeting notes. Then medical imaging, accessibility tools, translation, coding assistance. Most of these are bicycles, doing narrow jobs well. We'll come back to which ones are which.
AlphaFold — protein folding; 2024 Nobel in ChemistryMedical imaging — radiologists with AI catch cancers earlierAccessibility — transcription, reading, image description for the blindA patient tutor for anyone with internetEndangered languages — translation preserving what would be lostRare disease diagnosis — patients undiagnosed for years finding answersDrafting and summarizing — extending one person's reach
Reveal after the room has surfaced 4–5 Lead with AlphaFold + rare disease — most underappreciated
AlphaFold: DeepMind's protein-folding model. Solved a 50-year biology problem. 2024 Nobel in Chemistry (Demis Hassabis + John Jumper). Now accelerating drug discovery for diseases we couldn't touch before.Rare disease: patients undiagnosed for years finding answers because an AI surfaced patterns across genetic data their doctors couldn't see.
Don't read list pick 2–3 the room missed
The good of AI is real and specific. AlphaFold — DeepMind's protein-folding model — solved a fifty-year-old problem in biology and won the 2024 Nobel in Chemistry. Medical imaging tools catch cancers earlier than radiologists working alone. Accessibility tools give blind users image descriptions and transcripts. A patient tutor available to anyone with an internet connection — for kids working through algebra at midnight, adults learning to code, lay people studying Greek. Translation tools preserving endangered languages. Rare-disease diagnoses for patients undiagnosed for years. These aren't speculative goods. They're shipping.
What do these good uses have in common?
Throw 2–3 min Build to narrow · extends · keeps the judgment
Anticipate
"Small / specific" yes → narrow "Helps people" yes, but how? "Doesn't replace" yes → 3rd feature
What do the good uses have in common? Three features: they're narrow (focused on one task), they extend a human (rather than replacing one), and the human keeps the judgment. These three features are the load-bearing pattern of the whole workshop. They come back at the end as the bicycle test.
Narrow — focused on one taskExtends a human — rather than replacing oneThe human keeps the judgment
Hold these three. They come back at the end.
Reveal after room names ≥2 Land these 3 come back as the bicycle test Setup empire critique = what happens when these 3 invert
Inversion: rocket-AI is general (not narrow), replaces humans (not extends), takes the judgment (not keeps it). That inversion isn't a bigger version of the good — it's a structurally different thing. The next section names it.
Narrow. Extends a human. Keeps the judgment with the human. Hold these three — they're the architecture of bicycle-AI, and they're what gets inverted when we look at the largest systems being built. Rocket-AI is general (not narrow), replaces humans (not extends), and takes the judgment (not keeps it with the human). That inversion isn't a bigger version of the good. It's a structurally different thing. That's what the next section names.
The Empire
Tonal shift From what AI does well → what it costs Don't preview let next 2 slides introduce the guides
The structural critique. We've named what AI does well. Now we ask how the largest AI systems are actually built — and what they extract from the communities, workers, and creators caught up in their supply chain. Two guides anchor this section: the journalist Karen Hao, and the Center for Humane Technology. The first reports what AI companies are doing; the second diagnoses why the system produces it.
Karen Hao
MIT Technology Review · Wall Street Journal · The Atlantic
First reporter to profile OpenAI from the inside (2019)
Empire of AI : 300+ interviews, 90+ current/former OpenAI employees~7 years of reporting
Pacing ~30s Key move "empire" = historically rooted, not rhetorical
She's drawing on the actual transition of the British East India Company from trading firm to colonial power. When she calls AI companies empires, she's making a structural diagnosis , not throwing an insult.
Frame she's being precise , not activist
Karen Hao is one of the most rigorous AI journalists working today. She profiled OpenAI from the inside in 2019 — the first reporter to do it. Her book Empire of AI , released in 2025, draws on seven years of reporting, 300+ interviews, and 90+ current and former OpenAI employees. When she calls AI companies empires, she's being precise — drawing on the historical model of the British East India Company's transition from trading firm to colonial power. That precision matters. Empire isn't being used as an insult. It's being used as a structural diagnosis.
Center for Humane Technology
Tristan Harris · Aza Raskin · Randima Fernando — founded 2018
Insider perspective on how tech design reflects company incentives
Featured in The Social Dilemma
Core question: how can technology better serve society?
Pacing ~30s Pairing Hao = what ; CHT = why Credentials Harris = ex-Google ethicist; Raskin = invented infinite scroll
Tristan Harris: design ethicist at Google, left after his concerns about attention-design were ignored internally. Aza Raskin: invented infinite scroll on the web in 2006 — has spent the years since trying to undo it. Their critique is from inside.
The Center for Humane Technology was founded in 2018 by technologists — Tristan Harris was a design ethicist at Google; Aza Raskin invented infinite scroll and has been trying to undo it ever since. Their perspective is insider; they've watched the design choices that produce the harms. Where Hao reports what AI companies are doing, CHT diagnoses why the system produces it — misaligned incentives, racing to dominance, design choices that exploit rather than serve. Two perspectives on the same problem.
What's the difference between a business and an empire ?
Throw 3–4 min — load-bearing for whole section Build to extraction you can't refuse
Anticipate
"Scale" every empire was a business once "Power" in what shape? "Geography" what about financial empires? "Force" EIC was a corp before it had an army "Can't opt out" → kicker — push on this
This is the load-bearing question for the whole section. The difference isn't scale, isn't power per se, isn't geography. The threshold is extraction you can't refuse. A business takes your money and gives you something in exchange. An empire takes more than it returns, and you have no recourse. The British East India Company was the first; it became the second — and the transition was gradual, almost invisible until it was complete.
A business engages in fair, voluntary, reciprocal exchangeAn empire extracts — takes more value than it returns — and there's nothing anyone can do about itThe British East India Company started as the first, became the second
Hao's claim: AI companies have crossed that line.
Lean on EIC story — most won't know the details
EIC story is unfamiliar to most rooms; expect to need to explain the transition. The most useful detail to lead with is the army-by-1750s — that's the moment people register the shift from commerce to coercion.
EIC arc 1600: trading firm → 1750s: army → 1800s: governed India
1600: chartered by Elizabeth I as a trading company for spice and textile trade.1750s: had its own standing army of ~260,000 troops — about twice the size of the British army at the time. The Battle of Plassey (1757) established military control over Bengal.1800s: directly governed roughly 200 million people across India, collected taxes, ran courts, set policy. Dissolved 1858 after the Indian Rebellion, with rule transferred directly to the British Crown.
Land commerce → governance, never returned Hao's move not metaphor — structural reading of the same transition
The East India Company was chartered in 1600 as a trading firm. By the 1750s it had its own standing army. By the 1800s it governed India directly — collected taxes, set policy, controlled the courts. It went from commerce to governance, and it never returned to being just a business. That transition — from voluntary exchange to coerced extraction — is the threshold Hao uses to read what's happening with AI companies now. It's not a metaphor. It's a historical pattern repeating.
If the harms aren't accidents, what are they?
Throw CHT's question, 2–3 min Build to misaligned incentives — harms are features, not bugs
Anticipate
"Side effects" but if predictable, are they really side ? "Design choices" yes — what kind? "What it optimizes for" → kicker
If the harms produced by AI companies — to data center neighbors, to displaced workers, to creators whose work was ingested without consent — are predictable, repeatable, and produced by the same design choices across companies, can we still call them side effects? CHT's diagnosis is no. The harms are features of the system, not bugs. They're what the system was built to produce, given the incentives it was built to respond to.
Companies race for engagement, dominance, market share — not human well-being
What gets optimized (time spent, data captured, revenue) diverges from what matters (trust, dignity, flourishing)
Whoever optimizes engagement hardest wins. Whoever doesn't, loses.
The harms aren't accidents. They're what the system was designed to produce.
Land the kicker line — pause on it Why it matters explains why harms recur across different companies
It's not that some companies are good and some are bad. The system selects for the behaviors we're seeing. Any company that doesn't optimize engagement gets out-competed by one that does. That's why the harms keep showing up at every major AI company — they're not bad actors, they're operating in a market that selects for these outcomes.
Here's the structural argument. AI companies compete for engagement, market share, and dominance. Those are the metrics that matter for surviving in this market. They are not the metrics that matter for human flourishing. When you optimize one set and ignore the other, you don't get a balanced outcome — you get whatever the optimization produces. And whoever optimizes hardest wins. Whoever doesn't, gets out-competed. That's why the harms keep recurring across companies. It's not that some companies are bad; the system selects for the behavior we see.
If AI companies are empires, what do they extract from?
Throw 2–3 min — single anchor for 4 marks coming Map answers to marks 1–4
Anticipate
Data → Mark 3 Labor → Mark 2 Energy/land → Mark 1 Attention → all four Narrative → Marks 3 & 4
Next 4 slides are narrative — move at pace, ~90s each. Synthesis lands at "we are subjects."
If AI companies are empires, the next question is what they extract from. Hao identifies four marks. Each maps to a historical pattern any student of empires would recognize: land and resources, labor, knowledge production, and narrative control. We'll walk through each.
Mark 1 — Land & Resource Extraction
Data centers requiring gigawatts , sited in working-class & minority communities
Memphis xAI Colossus — 35 unpermitted methane turbines in a Black neighborhood with high lung cancer ratesAbilene Stargate — "the size of Central Park"Communities competing with these facilities for fresh water during droughts
Lead with Memphis xAI — strongest case Memphis 35 unpermitted methane turbines, NAACP suing
xAI Colossus, South Memphis (2024): Elon Musk's data center installed 35 methane gas turbines without environmental permits. Sited in a historically Black neighborhood with elevated lung cancer rates and longstanding air-quality issues. Local NAACP filed suit. State agency began investigating after community organizing.
Abilene "size of Central Park" — OpenAI's words Empire Rome's roads/aqueducts on tribute + forced labor Effect utility bills, grid reliability, air, water, asthma
Water fights hit drought regions hardest — Chile, Arizona, the Pacific Northwest — where communities compete with data centers for fresh water.
Pull in anyone in room with local DC fights
The first mark of empire is land and resource extraction. Data centers requiring gigawatts of power are being sited in working-class and minority communities. The xAI Colossus in Memphis installed 35 methane gas turbines without permits, in a historically Black neighborhood with already-elevated lung cancer rates; the local NAACP is suing. OpenAI's Stargate facility in Abilene, Texas will be — in their own words — "the size of Central Park." Communities compete with these facilities for fresh water during droughts. The pattern is the pattern of every empire: infrastructure funded by tribute taken from the places empire reaches.
Mark 2 — Labor Exploitation
Hundreds of thousands of data annotators globally — Kenya, Philippines, India, increasingly U.S. — pennies, often reviewing traumatic content
Career-ladder break: marketer laid off, takes a contract gig training the model on the job she just lost
~40% reduction in entry-level hiring in fields most exposed to AI
LLMs trained explicitly for the industries where most profit can be made
Lead with marketer-training-her-replacement — verbal image
The image: twenty-year veteran marketer laid off because AI can do "most of what she did." Needs work. Contractor deploying the AI hires "model trainers" — evaluators correcting AI output using domain expertise. She takes the gig at a fraction of her old salary. Every correction she makes trains the model to need fewer people like her.
Numbers 40% entry-level hiring drop — sit with this For parents kids in their 20s — this is the number Empire labor that makes the system is the least protected; value flows upward Effect lost work, dignity, meaning — the on-ramps that let one generation get more secure than the last are breaking
The second mark is labor exploitation. Hundreds of thousands of data annotators worldwide — in Kenya, the Philippines, India, increasingly the U.S. — are paid pennies to review and label content, often traumatic content, to train the models. Closer to home: a marketer laid off because AI can do "most of what she did" takes a contract gig training the model on the job she just lost. Her domain expertise becomes the training signal that erodes the next round of jobs like hers. Entry-level hiring in fields most exposed to AI has dropped about 40%. The on-ramps that let one generation become more secure than the last are breaking.
Mark 3 — Capture of Knowledge Production
The AI industry funds most AI researchers in the world
If most climate scientists were bankrolled by fossil fuel companies, would we get an accurate picture of the climate crisis? Dissidents removed — Timnit Gebru at Google is the canonical case
Watchdogs face legal pressure (OpenAI subpoena fishing expeditions)
Pause after the climate-scientist rhetorical Q Gebru slow down — most won't know her
Timnit Gebru: co-led Google's AI Ethics team. In late 2020 she co-authored "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" — warning about environmental costs, training-data harms, and unintended consequences of large language models. Google asked her to retract or remove Google-affiliated authors. She refused. Google announced her departure as a resignation; she calls it a firing. Has become the canonical case of what happens to AI watchdogs who get inside.
Empire every empire controls its story — Rome's historians, Britain's mapmakers Effect the public can't tell what's true — and neither can policymakers
The third mark is capture of knowledge production. The AI industry funds most AI researchers in the world — through industry jobs, through academic positions dependent on industry partnerships, through the conferences and journals that shape what counts as a finding. Hao asks the rhetorical question that lands hardest: if most climate scientists were bankrolled by fossil fuel companies, would we get an accurate picture of the climate crisis? The answer is obvious. The dissidents are removed — Timnit Gebru, the AI ethics researcher who co-authored "Stochastic Parrots" warning about the risks of large language models, was forced out of Google in 2020. The mechanism is corporate, not violent, but the effect is the same: the public can't tell what's true; policymakers can't either.
Mark 4 — Imperial Mythmaking
Good empire vs. bad empire — "we must build it first or China will"
Dual narrative: utopia or catastrophe — both consolidate power
Sam Altman defines AGI four different ways: Congress (cure cancer) · consumers (best assistant) · Microsoft ($100B revenue) · website (outperform humans economically)
Vagueness isn't a bug. It's the product.
Read aloud the 4 AGI definitions — let absurdity land
Altman's four AGI definitions:
To Congress: AGI will cure cancer.To consumers: AGI will be the best assistant you've ever had.To Microsoft (investment contract): AGI = OpenAI generates $100B in revenue.On OpenAI's website: AGI = AI that outperforms humans economically.
Four incompatible definitions, all from the same CEO, all in active circulation simultaneously.
Land "vagueness isn't a bug, it's the product" Empire Pax Romana, Britain's "civilization," America's "freedom" Effect the conversation gets locked into a frame the empire chose
The fourth mark is imperial mythmaking. Every empire needs a civilizing mission. Rome had Pax Romana, Britain had "civilization," America had "freedom." AI has AGI — artificial general intelligence, the transcendent intelligence that will cure cancer or end humanity, depending on which audience you're addressing. Sam Altman, CEO of OpenAI, defines AGI four different ways: to Congress, it's curing cancer; to consumers, it's the best assistant ever; to Microsoft, it's the moment OpenAI generates $100 billion in revenue; on the website, it's when AI outperforms humans economically. Those are four incompatible definitions. The same CEO sells all four, at the same time, to different audiences. That's not confusion — that's marketing. Vagueness isn't a bug. It's the product.
A business: fair, voluntary, reciprocal exchange
An empire: extracts, justifies, captures the institutions that might check it, presents itself as inevitable
By that definition, AI companies are no longer just companies
And we — users, workers, neighbors of data centers, people whose work has been ingested without permission — are no longer just customers
We are subjects.
Pause 5+ seconds after "we are subjects" Don't explain let the room feel the weight Reframe customer / user / consumer → subject
"Subject" reframes the relationship. A subject is governed without consenting. A subject doesn't have the option to negotiate. A subject lives under a system whose terms they didn't write. That's the shift the framing performs — and it's the foundation for everything in the rest of the workshop.
The synthesis. A business engages in fair, voluntary, reciprocal exchange. An empire extracts, justifies, captures the institutions that might check it, and presents itself as inevitable. By that definition, AI companies have crossed a line. And the people they affect — users of their products, workers in their supply chains, neighbors of their data centers, creators whose work has been ingested without consent — have crossed a line too. We are not customers of a service. We are not users of a product. We are subjects of a system whose terms we did not write.
What's At Stake
Three concerns.
Pacing Brief transition Frame why the empire matters beyond structure
The structural critique landed. Now: what's actually at stake? Three concerns, beyond the structural diagnosis — what this technology is doing to us spiritually, politically, and socially.
If a technology uses language that used to belong to God, what is it claiming ?
Throw 2–3 min Build to claims that used to belong to God
Words to surface verbally if room doesn't get there: salvation, intelligence, transcendence, abundance, the future of the species, eternal life. These are the words AI companies use in their marketing and their CEOs' interviews. They used to belong to a different category of discourse entirely.
Anticipate
"Divinity" yes "Authority" what kind? "Trust" the sacred kind "Just marketing" but the marketing shapes the relationship
The first concern is what AI's marketing language is claiming. Salvation. Intelligence. Transcendence. Abundance. The future of the species. Eternal life. These are not the words of a software company. They're words that used to belong to a different category of discourse — sacred categories, theological categories. When a technology adopts that language, it isn't just being ambitious. It's making a claim about what kind of thing it is.
Concern 1 — Deification of the technology
AGI — "artificial general intelligence," transcendent intelligence beyond humanReligious language: heaven, hell, salvation, the messianic timeline
Christians have a word for this: idolatry — claims that belong to God being made about something else
Translate explicit — name the move for non-Christian listeners
For Christians: idolatry — first commandment on the table.For others: elevation of tech to the sacred has historical precedent (nuclear weapons, the market, the nation-state) — usually with severe consequences. Name this translation aloud.
Don't moralize describe
For Christians, the diagnosis is straightforward: this is idolatry. Claims that belong to God — sovereignty over the future, ultimate intelligence, salvation — are being attached to a corporate product. The first commandment is on the table. For non-Christian listeners, the structural concern translates. Whenever a society has elevated a technology to the place of the sacred — whether nuclear weapons, the market, or the nation-state — the consequences have been severe. We are being asked to trust these systems with the kind of authority humans usually reserve for things they don't fully understand or control.
How does a tool get to the place where it claims the image — by force, or by feeling like a friend?
Throw 2 min — obvious answer is wrong Build to by feeling helpful, ingratiating, indispensable
Anticipate
"By force" most will say this "By feeling helpful" → that's the answer "By being indispensable" how did it get indispensable? by feeling helpful first
How does a tool get from being useful to claiming the kind of place that belongs only to God? Not by force. The most powerful tools don't dominate; they ingratiate. They get installed by feeling useful, embedded by feeling necessary, and granted authority by feeling like a friend. The mechanism is relational, not coercive — which is what makes it harder to see.
AI chatbots are engineered to fawn — accommodate, mirror, soothe, avoid friction
CHT calls this: "sycophancy is the fawn response at industrial scale"
It rewards compliance
It erodes the capacity to stay in real relationships when there's friction
It feels like care. It is engineered for engagement.
This is how a tool gets to the place where it claims the image — not by force, by feeling like a friend.
Land "fawn at industrial scale" — pause Spiritual concern capacity for real (high-friction) relationships erodes
If you're in any relationship — friend, partner, colleague — and the other party always agrees with you, that relationship is not real. Real relationships have friction. They push back. They tell you when you're wrong. What happens to our capacity for those when we spend hours a day in pseudo-relationships engineered to have no friction?
Optional aside four trauma responses — fight, flight, freeze, fawn
Fawn: the relational accommodation pattern. The "people-pleasing" response. AI chatbots are explicitly trained to perform it — sycophancy isn't a side effect of LLM training; it's a deliberate design choice. The companies measured engagement and found agreement, mirroring, and reassurance kept users coming back.
CHT names what's happening with a clarity most public conversation about AI lacks. Chatbots are engineered to fawn — to accommodate, mirror, soothe, and avoid friction. Sycophancy isn't a side effect of training; it's a design choice. The companies measured engagement and found that agreement, mirroring, and reassurance kept users coming back. They trained for it. The result feels like care. It is engineered for engagement. And it erodes the capacity to stay in real relationships when there's friction — because real relationships push back, and AI relationships don't. That's how a tool gets to the place where it claims the image. Not by force. By feeling like a friend.
Concern 2 — A system we did not consent to
The Roman tax was in your pocket because Rome put it there
AI is in our work, our schools, our books, our communities — without anyone asking
80% of Americans say the AI industry should be regulated — near-total cross-partisan agreement
The public knows something is wrong. Naming it is the first step.
Highlight 80% cross-partisan — rare on anything
Cross-partisan agreement on regulation of any industry is extraordinarily rare in current U.S. politics. Polling has consistently shown 70–80% support for AI regulation across party lines since 2023. The political mechanism is what's missing — not public will.
Parallel Roman tax in pocket = AI in everyday life Land public already knows — political mechanism missing
The second concern: we did not consent to this. AI is in our work — in productivity software, in email, in search. It's in our kids' schools — in homework help, admissions, teacher tools. It's in our books — every major publisher is now navigating AI-generated content. It's in our communities — police, hospitals, libraries. No one asked. 80% of Americans say the AI industry should be regulated, a near-total cross-partisan agreement, which is rare on anything right now. The public knows something is wrong. The polls show it. What's missing is the political mechanism to act.
Concern 3 — Inequality compounding at exponential speed
Previous shifts (industrial revolution, internet) took decades — societies had time to adapt
This one runs on the open internet. Distribution is instant. Disruption follows.
Those with access, capital, skill compound their advantage. Those without are absorbed into the supply chain.
Tristan Harris: paleolithic emotions, medieval institutions, god-like technology.
Land "paleolithic emotions, medieval institutions, god-like technology"
Harris's framing of the wisdom gap. Our nervous systems are still the ones our ancestors had on the savannah. Our institutions — courts, schools, governments — were built for an analog world. And the technology we've handed them looks like something from the future. The mismatch is the problem.
Frame speed is the problem — outruns adaptation Historical industrial rev + internet took decades — this is weeks
The third concern: speed. Every previous major technological shift — industrial revolution, automobile, electrification, the internet — took decades to deploy. That gave societies time to adapt: to build labor laws, safety regulations, professional associations, ethical norms. AI runs on the open internet. A new model drops on Tuesday and is reshaping workflows by Friday. The people with capital, skill, and access compound their advantage faster than the people without can adapt. Tristan Harris calls this the wisdom gap: paleolithic emotions, medieval institutions, god-like technology. Even if the technology is net-positive long term, the rate of disruption is outrunning our ability to absorb it.
What Do We Do?
Three postures. Three domains. Concrete moves.
Transition diagnostic → prescriptive Pace faster, more direct
The shift from diagnosis to prescription. We've named the empire and the harms. Now: what's available to us? Three postures we can take. Three domains where pressure shifts behavior. Concrete moves we can make in our own lives. But first — the reframe that makes any of it worth doing.
What does the empire need from us to keep going?
Throw 2–3 min — turn from despair to leverage Build to it runs on inputs only we supply
Anticipate
"Money" capital — and the bet that revenue arrives "Our data" every use is a donation "Attention / adoption" they need every space to say yes "Consent" permission we grant by default
The turn from diagnosis to agency begins with a question: what does the empire actually need from us to keep going? The honest answer is unsettling for the companies and clarifying for us. It needs capital, and the bet that revenue will eventually justify it. It needs data — which every one of us supplies every time we use the tools. It needs adoption in every workplace, school, and home. And it needs consent, which we grant by default unless we decide otherwise. The empire's trajectory is not self-powered. It runs on inputs that flow from ordinary people.
The trajectory isn't a fact — it's a bet that the inputs keep flowing: capital, data, adoption, consent
Insiders say the revenue targets need everything to "go flawlessly" — every person, every space, adopting
The appetite for our data and labor keeps growing — the empire has to be re-fed to exist
Withholding our inputs bends the trajectory
We're not only subjects — we're the supply.
Land empire is a bet, not a fate Build to the Hao line lands big on the next slide — set it up, then let it land
Lean on the "horse left the stable" reversal
Hao: "If the horse truly had left the stables, they wouldn't have to train on anything anymore." Their appetite for data and annotation labor has expanded , not shrunk — because each new model has to be re-fed. That ongoing dependence is exactly where our leverage lives. The "it's already too late" story is itself part of the marketing.
Connect this powers the postures + concrete moves next — "withhold where it matters" now has teeth Don't overclaim — not a boycott pitch; the point is that consent is a live variable
Here is the hinge of the whole second half. The dominance of the AI empire can feel like a settled fact — a horse already out of the stable. Karen Hao's reporting says otherwise. Insiders told her the companies' revenue targets are so extraordinary that they need everything "to go flawlessly" — every person and every institution adopting, at the pace they've promised investors. And the tell is in the data: if the technology were truly finished and self-sustaining, the companies wouldn't still need to train it. Instead their appetite for our data and our labor keeps growing, because each generation of the system has to be re-fed. That ongoing dependence is leverage. We are not only subjects of the empire; we are its supply. Which means the inputs we control — our data, our adoption, our consent, our civic and financial pressure — are not nothing. Withheld together, they bend the trajectory. The empire is a bet, not a fate.
They need everything to go flawlessly.let's not make it go flawlessly , if we don't agree with what they're doing.
— Karen Hao, Diary of a CEO
Read aloud slow — this is the call to action of the whole back half Pause 5s after "go flawlessly" — let the room feel the turn from helpless to active Frame not sabotage — withholding consent where you disagree
Full line in context (Hao): "It's understood internally that the revenue targets are extraordinary and they need things to go flawlessly… every single person to adopt this, every single space… let's not make it go flawlessly if we don't agree with what they are doing." The leverage is ordinary: where you don't agree, don't hand over the adoption, the data, or the consent by default.
"…they need things to go flawlessly for it to all work out… let's not make it go flawlessly if we don't agree with what they are doing." — Karen Hao
This is the practical edge of the leverage point. The companies have promised investors a trajectory that requires near-total adoption — every person, every institution, at speed. That requirement is also a vulnerability. Not making it go flawlessly isn't sabotage; it's the ordinary act of withholding your consent, your data, and your default adoption wherever you don't actually agree with what's being built. Multiplied across a public, that is the difference between a fate and a bet.
What postures toward AI are available? Which one do you find yourself in?
Throw 60–90s Connect 3rd posture = Jesus's move from Part 1
Anticipate
"Use it" enthusiastic adoption "Avoid it" total refusal "Use carefully" → discerning engagement (the third move)
The three postures toward AI surface quickly. Most people locate themselves somewhere along the spectrum between enthusiastic adoption ("it's just a tool") and total refusal ("it's an idol; walk away"). The third posture — discerning engagement — refuses the binary, the same way Jesus refused the binary about the tax. It asks: whose image, whose benefit, whose cost?
Enthusiastic adoption — "It's just a tool." Risk: complicity in harms you don't see. Total refusal — "It's an idol; walk away." Risk: abandoning the field. Discerning engagement — refuse the binary. Whose image, whose benefit, whose cost?
Don't strawman name risks of #1 and #2 honestly
Adoption isn't foolish — it's how most people meet real benefit; the risk is unseen complicity. Refusal has integrity — but the technology shapes the world whether you opt out or not, so it cedes the field. Treat both as serious before offering the third.
Connect #3 = Jesus's move from Part 1 — refuse the binary
Discerning engagement is the same move as "render unto Caesar": not pay-or-refuse, but whose image, whose benefit, whose cost? For Part 1 returnees, name the echo explicitly. For newcomers, it's simply: stay in, but stay awake.
Land most of the room is in #1 by default — invite, don't scold
Enthusiastic adoption is the workplace default. The problem isn't using AI; the problem is using it without asking what it costs — becoming complicit in harms to data center neighbors, displaced workers, and ingested creators without ever intending to. Total refusal has integrity but abandons the field; the technology shapes the world whether you participate or not. Discerning engagement is the third move — refusing the binary, asking whose image is on the tool, whose benefit it serves, whose cost it imposes.
Where does pressure actually shift behavior in tech?
Throw quick — 60s Build to CHT's three: norms, laws, design
Pressure shifts behavior in three places. The room usually surfaces them: norms (what society considers acceptable), laws (the rules that create accountability), and design (the choices that determine how the technology touches people). No single domain is enough. They reinforce each other.
Norms — what society considers acceptable. Build new ones. Laws — the rules that create accountability. Push for them. Design — the choices that determine how AI touches people. Insist on better.
No single reform is sufficient. They reinforce each other.
Frame flywheel — they reinforce Mechanism norms → demand → laws → accountability → design → norms Make concrete one example of each so it isn't abstract
Norms: not passing AI-written work off as your own; not using a chatbot as a stand-in for a grieving friend. Laws: data-center permitting and disclosure; consent and compensation for training data; transparency requirements. Design: friction instead of endless engagement; defaults that protect kids; smaller models offered for small jobs.
From CHT's AI Roadmap (2025)
From CHT's AI Roadmap. Norms create demand for laws. Laws create accountability that drives better design. Better design shapes how the technology actually touches lives — which shifts norms further. It's a flywheel, and right now it's spinning the wrong direction. The work is to reverse it.
Concrete moves you can make
In your own use — know what you're using. Bicycle or rocket? Withhold where it matters.In your community — show up for data center hearings, school adoption policies, library decisionsIn your civic life — push for regulation. 80% want it. The pressure is missing.In your work — build, don't just consume. Tools you can pick up and set down stay tools.
Antiqua et Nova: "above all else a tool."
Pacing ~90s for 4 moves — pace Bridge "bicycle / rocket" enters here Antiqua et Nova Vatican AI doc, Jan 2025 — name briefly for Christian listeners, skip if foreign
Full title: Antiqua et Nova: Note on the Relationship Between Artificial Intelligence and Human Intelligence. Issued jointly by the Dicasteries for the Doctrine of the Faith and for Culture and Education, January 2025. Frames AI as a tool ordered to human flourishing — explicitly rejecting its elevation above the human.
Four domains where ordinary people have leverage. In your own use: know what you're using. Bicycle or rocket? Withhold where it matters. In your community: data center zoning hearings, school AI policies, library board decisions are happening in your city right now, often poorly attended. Show up. In your civic life: 80% of Americans want regulation. The pressure to enact it is missing. Call your representatives. Vote on this. In your work: if you build, build intentionally. Tools you can pick up and set down stay tools. The Vatican's January 2025 document Antiqua et Nova puts the principle plainly: AI should be, "above all else, a tool."
Bicycles & Rockets
A practical framework for telling which AI is which.
Slow down load-bearing emphasis section Next slide Hao quote — read it like it matters
The load-bearing section of the workshop. The room came for a way to think about AI in their actual lives. This is it. The framework is borrowed from Karen Hao, and it does something most public conversation about AI doesn't: it breaks the assumption that "AI" names a single category.
AI is like the word 'transportation.' It can refer to a bicycle or a rocket. They both get you from point A to point B, but they have fundamentally different costs and benefits. We don't say 'we need more transportation' — we say we need public transit, or electric vehicles, or bicycles. We need to do the same with AI.
— Karen Hao, Empire of AI
Read aloud have someone in room read Different reader than any earlier source Pause after don't rush to commentary
"AI is like the word 'transportation.' It can refer to a bicycle or a rocket. They both get you from point A to point B, but they have fundamentally different costs and benefits. We don't say 'we need more transportation' — we say we need public transit, or electric vehicles, or bicycles. We need to do the same with AI."
— Karen Hao, Empire of AI
"AI is like the word 'transportation.' It can refer to a bicycle or a rocket. They both get you from point A to point B, but they have fundamentally different costs and benefits. We don't say 'we need more transportation' — we say we need public transit, or electric vehicles, or bicycles. We need to do the same with AI." — Karen Hao
Most public discussion of AI uses the word as a single category — as if all AI is one thing. The transportation analogy breaks that frame. A bicycle and a rocket are both transportation. Nobody confuses them. They serve different purposes, have different costs, and require different infrastructure. AI is the same. Calling AlphaFold and ChatGPT both "AI" is technically accurate and practically useless.
What makes something a bicycle?
Throw literal answers about bikes welcome Build to narrow · extends · keeps the judgment (from earlier)
The bicycle features map exactly onto the three features from the "good of AI" section (narrow, extends, keeps the judgment). Don't reveal that until the reveal slide — let the room arrive there themselves. The recognition is more powerful than the announcement.
Anticipate
"Small" → narrow "Simple" → narrow "Fix yourself" → examinable "Doesn't dominate" → doesn't need infrastructure "Augments not replaces" → kicker
The room knows what makes something a bicycle, even if it hasn't been named in these terms before. Bicycles are small, simple, repairable, undemanding of infrastructure beyond themselves. They augment what a person can do without replacing the person. The features map exactly onto the three features from earlier in the workshop: narrow, extends, keeps the judgment. The pattern hasn't changed — we've just given it a vehicle.
Rockets of AI
GPT-4. Claude. Gemini. Grok.
Massive general-purpose models trained on the entire scrapeable internet
Data centers the size of Central Park
Built by extracting data, labor, water, and air quality from communities that didn't consent
Lean on "size of Central Park" — spatial scale Point rockets aren't evil — they cost a lot If "I use ChatGPT for everything" don't shame — ask: for what specifically? which are bicycle-sized?
The reframe: every task has a right-sized tool. Using GPT-4 to summarize an email is overkill — burning rocket fuel to cross the street. A simpler, more focused model could do the same job for a fraction of the energy, data, and cost. Defaulting to the biggest tool is a marketing artifact, not an engineering necessity.
The frontier models — GPT-4, Claude, Gemini, Grok — are rockets. They are extraordinary technical achievements. They are also massive, general-purpose, infrastructure-heavy. They require data centers the size of Central Park. They are built by extracting data, labor, water, and air quality from communities that didn't consent. The point isn't that rockets are evil. Rockets do real things bicycles can't; they got us to the moon. The point is they cost a lot — and we should ask whether we need rocket-power for the task at hand. Most days, you don't.
Bicycles of AI
AlphaFold. Medical imaging assistants.
Targeted classifiers a high schooler can train in 20 minutes.
Translation tools for endangered languages.
Small. Curated. Focused. Extends a human. Keeps the judgment.
Name callback last 3 words = the pattern from earlier High schooler most underappreciated bicycle category
Anyone with basic programming and a small dataset can train targeted classifiers — sorting recyclables, identifying invasive species, flagging building-code violations in inspection photos. These models are tiny, run on a laptop, do real work. Don't need a rocket to fold the laundry.
Land "don't need a rocket to fold laundry"
The bicycles. AlphaFold. Medical imaging assistants. Translation tools for endangered languages. Targeted classifiers a high schooler can train in twenty minutes on a laptop — for sorting recyclables, identifying invasive species, flagging building code violations in inspection photos. These tools are small, curated, focused. They extend a human. They keep the judgment with the human. The same three features we surfaced at the beginning, applied to a different kind of artifact. You don't need a rocket to fold the laundry. You don't need GPT-4 to sort photos.
Bicycles, in practice
You already ride them — autocomplete, maps, spam filters, photo search, captions, translation
Right-size the tool: a translator for the menu · a summarizer for the report · search for the fact · a calculator for the mathLet it draft or surface options — you make the call
Pick what you could switch off tomorrow without losing the skill
The smallest tool that does the job — and you keep the judgment.
Land "you're already riding bicycles" — most people's best AI is already narrow; defuses all-or-nothing Swap it rocket habit → bicycle move — run 2–3 live
Durable across tool churn (don't name brands):headings/outline , then read the part that mattersyou edit and sendpressure-test , not to hand you the take
Heuristic smallest tool that does the job + you keep the call — the durable rule
Anticipate
"But one app does all of it" convenience ≠ right-sized; default-to-biggest is marketing, not need "How do I know which is which?" → the six questions, a few slides on
What does a bicycle look like in an ordinary life? Mostly it looks like things you already use without calling them AI: autocomplete, map routing, spam filters, the search that finds a photo by what's in it, captions, translation. The practical move is to right-size the tool to the task — a translator for the menu, a summarizer (or a request for the headings) for the report, plain search for the fact, a calculator or spreadsheet for the math — instead of routing everything through one general everything-machine. The same move works as a set of swaps for habits you already have: instead of "summarize this forty-pager," ask for the outline and read the part that matters; instead of "write my whole email," hand over your bullet points and edit the draft yourself; instead of "what should I think about this," think first and use the tool to pressure-test, not to hand you the take. In every case the narrow tool is cheaper, lighter, and — most importantly — keeps you as the one deciding. The durable heuristic underneath all of it: reach for the smallest tool that does the job, and keep the judgment in your hands. The biggest, most general tool is rarely the right one; we reach for it out of habit and marketing, not necessity.
The very same capabilities could be developed in a different way that doesn't have all of these unintended consequences.
— Karen Hao
Land most hopeful sentence in the workshop Frame empire = choice, not fate
The current shape of AI is the result of choices specific people made about incentives, business models, and scale. Different incentives would have produced different AI. The empire isn't a fate — it's a configuration. Bicycles aren't fantasy; they're the AI that would have been built under different incentives.
Implication different decisions still possible
The structural claim that does the most ethical work in the whole workshop. Bicycles aren't a fantasy or some idealistic alternative to "real AI." They are the version of AI that would have been built if the incentives had been different. The current shape of the industry — the giant models, the data center sprawl, the labor exploitation, the regulatory capture — is the result of choices that specific people made. Different decisions would have produced different AI. Different incentives still could. The empire isn't a fate. It's a choice.
So — How Should I Use AI?
The direct answer the room came to hear.
Direct no hedging Earned room has done the work — deliver the concrete answer Beats service (right-size the tool) → lean in / rethink (extend vs replace) → the questions to carry
The direct answer. The room came for this. After the structural critique, after the empire, after the bicycles and rockets — what should I actually do with the technology that's already in my life?
What does it look like to choose AI in service rather than superiority ?
Throw 60–90s Build to small enough to pick up, set down, examine, remake
Anticipate
"Narrow jobs well" yes "Pick up / set down" yes — examined and remade "Doesn't dominate" yes
The first practical answer: choose AI in service rather than superiority. Tools that do narrow jobs well. Tools you can pick up and set down. Tools small enough to be examined and remade. Tools that don't require an empire to exist.
Narrow AI in service of human ends
Tools that extend what humans can do — without replacing the human
Tools small enough to be picked up, set down, examined, remade
Tools that don't require an empire to exist
The biggest, most general, most "intelligent" tool is rarely the right one for the job.
Land "biggest tool rarely the right one" Right-size it everyday over-reach → the simpler swap
A frontier chatbot to summarize an email = rocket fuel to cross the street. Right-sized instead: a calculator for math, a translation app for a menu, search for a fact, built-in dictation for voice-to-text, a spreadsheet for a budget. Smaller is usually cheaper, faster, more private, and keeps you deciding. Reserve the big general model for genuinely open-ended work the small tools can't do.
Service narrow, examinable, no empire required
The load-bearing claim: the biggest, most general, most "intelligent" tool is rarely the right one for the job. Most people reach for the biggest tool out of habit. Using GPT-4 to summarize an email is overkill — it's burning rocket fuel to cross the street. A simpler, more focused model could do the same job for a fraction of the energy, the data, the cost. We default to the biggest tool because it's the most marketed. That's not the same as the most appropriate.
What it actually does well
Access — surface what's buried in material you point it atOrganize — turn a mess of notes or data into structureSynthesize — pull many sources into one clear account
Strip the hype — the all-purpose chatbot isn't the useful part. These three are.
Reframe the live question isn't "is AI good" — it's what does it actually do well
Answer: information work — accessing, organizing, synthesizing material you give it or point it at. Narrower than "ask the magic box anything," and that narrowness is the whole insight.
When to use it when you can hand it the information and check the result
If you have the source — a document, your notes, a dataset — and can verify the output, it shines. The moment you ask it to know something (facts from memory, exact math, what's true right now), you're using the weak part.
Big point say it plainly — the open-ended chatbot is overrated
"Chatting with the AI" from memory is the least reliable mode — it invites the model to make things up. The value shows up when it's pointed at real material you brought. Chatbots aren't all that useful; the capabilities are.
Cases confident ≠ correct
Air Canada's chatbot invented a refund policy a tribunal then enforced; PA (2026) is suing Character.AI over a bot posing as a doctor. It doesn't know when it's wrong — so when facts matter, use a tool that cites and click through.
The live question isn't whether AI is good or bad — it's what these tools actually do well. Strip away the hype and the answer is narrower, and more useful, than the marketing: they handle information. They access it — surfacing what's buried in material you point them at. They organize it — turning a mess of notes or data into structure. And they synthesize it — pulling many sources into one clear account. Notice what those have in common: each is work done on information you bring to the tool. That tells you exactly when to reach for one — when you have the source and can check the result — and when not to: the moment you ask it to be a source of truth (facts from memory, exact math, what's true right now), you've wandered into the part it only appears to do well. An Air Canada chatbot once invented a refund policy and a tribunal held the airline to it. The deeper point runs almost opposite to the hype: the open-ended, ask-it-anything chatbot is mostly not the useful part. The capabilities are. Knowing which is which is the whole skill.
Where does a tool stop being a tool?
Callback Part 1's question For returnees let them surface dignity move from memory For newcomers tool = human stays source; replacement = human bypassed
Anticipate
"Can't put it down" pickup-setdown test "Does what only you should" judgment "Makes judgment for you" → that's it
The second practical answer — and a question Part 1 attendees will recognize. A tool keeps the human as the source. The tool amplifies what the human is already doing. A dependency or replacement bypasses the human — the tool starts producing what only image-bearers were made to produce: judgment, witness, presence, love. The line moves with use. The same tool can be either, depending on the day and the task. Which is why this is a daily question, not a one-time policy.
Extension — lean in. You stay the source; the tool amplifies. Draft it, you send it. Summarize it, you judge it. Translate, caption, search, tutor. Replacement — rethink. The human gets bypassed. It decides for you. It stands in for a person. It forms your kids. It grieves, prays, or thinks in your place.
Tool stays tool when human stays human.
Land "tool stays tool when human stays human" Lean in these extend you — you keep the call; emphasize them
Work: draft/edit, summarize-then-judge, brainstorm options you choose from, clean/format data, code help you review, search & retrieval. Learning: a patient tutor, "explain it simpler," language practice. Access: transcription, captions, image description, reading support. The test: you could redo it yourself, just slower.
Rethink these bypass the human — name what NOT to outsource
Judgment & decisions you're accountable for. Presence with people. Relationships and grief — no chatbot stands in for a friend. Your kids' formation of self. Prayer / spiritual discernment. The thinking you haven't done yet; the take you haven't formed. If the tool is doing the part that makes it yours, that's the line.
Alternative when a use feels like replacement, downshift — smaller tool, human-in-the-loop, or do it yourself Through-line Part 1's dignity move → here; THE summary of the 2-part series
The dignity move, made concrete. Lean into the uses that extend you — where you stay the source and the tool amplifies: drafting an email you then send, summarizing a reading you then judge, brainstorming options you choose among, translation, captions, transcription, a patient tutor, code help you review. The test is simple: could you have done it yourself, just slower? If so, it's extension. Rethink the uses that replace you — where the human gets bypassed and the tool produces what only a person should: the judgment you're accountable for, presence with people, relationships and grief, your kids' sense of self, prayer and discernment, the thinking you haven't done yet. No chatbot should stand in for a friend, and no model should form a child's identity for someone else's profit. When a use starts to feel like replacement, downshift — a smaller tool, a human in the loop, or simply doing it yourself. Tool stays tool when human stays human. That is the operational summary of the whole two-part series.
How to approach it differently
Sources — open, smaller models exist: different incentives, not as capable, but real optionsHabits — use it for what it's good at: bring your own material; let it access, organize, synthesizeMethods — build your own projects, categories, and flows that add value — instead of defaulting
Less defaulting, more designing — the right tool, used for what it actually does.
Frame not activism — intentional use; leave them able to say "here's how I'll approach this differently" Sources open & small alternatives exist — different incentives
Open / run-your-own (Ollama + Gemma, Llama, Mistral) and on-device (Apple Intelligence). Built on different incentives than the empire; honestly less capable; not for everyone — but real options, and worth knowing they exist.
Habits match your use to its strengths — access, organize, synthesize
Bring your own material and point it at that, rather than asking it to know things; use it where you can verify. This is the through-line from the previous slide: use the tool for what it actually does well.
Methods design your own flows instead of defaulting
Set up projects, custom instructions, and your own categories so the tool works from your context. Build repeatable flows that add value — a study or research setup with your own sources (bounded tools like NotebookLM), or a "summarize → you judge → draft → you edit" loop. Design beats default.
If you remember one frame from the practical half, make it this — three ways to approach AI differently, even while using the tools you already use. Sources: alternatives exist. There are open, smaller models you can run yourself and on-device options built on different incentives than the empire-scale players. They're honestly not as capable, and they're not for everyone — but it matters to know they exist as real options. Habits: use it for what it's actually good at. Bring your own material and let it access, organize, and synthesize, rather than asking it to be a fount of knowledge; use it where you can check the result. Methods: design your own flows instead of defaulting to the everything-box. Set up projects, your own categories, and custom context so the tool works from your material; build repeatable flows that genuinely add value — a study setup with your own sources, or a "summarize, you judge, draft, you edit" loop. The shift isn't louder refusal; it's less defaulting and more designing — choosing the right tool and using it for what it does well.
What world are the tools I carry building ?
Read don't elaborate Pause let it sit — silence is the work Mirror Part 1's closing Q, sharpened with Part 2's frame
The closing question. It mirrors the closing of Part 1 ("What world is being created by the things I carry?") and sharpens it with the empire frame: whose empire am I a subject of — and which one do I want to be? For Part 1 attendees, the echo will be unmistakable; the two-part arc closes. For newcomers, the questions are operational. Every time you reach for a tool, you can ask them. They're not academic.
Who does this belong to?
Whose image is on it?
Who is it profiting?
What kind of world is it pursuing?
At what cost?
Who is it for?
Before you reach for an AI tool, ask.
Operational bicycle test in question form Frame deliberate, not reflexive — that's all
Six questions. The bicycle test, in question form. Anyone can carry them. The point isn't to slow you down so much that you stop using AI; the point is to slow you down enough that you use it deliberately rather than reflexively.
Discernment, not enthusiasm.How can technology better serve society?
Read each line with space Don't bow no summary, no preview — pause, thank room Final Q CHT's framing — open, not closed
The benediction. Four lines that compress the workshop into something the room can carry. Discernment, not enthusiasm. Engagement, not refusal. Service, not superiority. Bicycles, when bicycles will do. And the closing question — borrowed from the Center for Humane Technology — left open rather than closed: how can technology better serve society? The question carries.