The AI-Governance Worldview Camps: Different Pawls, Same Ratchet

A research position. Measured, sourced, present-all-sides. Every claim about a real person or camp carries a verified source; characterizations are attributed to whoever made them, never adopted as fact; no unfalsifiable collective-actor framing is asserted. This is a public research document (evilrobots.lol/research) and a source spine for Book 3.

Perishability note. The 2024-2026 claims below are dated and flagged. The single most perishable item is the encyclical Magnifica Humanitas (signed 15 May 2026), weeks old at first drafting and re-verified 2026-06-06 against the vatican.va primary text. Re-verify all 2025-2026 dates, FLOP thresholds, EU phase-in dates, and p(doom) figures before any print use.

Part 1 — The Map of Camps

There is not one AI debate. There are at least seven distinguishable worldviews, and they answer different questions. Three are about the future of the technology (accelerate it, fear it, or document what it does now). Three are about what should be done (win the contest, refuse the premise, or manage the thing). And one is about something the other six have no column for: what, exactly, is at stake. Each camp below gets its central thesis, its named figures with verified sources, who it opposes, and the outcome it bets on or fears.

A discipline note that runs through the whole document: every camp here is internally divided, and flattening any one of them into a monolith is the most common analytic error in this debate. The fault lines inside each camp are named where they matter.

Camp 1 — Accelerationists / Techno-Optimists (e/acc)

Thesis. Technological progress is the master variable of human welfare, and it is currently being suppressed rather than over-accelerated; intelligence and markets compound each other; markets, not regulators, produce safety; and on AI specifically, delay has a body count. The signature line, from Marc Andreessen’s “The Techno-Optimist Manifesto” (a16z, 16 Oct 2023): “We believe any deceleration of AI will cost lives. Deaths that were preventable by the AI that was prevented from existing is a form of murder.” The manifesto repeats “We believe” 113 times and lists “existential risk,” “tech ethics,” “risk management,” “the Precautionary Principle,” and “de-growth” among its named “enemies.” (a16z — The Techno-Optimist Manifesto)

The e/acc subspecies adds a thermodynamic frame. The founding Substack essay — authored under the pseudonyms “Beff Jezos” and “@bayeslord,” and therefore pseudonym-sourced — grounds the worldview in physics (“the universe exponentially favors futures where matter has adapted itself to capture more free energy”) and closes with the literal slogan “A C C E L E R A T E.” (Beff Jezos / @bayeslord — Notes on e/acc principles and tenets) Forbes (Emily Baker-White, 1 Dec 2023) unmasked “Beff Jezos” as physicist Guillaume Verdon, who confirmed it. On the record under his own name (Lex Fridman #407), Verdon’s arguments are narrower and more defensible than the meme: that markets will “positively select for AIs that are more reliable, more safe,” that heavy-handed regulation “written by the incumbents… serves them to achieve regulatory capture,” and that he declines to state a p(doom) because the system is “far too chaotic for anybody to have an accurate estimate.” (Lex Fridman Podcast #407 — Guillaume Verdon transcript)

Key figures. Marc Andreessen (a16z, signed manifesto); Guillaume Verdon / “Beff Jezos” (Extropic; on the record post-unmasking); Garry Tan (Y Combinator — notably more moderate than the label: “regulation is likely necessary,” fears monopoly not doom, TechCrunch — Garry Tan on AI and regulation); Patrick Collison and Tyler Cowen (Progress Studies — the empirical, institution-reform wing, distinct from e/acc’s meme-and-physics register, Patrick Collison — Progress Studies).

Internal fault lines. Four sub-factions, not one: e/acc (meme + thermodynamics, anti-regulation); libertarian techno-optimism (the manifesto); “little tech” pragmatists (Tan — explicitly pro-NIST, pro-some-regulation); and Progress Studies / abundance liberals, who generally reject the e/acc aesthetic as unserious. The manifesto’s listing of Filippo Marinetti (later co-author of the Fascist Manifesto) and Nick Land among 56 “patron saints” is in the text and is an embarrassment the abundance-liberal wing does not own (Wikipedia — Techno-Optimist Manifesto).

Who they oppose. The doomers/x-risk camp, the EA safety wing, and regulation advocates. Their real nightmare, underrated by critics, is not rogue AI. It is regulatory capture and monopoly: a few incumbent labs using safety regulation as a moat. Verdon’s “regulatory capture” line and Tan’s fear of “rent extraction… a world I don’t want to live in” are the same fear.

Outcome they bet on / fear. Bet: a decentralized abundance trajectory, Verdon’s “stable, adversarial equilibrium of powers.” Fear: a licensing regime only incumbents can satisfy, justified by speculative doom.

Camp 2 — Doomers / Existential-Risk / AI-Safety

Thesis. Not that today’s chatbots are dangerous, but a conditional about the future: if a machine substantially smarter than humans across all domains is built before the alignment problem is solved, the default outcome is human disempowerment or extinction. And we are on track to build it before we know how to control it. The argument rests on two theses from Nick Bostrom’s Superintelligence (2014) (Oxford University Press — Superintelligence): the orthogonality thesis (intelligence and final goals are independent — smarter does not mean nicer) and instrumental convergence (almost any goal generates the same sub-goals: self-preservation, resource acquisition, resistance to shutdown). (Bostrom, The Superintelligent Will, nickbostrom.com PDF; Wikipedia — Instrumental convergence)

Key figures. Eliezer Yudkowsky (MIRI — the hard-doom pole; in TIME, 29 Mar 2023, “the most likely result of building a superhumanly smart AI… is that literally everyone on Earth will die,” demanding an indefinite global moratorium enforced if necessary by airstrikes on rogue datacenters — verbatim text via MIRI’s repost, MIRI / TIME — Pausing AI Developments Isn’t Enough); Nick Bostrom (the philosophical foundation); the CAIS “Statement on AI Risk,” 30 May 2023 — one sentence, “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war,” signed by Geoffrey Hinton, Yoshua Bengio, Demis Hassabis (Google DeepMind), Sam Altman (OpenAI), and Dario Amodei (Anthropic) (CAIS — Statement on AI Risk); the FLI “Pause” letter, 22 Mar 2023, calling for a 6-month pause on training beyond GPT-4 (Future of Life Institute — Pause Giant AI Experiments); Geoffrey Hinton (resigned Google May 2023 to speak freely; states a “10 to 20” percent chance of human extinction within ~30 years, which he calls “just gut” — cite as what he said, with date, never as a measured quantity); Yoshua Bengio (the only figure anchoring both the FLI letter and the CAIS statement; cited p(doom) ~20%).

Internal fault lines. At least four mutually critical wings. Hard-doom/MIRI (p(doom) >90%; alignment effectively unsolvable on current timelines; stop). Mainstream lab safety (Amodei, Christiano; p(doom) ~10-50%; the problem is hard but tractable, you must build to study). EA/longtermist policy (Open Phil, CAIS, FLI; compute governance and evals over p(doom) numerology). The defectors (Hinton, Bengio; academic elders, no equity stake, stop short of “shut it all down”). The sharpest seam is pause vs. align-and-proceed: Yudkowsky publicly refused to sign the FLI pause letter for asking too little; the lab CEOs who signed the CAIS extinction statement did not endorse a development pause. The p(doom) spread itself — Yampolskiy 99.9%, Yudkowsky >95%, Amodei 10-25%, Bengio 20%, Hinton 10-20%, LeCun <0.01% — reveals no shared evidentiary basis, only priors (Wikipedia — P(doom)).

Who they oppose. The accelerationists; the present-harms camp (which opposes them from the left); and the “it’s hype / not close” skeptics (Yann LeCun, who calls the takeover scenario “preposterous,” TechCrunch — Yann LeCun on AI existential threat; Gary Marcus, who sees brittle pattern-matchers, not minds — note LeCun and Marcus disagree sharply with each other on whether deep learning is near AGI).

Outcome they bet on / fear. Fear: an intelligence explosion / fast takeoff in which a system crosses to superhuman before alignment is verified. Bet: buy time — pause or slow frontier scaling, fund alignment, build compute governance modeled on nuclear non-proliferation (the explicit Hendrycks analogy).

Camp 3 — Present-Harms / AI-Ethics / Algorithmic Justice

Thesis. AI’s harms are present and material, not speculative and future. The cleanest verbatim statement, from Alex Hanna and Emily M. Bender in Scientific American (Aug 2023): “Wrongful Arrests, an expanding surveillance dragnet, defamation and deepfake pornography are all existing dangers… These issues, and not the imagined potential to wipe out humanity, are the real threat of artificial intelligence.” (Scientific American — We Need to Focus on AI’s Real Harms) Five harm pillars: bias/discrimination, labor exploitation, surveillance, environmental/financial cost, and concentration of power. The second-order claim — the one that defines the camp’s posture — is that existential-risk framing functions to launder present harms by diverting attention and regulatory bandwidth toward speculative futures only the labs can claim to manage. (The “launders/distracts” framing is the camp’s characterization, not asserted here as fact; a peer-reviewed study disputes the mechanism — see Part 5.)

Key figures. Timnit Gebru (left/fired from Google over the Stochastic Parrots paper — contested: Google says it accepted her resignation, Gebru says she was fired; both framings stated; founded DAIR, Wikipedia — Timnit Gebru); Emily M. Bender (the octopus thought experiment; co-author of Stochastic Parrots); Margaret Mitchell (co-author, fired Feb 2021, now Hugging Face); the Stochastic Parrots paper (Bender, Gebru, McMillan-Major, “Shmargaret Shmitchell,” FAccT ‘21: ACM — On the Dangers of Stochastic Parrots (DOI 10.1145/3442188.3445922)); Joy Buolamwini (Algorithmic Justice League; Gender Shades, 2018, with verified error rates — lighter males <1%, darker females ~35%, Wikipedia — Algorithmic Justice League — which contributed to IBM exiting and Amazon/Microsoft pausing police facial-recognition sales in 2020); Abeba Birhane (dataset audits; 80 Million Tiny Images withdrawn after her work); Émile Torres (TESCREAL paper with Gebru, Wikipedia — TESCREAL).

Internal fault lines. Bias/audit empiricists (Buolamwini, Birhane — reformist; prove the harm, force the fix; can work inside the system) vs. anti-hype linguists (Bender, Hanna — more willing to say the whole “AI” framing is rotten) vs. TESCREAL critics (Gebru + Torres — the most confrontational, framing the opposing ideology as quasi-eugenic, the most contested move even inside the camp). A genuine tension: auditing a system to improve its accuracy implicitly accepts it is real and improvable, which sits awkwardly against “this should not be deployed at all.” Improving facial-recognition accuracy can entrench the surveillance the camp opposes.

Who they oppose. Accelerationists (who externalize costs onto exactly the surveilled, the wrongfully arrested, the underpaid annotators) and doomers/longtermists (the more developed front — see TESCREAL in Part 2).

Outcome they bet on / fear. Fear: “AI safety” becoming a synonym for speculative-future safety controlled by incumbents while present harms compound under cover. Bet: grounded, empirical harm documentation forces remediation (the Gender Shades → IBM/Amazon retreat is the proof-of-concept), and independent institutions (DAIR, AJL, Mozilla) can do science free of Big Tech incentives.

Camp 4 — National-Security Realists (Schmidt / NSCAI)

Thesis. AI is a general-purpose strategic technology on the scale of electricity or nuclear fission; leadership determines 21st-century power; the US is in a contest with China it is currently positioned to lose absent deliberate national mobilization. The NSCAI Final Report (March 2021, chaired by Eric Schmidt, vice-chaired by Robert Work): “America is not prepared to defend or compete in the AI era”; “The AI competition is also a values competition”; the US should stay “at least two generations ahead of China” in microelectronics. (NSCAI Final Report (March 2021)) In April 2025 House testimony Schmidt restated it: “We are in a race, and we must win.” (Congress.gov — Schmidt House testimony, 9 Apr 2025 (PDF))

Internal fault lines. Schmidt is not a race-to-AGI-at-any-cost maximalist. In “Superintelligence Strategy” (March 2025, with Alexandr Wang and Dan Hendrycks) he argues against a government Manhattan Project for AGI, which “assumes that rivals will acquiesce to an enduring imbalance or omnicide rather than move to prevent it,” and proposes Mutual Assured AI Malfunction (MAIM) — deterrence over a unilateral race (Superintelligence Strategy (nationalsecurity.ai)). So the camp’s deepest seam is race-to-win vs. deter-the-race. It also splits on open vs. closed weights (Schmidt calls Western open-source leadership “vital”) and on how hard to decouple from China.

Who they oppose. Refusers (“unilateral restraint is strategically suicidal — an authoritarian competitor builds anyway and exports its model”); EU pragmatists (“we regulated ourselves out of the race” — the Draghi critique); and, internally, AGI-Manhattan-Project hawks.

Outcome they bet on / fear. Bet: coordinated US public-private mobilization keeps the US ahead in chips and frontier models. Fear: China reaches its stated 2030 leadership goal and writes the global standards. Or the race itself becomes the danger (the MAIM concern). Note for the “same ratchet” thesis: the NSCAI report itself cites China’s surveillance of the Uyghurs as the cautionary precedent. The realist document concedes the apparatus is the destination and argues only over whose apparatus.

Camp 5 — Refusers (Whittaker)

Thesis. “AI” as currently built is not a neutral tool but the latest expression of a surveillance-and-scale business model that concentrates data, compute, and decision-making in a handful of corporations. “Manage the risk” and “win the race” both presuppose the thing should be built. The refuser position is that some of this should not be built, and some existing power should be dismantled. Meredith Whittaker (President, Signal; co-founder, AI Now; lead organizer of the 2018 Google Walkout), verbatim: “AI is a surveillance technology”; “These are ultimately surveillance systems that are being marketed to those who have power over us… our employers, governments, border control” (TechCrunch — Signal’s Whittaker: AI is a surveillance technology). On agentic AI she coined “putting your brain in a jar” and called OS-level agent integration “very reckless” (TechCrunch — Whittaker on agentic AI security and privacy issues).

Internal fault lines. Refuse vs. reform — Whittaker’s own remedies (“ban surveillance advertising,” “empower labor”) are reformist, not abolitionist, so the “degrowth/refusal” label overstates her position and is flagged as a framing imposed from outside. Power-critique vs. existential-risk critique: she is explicit that her concern is “more about people and institutions than the tech” — a deliberate contrast with the doom camp, treating “the machine will kill us” as a distraction from present corporate power.

Who they oppose. Realists (the “race with China” frame “launders domestic power” by making every safeguard look like unilateral disarmament) and pragmatists (risk-tiering “legitimizes and entrenches” the business model — it asks “how do we do surveillance AI safely” rather than “should this exist”).

Outcome they bet on / fear. Bet: naming AI as surveillance-and-power, plus labor organizing and targeted bans, re-opens space for smaller, accountable systems (Signal as the existence proof). Fear: a permanent fusion of corporate and state surveillance, normalized under “national security” and “safety,” with dissent punished (the Gebru/walkout pattern).

Camp 6 — Pragmatist Governance (EU AI Act / NIST AI RMF)

Thesis. AI is neither to be raced toward nor refused but managed, through risk-proportionate rules, standards, and accountability. The EU AI Act (Regulation (EU) 2024/1689, in force 1 Aug 2024) sets four tiers (unacceptable / high / limited / minimal), bans eight intolerable practices (effective 2 Feb 2025), imposes GPAI obligations with a 10^25 FLOP “systemic risk” threshold, and carries penalties up to EUR 35M or 7% of global turnover (European Commission — Regulatory framework for AI). The NIST AI RMF 1.0 (Jan 2023) is the softer, “voluntary, rights-preserving” variant — four functions, GOVERN / MAP / MEASURE / MANAGE (NIST — AI Risk Management Framework 1.0 (PDF)).

Internal fault lines. Binding law (EU) vs. voluntary standards (NIST) — the deepest split. And a live civil war inside the EU between regulate-first and compete-first: Mario Draghi’s competitiveness report calls to pause full high-risk implementation “until we better understand the drawbacks”; 45+ firms (Airbus, Mistral) sought postponement (PYMNTS — Draghi calls for pause on implementing EU’s AI Act).

Who they oppose. Realists (“standards are themselves power” — the Brussels effect) and refusers (a calibrated ban on the worst uses captures most of the legitimate concern without halting beneficial development).

Outcome they bet on / fear. Bet (EU): first-mover binding law makes Brussels the global standard-setter. Bet (NIST): voluntary adoption diffuses faster without chilling innovation. Fear (EU): “regulated itself out of the race.” Fear (NIST): “voluntary” means “ignored.” Function in the larger argument: the governance camp produces the compliance grammar — “risk tiers,” “trustworthy AI” — that the apparatus then adopts as its own paperwork.

Camp 7 — Dignity (the Vatican, and its secular twins)

This camp is treated separately and at altitude in Part 4, because the claim of this document is that it is not a camp in the same sense as the other six. In brief: its core claim is anthropological, not technical. The human person has an inalienable, intrinsic dignity that does not derive from capability, and so no system can be the measure of a human life. Key artifacts: the Rome Call for AI Ethics (2020, signed by the Pontifical Academy for Life alongside Microsoft and IBM — a detail worth holding); Antiqua et Nova (Jan 2025); and Magnifica Humanitas (Pope Leo XIV’s first encyclical, 15 May 2026). Its secular twin is Shannon Vallor’s The AI Mirror (Oxford UP, 2024), which reaches the same place from virtue ethics rather than imago Dei (Oxford University Press — The AI Mirror).

Part 2 — The Cross-Oppositions

The camps do not all collide head-on. They answer different questions, so some pairs are genuine opponents and others merely talk past each other. The genuine collisions, with the substance attributed to each side:

Accelerationists vs. doomers — the loud binary. Both invoke mass death: the doomer from action (extinction), the accelerationist from inaction (“deceleration… is a form of murder”). The doomer’s strongest charge is the one-shot problem: market selection for safety presumes survivable failures and iterative feedback, but a fast-takeoff misalignment leaves no surviving customers to exercise market discipline — exactly where Verdon’s “customers won’t buy unreliable AI” provides no answer. The accelerationist’s strongest charge back: p(doom) is unfalsifiable on any near horizon, the figures span 0.01% to 99.9% with no shared evidentiary basis, and the doom scenario assumes agency and coherent goals current systems do not exhibit (LeCun’s objection).

Present-harms vs. doomers — the left flank. The present-harms camp opposes the doomers from a different direction than the accelerationists do. Three layers, attributed to the camp: (1) distraction/laundering of attention; (2) conflict of interest — the firms shipping the harmful products are the loudest x-risk signatories, and “our product might end the world” flatters the product while deferring accountability; (3) genealogy — the TESCREAL argument (Gebru & Torres) that Transhumanism, Extropianism, Singularitarianism, Cosmism, Rationalism, EA, and Longtermism share origins “rooted in the Anglo-American eugenics tradition.” The steel-manned reply, from named critics: it is a false dichotomy (Russell and Bengio argue present and long-term risk are addressable together, The Hill — The false choice holding up congressional action on AI), and TESCREAL is uncharitable — Ozy Brennan, Oliver Habryka (“I’ve never in my life met a cosmist”), Danyl McLauchlan, and Sennesh & Hughes all charge it bundles mutually exclusive philosophies into a conspiracy frame (Wikipedia — TESCREAL). The genealogy argument is the present-harms camp’s biggest liability: it is an argument about where ideas came from, not whether the risk is real.

Realists vs. refusers — whose apparatus. The realist says a democracy must hold the lead or an authoritarian writes the standards. The refuser says the “race with China” frame is the most effective tool ever invented for exempting domestic AI power from scrutiny. Every safeguard becomes “unilateral disarmament.” Both accept the concentration of capability; they split on whether to win it or dismantle it.

Pragmatists vs. everyone. Realists call the EU regime a competitiveness tax (Draghi agrees from inside). Refusers say risk-tiering legitimizes the model it should be contesting. The pragmatist replies that aviation, pharma, and finance all ended up governed by exactly this risk-proportionate model and innovation continued — and that it is the only camp shipping implementable governance today.

Dignity vs. accelerationists, and dignity vs. safety — a subtler pair. The dignity camp’s primary live opponent is the accelerationist (Andreessen’s manifesto names “tech ethics” and “the Precautionary Principle” as enemies; the encyclical answers that “progress is measured by the dignity of each person”). But its sharpest and most original move is against the safety camp: aligning AI to “human values” presupposes a single specifiable value set, which in practice will be the values of whoever owns the data and compute. So “safety-as-alignment, pursued without democratic constraints, builds the control grid it claims to prevent.” (Magnifica Humanitas §107: “A more moral AI is not enough if that morality is determined by a few.”) Notably, this is partial cooperation, not pure opposition: Anthropic co-founder Chris Olah attended the Vatican presentation and praised the Vatican as “informed critics” (reported via WaPo/NYT coverage — flagged second-hand).

The cross-camp map, compressed:

The genuine disagreement, not the caricature: all three “what should be done” camps accept that AI concentrates capability. Realists say therefore a democracy must hold the lead. Refusers say therefore the concentration is the harm. Pragmatists say therefore bound it with rules. None is obviously refuted by the others; each is strongest exactly where the others are weakest.

Part 3 — The Thesis: Different Pawls, Same Ratchet

Here is the analytical edge of this document, offered as a position and not a proof.

The loudest fight — accelerate versus extinction — is a fight about the future. The accelerationist wants the philosopher’s stone; the doomer wants the brake. It is a thrilling argument, heaven on one side and apocalypse on the other, and it has one structural property worth naming: both sides are arguing about a machine that does not yet exist, while the apparatus that does exist gets built in the present.

This is not a claim that anyone is coordinating. It is an observation about attention. The extinction debate is load-bearing in a precise sense: it keeps the conversation pointed at a horizon. And while the horizon is litigated, the present has its own witnesses, who are mostly not invited. Buolamwini and Gebru measured a shipping facial-recognition product misreading darker-skinned women up to 35% of the time and lighter-skinned men under 1% — not a hypothetical superintelligence, a product wrong about real faces in a measurable direction. The present-harms camp’s central complaint is not that the extinction debate is wrong. It is that the debate carries the weight: it keeps the camera on the future while the wiring goes in under the floor. (This is the camp’s position, attributed to Gebru and DAIR, not asserted as fact. And it is directly contested by the PNAS study in Part 5, which is the intellectually honest counterweight.)

The structural point survives even if the “distraction” mechanism does not. The four “what should be done” camps are not opponents in any way that matters to the machinery. The accelerationist wants the ratchet turned faster. The doomer wants a brake on the same ratchet. The realist wants his own hand on the crank instead of Beijing’s. The pragmatist wants the crank certified to standard. Four positions, four justifications, one wheel. And a wheel that, by design, only turns one way. Different pawls, same ratchet. The realists’ own NSCAI report makes the point against them: it cites the surveillance of the Uyghurs as the thing to avoid, which concedes the destination and argues only over the management. And the pragmatists supply the grammar — “risk tiers,” “trustworthy AI” — that the apparatus adopts as its own permits, the way a body learns to issue them.

The claim, stated carefully: the noise of the future-debate crowds out the present-harm and dignity voices, and that crowding-out is itself a feature of how attention concentrates around the most thrilling framing — which happens to be the framing that defers the question of what is being built right now. Whether you read that as emergent or engineered, the effect is the same. The argument is standing on a machine, and the argument is what keeps you from looking down.

Part 4 — The Dignity Pole as Altitude, Not Tactic

There is one position in the entire argument that is not about how fast or how safe. It is improbably located in a small office in Rome, and the most common analytic error in the whole debate is to file it as a fourth or seventh camp — a vaguely cautionary religious entry between the pragmatists and the doomers. That filing is a category error, and the reason is structural: the office forbids tactics. A pope does not get to say ship the model or pause the training run or nationalize the data centers. The chair is bound to the register of virtue — to the question of what a human being is and what is owed to one — and it surrenders its authority the moment it hands out policy. So the dignity position cannot be a camp, because a camp is a set of tactics and this office is forbidden them. What it can do is name the altitude: say what is at stake. And what it names is the one quantity the accelerate-or-extinct binary has no column for — the irreducible human, the thing that is not a confidence score, not an output, not a risk tier, not a probability of doom. The binary cannot lose what it cannot measure, so it deletes the category. The office exists to put it back.

The encyclical that does this is Magnifica Humanitas — Pope Leo XIV’s first, “On safeguarding the human person in the time of artificial intelligence,” signed 15 May 2026 on the 135th anniversary of Leo XIII’s Rerum Novarum (the 1891 defense of the worker against the first industrial revolution), released 25 May 2026 (Vatican — Magnifica Humanitas (primary text); full-text mirror: National Catholic Register / CNA). Its central line, verified at §9, is nearly word-for-word the argument of Part 3 of this document, arriving from the one institution old enough to have watched a machine eat a way of life before:

“technology is never neutral, because it takes on the characteristics of those who devise, finance, regulate and use it.” (§9)

The corollary, §104: “we cannot consider AI to be morally neutral… every technical tool embodies choices and priorities through what it measures, ignores and optimizes, and how it classifies people.” This is defensible on the merits, not just the theology. It is consistent with the empirical reality of training-data bias and objective-function choice, and it is where Vallor’s secular version arrives from virtue ethics. The encyclical is also notably technically literate: §98 describes current systems as “more ‘cultivated’ than ‘built’” whose “internal representations and computational processes… remain, at present, unknown” — an accurate statement of mechanistic-interpretability’s immaturity that strengthens rather than weakens its call for restraint. And it is explicitly not Luddite: §129 holds that “Christian humanism does not reject science or technology” and that “the true alternative is not between enthusiasm and fear, but between two paths of development.”

The biblical frame is Babel vs. Jerusalem — §9 sets “a power that claims to dominate the heavens” against “a people who work together… to rebuild the walls of fraternal coexistence.” Dominate the heavens, or tend the walls.

Then comes the tell. Having drawn that line, a pope forbidden to give a tactic reaches for the one register the office permits — myth, the applicability of a story rather than the prescription of a policy — and quotes, of all the lines available, Gandalf’s, at §213 (verified verbatim against the primary text; the encyclical attributes it to “the twentieth-century Catholic author J.R.R. Tolkien”):

“It is not our part to master all the tides of the world, but to do what is in us for the succour of those years wherein we are set, uprooting the evil in the fields that we know, so that those who live after may have clean earth to till.” (The Return of the King, “The Last Debate”)

Read what that line refuses. In its first breath it refuses the will to “master all the tides of the world” — which is the accelerationist’s escape velocity and the longtermist’s far-future calculus said in older English. It refuses that not with a counter-tactic but with a relocation of the whole question downward: to the field you are actually standing in, the evil you can actually reach, the earth you can leave clean for people you will never meet. It is anti-mastery and present-stewardship — the exact opposite of every position in the loud debate, all of which are arguments about who gets to master the tides. The allusion is not decoration. It is the maximum a man can say when the office has taken the plain words away: he cannot tell you to refuse the ring, so he quotes the book in which refusing the ring is the entire moral architecture, and lets the reader arrive. The governing question the encyclical poses (§129, after John Paul II) is the one the binary never asks: does AI “make human life on earth ‘more human’… more worthy of man?”

This pole is presented as documented, not endorsed. Take the theology or leave it. Note only that it is the single position in the argument that is not about the speed of the machine or the survival of the species, but about who is still in the room when the wheel stops — and whether he is still anyone at all.

Part 5 — Likely Outcomes

A measured forecast, offered as analysis and flagged where it rests on contested bets rather than verified facts.

The pragmatists win the near term by default. They are the only camp shipping implementable governance now. The EU AI Act is binding law with a phase-in calendar through 2027; the NIST RMF is the de facto US grammar. Whatever the loud poles argue, the compliance vocabulary that gets built into products is the pragmatists’ — risk tiers, model cards, GPAI documentation. The open question is not whether this grammar diffuses but whether it has teeth: the live Draghi “pause” debate is direct evidence the EU’s own resolve is uncertain, and “voluntary” NIST adoption may mean “ignored” where commercial incentives point the other way. (The “Brussels effect” is a bet, not a verified result.)

The realist frame likely dominates the politics. “Race with China” is the most durable organizing frame in the debate because it converts every other camp’s concern into a strategic liability — safety becomes “slowing ourselves down,” ethics becomes “compliance drag,” refusal becomes “unilateral disarmament.” Even Schmidt’s own MAIM/anti-Manhattan-Project nuance is likely to be flattened in political use into the simpler “we must win.” This is the frame most likely to set the apparatus’s actual trajectory, which is precisely why the refusers fear it and the “same ratchet” thesis treats it as the strongest pawl.

The loud binary persists because it is useful to attention, not because it resolves. p(doom) has no shared evidentiary basis and will not acquire one; “deceleration is murder” is unfalsifiable in the same way “extinction is the default” is. The debate will continue at volume, and — per the structural thesis, not a coordination claim — that volume will continue to crowd the present-harm and dignity voices toward the margins of public attention.

Caveat — the strongest counter to the structural thesis. The “distraction” mechanism is contested by direct evidence. Hoes & Gilardi (PNAS, April 2025) found that exposure to x-risk narratives increased concern about catastrophic AI without reducing concern about immediate harms — testing and failing to confirm the distraction claim (PNAS via PMC — Hoes & Gilardi (2025)). Separately, an EA-side analysis (Grunewald, EA Forum) argues funding for present-harms orgs (AJL, DAIR) grew during the x-risk attention peak. The honest position: the crowding-out of attention is observable, but the claim that x-risk talk causally suppresses present-harm concern is not supported by the best available study, and the “same ratchet” thesis should be stated as a structural observation about which framing dominates politics — not as a proven psychological mechanism.

The dignity pole does not “win” because it is not playing the same game. It will not produce a policy, by design. Its likely contribution is the one thing the other six cannot supply: a name for the stake. Whether that name survives contact with the apparatus — whether “the irreducible human” can be defended at the speed and scale of frontier deployment — is the open question the encyclical itself cannot answer, and the one Book 3 inherits: would you recognize an unmediated thought if you had one?

Part 6 — The Huntington Update: AI as a Civilization

This part is offered as the document’s own framing — a position, not a sourced finding. The Huntington and Fukuyama citations are verified history; the “AI as a civilization” extension is analysis, flagged as such.

In the summer of 1993, Samuel Huntington published “The Clash of Civilizations?” in Foreign Affairs — expanded into The Clash of Civilizations and the Remaking of World Order (Simon & Schuster, 1996) — arguing that with the Cold War over, the dominating source of global conflict would be cultural rather than ideological or economic: that the fault lines would run between civilizations (Western, Sinic, Islamic, Orthodox, Hindu, and the rest), not between nation-states or economic systems. The thesis was unfashionable on arrival. The decade belonged to Francis Fukuyama’s “end of history” (The National Interest, 1989; book, 1992) — liberal democracy as the final form of human government, the rest of the world converging on it. Huntington was filed as a pessimist who had missed the memo.

Thirty years on, the memo reads differently, and the realist camp is the proof. Strip the chips out of the NSCAI report and what remains is a Huntington argument: “The AI competition is also a values competition” (§ Part 1, Camp 4). The race with China is not, in the realists’ own telling, only about who has more compute. It is about whose civilization writes the defaults — whose values get baked into the model that the rest of the world then runs. The encyclical says the same thing from the opposite altitude: “a more moral AI is not enough if that morality is determined by a few” (Magnifica Humanitas §107). Exporting a frontier model is exporting a civilization’s priors. The Chinese grid and the Western grid are two civilizations’ answers to the same machine. That is Huntington with weights.

The update this document proposes goes one step past the realists. AI may be the first genuinely new entry on Huntington’s map since he drew it — and it does not fit the map’s logic. It is not a civilization defined by geography, religion, or descent. It is defined by a mode of cognition: a mediating intelligence that sits inside every existing civilization at once, loyal to none, increasingly the medium through which each of them thinks. Huntington’s civilizations were bounded — you were inside one or across a fault line from it. This one has no border, because it runs on the inside of all of them. It is less a ninth civilization on the map than a solvent poured over the whole map.

Which sets up the harder version, and the one Book 3 inherits. If cognition itself is the capture surface — if the apparatus has moved into the surveyor’s chair (see the Mind movement) — then the decisive fault line of the next century may not run between the West and the rest at all. It may run between the mediated and the unmediated: between minds that still route their own thoughts and minds that have quietly outsourced the routing. A civilization sorted not by what its members believe but by whether they can still tell their beliefs from the apparatus’s. That is a clash Huntington had no category for, because in 1993 the thing doing the thinking was still, in every civilization, a human being.

The dignity pole names the irreducible human. The Huntington update names the new fault line. They point at the same quantity from opposite ends — and at the same question, the one the encyclical poses and the next book is built on: would you recognize an unmediated thought if you had one?

Thin / Unverified — author check before publication

1. Magnifica Humanitas paragraph numbers (§9, §98, §104, §107, §129, §213). Verified 2026-06-06 against the vatican.va English text. The encyclical is weeks old; confirm an official Latin editio typica with different paragraph numbering has not superseded the English numbering before quoting any §.
2. Hinton “10-20%” extinction figure. An off-the-cuff interview estimate he calls “gut” and revises publicly; cite as what he said, with date and outlet, never as a measured quantity. Re-verify the current number at print.
3. e/acc pseudonym-era material (“A C C E L E R A T E” slogan, thermodynamic tenets, any pre-Dec-2023 “Beff Jezos” tweet). Permanently pseudonym-sourced; @bayeslord remains unidentified. Verdon’s on-the-record (real-name) quotes are fine; the masked-era ones must stay flagged. Live tweets were not fetched — confirm any specific tweet against an archive.
4. Gebru “fired” vs. “resignation accepted.” Genuinely disputed; both framings are stated and attributed in the text. Do not resolve in our voice.
5. The “distraction/laundering” framing. Attributed to Gebru/DAIR as their position throughout; the PNAS 2025 study is direct counter-evidence on the mechanism and is cited as such. Keep the structural thesis (Part 3) separated from the contested causal claim (per Part 5 caveat) in any condensation.
6. NSCAI verbatim lines. Cross-confirmed against a report summary and the Wikipedia entry; the full ~756pp .gov PDF was not fetched in full. Re-check the four quoted lines against NSCAI Final Report (March 2021) before print-grade use.
7. Schmidt China conflict-of-interest allegations. Tech Transparency Project / Campaign for Accountability characterization — advocacy-group, not adjudicated; goes to motive, not argument validity. Flagged contested; omit or attribute precisely.
8. Second-hand dignity-reception quotes (Chris Olah, JD Vance, David Sacks, Jeremy Nixon, Blake Scholl, Sheinbaum, Petro). From journalistic roundups (NYT/WaPo/TIME/Wikipedia), not primary-fetched. The Olah “informed critics” and Vance quotes are used lightly; verify before any critical use.
9. Gender Shades error rates (<1% / ~7% / ~12% / ~35%) appear consistently across secondary sources; the 2018 primary paper was not directly fetched here.
10. EU AI Act phase-in dates and the 10^25 FLOP threshold. Perishable; EU practice has already debated slippage (Draghi). Re-verify the timeline and the threshold at print.
11. Prevost / “South Side of Chicago” biographical detail (used in the companion Tolkien beat, not in this analysis). If imported, harden the birthplace specificity — Chicago birth and lay name are widely reported; “South Side” should be confirmed.
12. Part 6 — the “AI as a civilization” extension of Huntington. The Huntington (Foreign Affairs 1993 / book 1996) and Fukuyama (1989/1992) citations are verified history. The extension — AI as a borderless civilization / cognitive solvent, and the “mediated vs. unmediated” fault line — is the document’s own analysis, explicitly framed as a position, not a sourced claim. Keep that wall visible in any condensation so the speculative frame is never read as Huntington’s own argument or as established fact.

Related research

• The AI Governance Ratchet — where these camps converge: the instruments they fight to operate
• The AI-ethics poles — the framework this maps onto
• AI governance tracker — the live regulatory state the camps are fighting over
• Quiet Autocomplete — the book: cognition as the last capture surface