The AI Fork Is About Agency
Prometheus, again
When Gutenberg’s movable type began its spread in the late fifteenth century, European humanists reached for the most extravagant vocabulary available. Cardinal Turrecremata called printing “the divine art.” Ulrich Zell declared it “the invention of inventions.” Erasmus imagined a world in which even the poor could acquire the books that once required princely wealth. Knowledge would ascend from the monasteries into the public square; literacy would rise and ignorance wither; scripture and classical learning would percolate into every household. The press was heralded as nothing less than an epistemic liberation engine.
Within a generation, the most printed categories were not the classics or the sacred texts. They were political pamphlets and agitprop, apocalyptic predictions, scurrilous theological attacks, astrology guides, miracle tales, sex stories and bawdy plays, forged revelations and sensational hoaxes. The Protestant Reformation — and its bloody aftermath — was impossible without rapid-fire print propaganda; peasant uprisings leveraged pamphleteering instead of sermons. The average viral product of the press was not humanist scholarship but the sixteenth-century equivalent of a meme, coded in satire, fear, or outrage. The humanists dreamed of a republic of letters. They got an economy of pamphlet wars.
Artificial intelligence is being received with the same millenarian optimism, and that reception invites comparison with earlier informational upheavals. This chapter closes Part VII by putting the catalyst in historical perspective: first a recurrent pattern among informational superpowers, then AI’s possible instantiation of it, and finally the fork this amplification presents — which does not run between believers and skeptics, but through each user’s relation to their own agency.
The Gutenberg Arc
Every civilization that receives a new informational superpower reacts the same way: it sees a path to collective uplift, expects universal enlightenment, and then discovers that the median human appetite has not changed. Fire becomes arson as well as metallurgy; printing becomes pamphleteering as well as philosophy; the internet becomes Wikipedia as well as algorithmically optimized sludge. The arc has five stages, and the press ran through all of them.
The utopian promise came first, in the rhetoric above. Then the amplification: not of wisdom, but of everything — the full distribution of human appetites, brilliance and vulgarity alike. Then the institutional shock. Information began moving faster than the institutions built to interpret it, and the authorities responded with the perennial reflex: mandatory printer licenses, prior censorship, state-controlled presses, church indexes of banned books, arrests, fines, and executions. Information escapes; power reacts. Alongside the crackdown came the moral panic, and its rhetoric is worth quoting because it could be pasted into a contemporary op-ed unchanged: “Books pour forth without moderation and without discrimination.” “Errors spread faster than the truth can refute them.” “The world is full of madmen who now print their ravings.” These are the sixteenth-century versions of misinformation, attention collapse, and the death of expertise. Whenever cognitive power becomes cheap, gatekeepers panic.
Last comes the equilibrium — and it is real but unequal. The press eventually stabilized into a mixed ecology: mass literacy, newspapers, scientific treatises, the Enlightenment, the scientific revolution itself. But the trash never disappeared; modern bookstores still overflow with tabloids, pseudoscience, and horoscope annuals. The printing press gave us both Newton and Nostradamus almanacs. The internet gives us both arXiv and OnlyFans.
That last comparison matters, because the standing lament about the internet — that it fell from its early techno-utopian promise into a morass of low-grade discourse — mistakes the arc’s middle for its end. The internet’s current state is not a failure relative to its ideals. It is the early Gutenberg phase of a vastly accelerated media cycle: an explosion of noise before the maturation of new filtering standards, which last time took the form of scientific norms, peer review, constitutional protections, and specialized institutions of knowledge curation. We are not witnessing decline; we are living through the predictable turbulence before stabilization. Democratized knowledge platforms do not uplift humanity uniformly. They amplify human nature — and reveal it.
Prometheus in Code
AI boosters now describe their moment in the same Promethean vocabulary: universal tutoring for every child, superhuman assistants democratizing expertise, flawless translation dissolving cultural barriers, automated research engines solving disease, climate, and scarcity. The dream underlying all of it is simple: cognition has become a commodity, therefore wisdom will become a mass good. It is the same assumption humanity makes at the birth of every cognitive tool — that expanding access to high-quality reasoning will expand the prevalence of high-quality reasoning. It never works like that.
AI’s immediate effect is not enlightenment but magnification. Whatever patterns a society contains — curiosity, tribalism, malice, brilliance, apathy — AI amplifies them at scale: hyper-accelerated misinformation, infinitely personalized propaganda, synthetic identity swarms, weaponized parasociality of the kind Artificial Intimacy dissects, algorithmic addiction shaping attention with surgical precision, high-speed scams and mass persuasion attacks. This is not because AI is evil. It is because automation amplifies the full distribution of human incentives, not the distribution optimists want. We did not get a world where the average person reads the Federalist Papers because printing became cheap. We got a world where the average person reads whatever aligns with their appetites.
The institutional shock repeats too, wider than before. Legal systems that operate on years collide with models that generate new strategic surfaces daily; education systems calibrated for human teachers face students with autonomous cognitive tools; media institutions built on scarcity lose relevance in a world of infinite generative supply; regulators attempt to freeze what is while the technology mutates beneath them. The tempo mismatch is not a bug — it is the transitional phase all media technologies impose. The difference now is velocity that makes the printing press look geological. And the moral panic has already arrived on schedule, its slogans barely updated: AI will destroy truth, dissolve the social fabric, produce nihilistic content; people cannot be trusted with these tools.
What the arc predicts for the equilibrium is the part the boosters least want to hear. AI will not uplift humanity equally, for the same reason universal literacy did not produce universal rationality. It will crystallize a power-law distribution of cognitive effectiveness: high-agency individuals becoming massively more capable, low-agency individuals outsourcing ever more of their cognition, the middle diverging into competence clusters and dependence traps. AI will widen the variance of human outcomes, not compress it. The printing press made philosophers more dangerous, preachers more persuasive, bureaucracies more efficient, and mobs more coordinated. AI is the next iteration.
The Cognitive Reservoir
Axio has a name for what each of these technologies did. Chaos as Foundation introduced the Chaos Reservoir — the total ensemble of possible patterns, from which everything lawful and meaningful is carved by Coherence Filters. Every prior media technology abruptly expanded a civilization’s access to that reservoir: more patterns available, faster than the existing filters could process. Early in each cycle, Chaos outpaces Coherence — noise dominates signal, institutions lack the interpretive machinery for the flood, established authorities panic. Only later do new filters mature and Coherence catch up, never suppressing Chaos, only balancing it.
AI does something categorically new to this dynamic. It creates a Cognitive Reservoir: a persistent, accessible layer of generative problem-solving power — the same object The Agency Criterion identified from the inside, the crystallized modeling activity of billions of minds compressed into weights. And the Cognitive Reservoir has a property no prior reservoir had: it does not merely supply patterns; it supplies constructed coherence on demand.
The consequences are specific. The boundary between Chaos and Coherence collapses at the point of use — the flood now arrives pre-organized, fluent, plausible. Individuals can outsource not just pattern-generation but pattern-organization, the filtering itself. And institutions built to filter Chaos cannot keep pace with a technology that manufactures Coherence. AI does not just expand what can be known; it expands what can be generated, moving society from a scarcity economy of cognition to a saturation economy of cognition, which every agent navigates according to their existing capacities.
That is why this cycle’s differentiator shifts. In Gutenberg’s cycle, access differentiated: literacy, then libraries, then bandwidth. When coherence itself is free, access differentiates nothing. What differentiates is effective agency — the capacity to direct the reservoir rather than submit to it. Users who direct it become powerful; users who submit to it become dependent. The dream that AI will uplift everyone equally assumes that effective agency is evenly distributed. Axio rejects this assumption, and the arc’s widening variance is the rejection made empirical.
The Fork
A framing that circulates among AI researchers — the pseudonymous commentator Teortaxes gave it its sharpest form — holds that a divergence has already begun among the cognitive elite, and that both camps have passed their respective points of no return: one group can no longer be productive without AI, while the other has formed a stubborn reality bubble in which AI doesn’t work. The meme attached to the claim draws it as a literal fork in the road: one path leading to a shining castle labeled posthumanism (dependence on AI), the other to a storm-wracked ruin labeled subhumanism (mindkilled on AI), with a lone bright kid standing at the junction.
The observation is real and the diagram is wrong, in an instructive way: as drawn, both roads lose. Dependence and mindkill are not opposite destinies; they are the same destiny — agency surrendered — reached by different routes. And the picture is missing a path. The emerging divide over AI is usually described as believers versus skeptics, accelerationists versus doomers, tool-users versus refusers. The relevant fork concerns agency: who preserves it, who displaces it, and who refuses the tool because refusal feels like sovereignty. AI is a cognitive amplifier — of drafting, abstraction, search, simulation, critique, translation, refactoring, adversarial review. It can also turn capable people into curators of plausible sludge. The difference lies less in the model than in the user’s control architecture.
The sharpeners. At its best, AI functions as the external dialectic catalyst: it compresses material, recombines ideas, generates objections, exposes weak joints, gives thought more surface area. It has coherence without agency — the previous part’s central verdict — and holding that distinction prevents two symmetric errors: treating AI output as authoritative intelligence, and treating lack of agency as lack of cognitive value. The human remains responsible for intention, judgment, taste, epistemology, and final authority. The machine can produce candidate structures, propose analogies, map tensions, pressure-test claims; it cannot decide what is worth valuing, what is true enough, or what should be authored under one’s name. The sharpener’s practice is disciplined symbiosis: use AI aggressively, but bind it to verification loops. Use it for variation, compression, critique, scaffolding, and adversarial pressure; use it to make your assumptions visible and force contact with objections; use it as opponent, editor, simulator, and index. Do not let it become your taste, your conscience, your epistemology, or your will. Keep unaided competence alive — periodically do the work without the machine — and maintain domain models strong enough to detect fluent nonsense. The full regimen, objection by objection, is the business of The Discipline of Thinking With AI.
The displacers. The AI-native failure mode is agency displacement: fluent output substituting for understanding, generated prose flattening voice, code accepted because it looks plausible, the user no longer maintaining an independent model of the work. Apparent capability rises while the evaluative faculty weakens. This — not overuse — is the real danger in heavy AI use. The risk is not that a serious user consults the tool too often; it is that convenience slowly replaces judgment. Automation bias is the predictable human tendency here: once the machine becomes normal, checking the machine begins to feel optional, inefficient, even irrational. The failure arrives gradually, disguised as productivity. And it will often be economically selected, because disciplined symbiosis is cognitively possible but institutionally fragile. Speed is easier to measure than judgment; output volume is easier to reward than retained competence. A manager can count tickets, drafts, and generated artifacts far more easily than he can measure whether a worker’s independent model of the domain is improving or degrading, so some institutions will prefer the appearance of cognition over the cost of actual expertise. That is not a user-interface problem; it is an incentive problem. Programming After Programming is the professional case study: a whole discipline crossing the mechanization threshold, where the practitioners who survive are precisely those who let value migrate from authorship to framing, verification, and warrant — sharpeners by necessity — while those who merely accept plausible output inherit the displacer’s trajectory at career scale.
The refusers. The opposite failure mode is cognitive refusal. Some people still argue against AI as though the current frontier were a 2022 chatbot demo. They treat hallucination as a disqualifying defect rather than an engineering constraint. They compare raw AI output against expert final work instead of comparing AI-assisted workflows against unaided workflows. That is refusal masquerading as discernment — independence preserved by declining a tool one has not learned to govern, cognitive refusal mistaken for sovereignty. It must be distinguished from principled refusal, which is a different thing entirely. Some people object to AI on legal, ethical, environmental, or economic grounds, and those objections deserve separate treatment: a person can acknowledge that the tool works while rejecting its provenance, business model, labor effects, or institutional incentives. That position is coherent. Capability denial is not.
Harder to Fool
The practical test is simple to state. After using AI, are you harder to fool, faster to clarify, better able to defend your claims, and more capable of acting from your own judgment? Or are you merely producing more fluent artifacts with less internal command over what they mean?
AI will not make weak agency strong. It will magnify the control structure already present. A disciplined mind gets leverage. An undisciplined mind gets velocity without steering. An avoidant mind gets purity at the price of reduced reach. The future belongs to the first group, because it is the only one that treats amplification as a responsibility rather than a shortcut or a contamination.
The Gutenberg arc says this turbulence is survivable — that the press did not uplift human nature but revealed it, and that new Coherence Filters eventually matured around the flood. AI may produce analogous turbulence, perhaps faster and deeper. But the comparison carries a caveat the historical cases did not: the press’s turbulence still included a century of religious war, and every prior reservoir was inert — patterns waiting for minds to use them. The Cognitive Reservoir is the first that talks back, and The Agency Criterion already identified strong incentives for binding it into genuine agents. Part VII has treated the machine as a catalyst in individual hands, and the fork it presents to each of us. What such amplification may do at civilizational scale — to risk, to power, and to the race between them — is a different question, and it begins with learning to say what we even mean by the odds of catastrophe: Making Sense of P(doom).