I keep returning to a half-formed intuition:
Maybe we oppose blueprints not because planning is bad, but because static plans can’t adapt. Real societies need organic behavior, local knowledge, and self-correcting mechanisms — elections, peer review, amendments, markets that punish error. Any formal contract or institution, meanwhile, is incomplete by necessity (Hart & Moore, Simon 1951): it cannot specify every future state of the world, so it always sacrifices something at the margins — someone excluded, some contingency unaddressed, some value compressed into a one-size-fits-all rule.
From there, a seductive next step:
What if Dataism — treating society as an information-processing system — plus explicit self-correction is the first blueprint without those defects? Not a frozen constitution, but a living loop: measure, update, re-optimize. Distributed nodes instead of a central planner. Continuous feedback instead of a one-shot utopian redesign.
I don’t think that works as stated. But the intuition points at a real tension in AI governance, alignment, and institutional design — and the useful output may be questions, not a manifesto.
This essay is a research agenda in prose.
What “anti-blueprint” is actually rejecting
The utopia canon I read for this (Scott, Popper, Ostrom) is often summarized as “don’t plan.” That’s misleading.
Popper rejects utopian social engineering paired with historicism — the claim that someone knows history’s destination and may override dissent to get there. He favors piecemeal reform: test, fail, revise, retain the power to reverse (The Poverty of Historicism).
Scott rejects legibility-first redesign: when states or platforms force complex social life into a single measurable grid, they destroy mētis — tacit, local know-how — and often trigger disaster when high modernist confidence meets coercive power.
Ostrom rejects the tragedy-of-the-commons default that only Leviathan or privatization can solve resource problems. Polycentric systems — nested rules, local monitoring, graduated sanctions — can adapt without a master plan.
Shared thread: good order is process-shaped, not blueprint-shaped.
Questions
- Is “anti-blueprint” really anti-finality, anti-monism, or anti-coercion? Those are different objections with different policy implications.
- When is a “design principle” (Ostrom’s eight, Beer’s viable system model) still a blueprint in Popper’s sense?
- Does AI change the calculus because models are explicitly trained to update — or does deployment at scale recreate Scott-style legibility violence faster than any Soviet five-year plan?
Incomplete contracts and the sacrifice you can’t contract away
Grossman & Hart (1986) and Hart & Moore (1990) formalized what lawyers already knew: contracts can’t list every contingency. What matters is who holds residual control rights — authority over decisions the contract didn’t specify.
Human employment works partly because implied terms fill gaps: reputation, guilt, social sanction, “you don’t knock over the vase.” Baker, Gill, and Korinek’s AIES paper argues AI alignment is structurally similar: you can’t write a complete spec for an autonomous agent, so who can intervene, renegotiate, and override is the alignment problem in institutional dress.
Recent hype adds a twist: Will AI agents write perfect contracts? Cheap drafting might shrink some incompleteness — but enforcement, evidence, probabilistic behavior, and norms outside the text remain. Completeness is not the same as justice.
Sen’s capabilities approach adds a normative lens: even when GDP rises, conversion factors (disability, discrimination, missing public goods) mean the same resources buy different lives. A contract that pays everyone equally can still sacrifice freedom if the metric ignores what people can actually do and become.
Questions
- If AI shrinks drafting incompleteness but enlarges measurement incompleteness (what gets logged vs. what matters), is net incompleteness actually falling?
- When residual control over a model sits with a lab, a regulator, or a user, who bears the sacrifice when the unspecified state arrives?
- Can “implied terms” for AI — social norms, liability, shame — replicate human contracting without anthropomorphizing systems that don’t care about reputation?
- Is the user’s intuition (“institutions always sacrifice something”) a theorem about scale, value pluralism, or limited cognition? Which defeat conditions would falsify it?
Self-correcting mechanisms: the anti-blueprint’s positive content
Yuval Harari popularized “self-correcting mechanism” in public lectures and interviews (e.g. Big Think, 2025; related themes on Lex Fridman #390): institutions that can identify and fix their own mistakes without waiting for collapse.
His examples:
- Elections — try a government, revise at the ballot box
- Science — journals publish corrections; progress via refutation
- Constitutional amendment — rules admit human fallibility (“We the People,” not infallible scripture)
- Contrast: dictatorships and dogmatic traditions lack internal correction
This is Popper for a general audience: conjectures and refutations, open society, falsifiability transferred from science to politics.
Hayek’s knowledge problem adds a complementary point: correction requires information flow. Prices, profit, and loss are feedback signals decentralizing knowledge that no planner possesses. Lynne Kiesling summarizes the institutional upshot: markets are discovery procedures, not static equilibria (Knowledge Problem, 2023).
Harari’s Homo Deus framing goes further: politics as comparative information processing — capitalism and democracy as distributed networks; centralized plans as brittle (chapter summary). Jay Galbraith’s information-processing view of organization design makes a similar move in management theory: higher uncertainty demands richer horizontal information systems, not thicker rulebooks.
Questions
- Is self-correction constitutive of legitimacy (Sen: freedom as participation) or instrumental (fewer famines, faster science)? When they conflict, which wins?
- Harari warns that AI may break democracy because fake humans flood the conversational public sphere (Lex #390, ~00:38). Is that an attack on correction mechanisms themselves, or on a precondition (trustworthy speech)?
- Popper’s correction is social — critics, voters, reviewers. ML “self-correction” at inference time (generate, verify, revise) is internal to the model. Are these the same family, or a category error that launder social legitimacy through engineering?
- Who decides that a correction occurred? Science has retractions; democracy has elections; markets have bankruptcy. What is the error signal in a platform model optimized for engagement?
The seductive hypothesis: Dataism + correction = the non-blueprint blueprint?
Dataism (Harari’s term, not an academic school) treats the universe as data flow and elevates maximizing processing/connection as supreme value — a successor religion after humanism (Homo Deus). Critics read it as techno-religion: transparency without politics, optimization without dignity (Crossings, 2020; Chapman Center).
The user’s hypothesis merges:
| Anti-blueprint insight | Dataist promise |
|---|---|
| Organic / local adaptation | Continuous learning from data |
| Feedback loops | Real-time metrics and A/B tests |
| Incomplete contracts | Dynamic renegotiation / RL updates |
| No single planner | Distributed compute + markets |
Why it’s tempting: it resembles cybernetic governance done right — Stafford Beer’s viable systems with feedback, not Glushkov’s OGAS fantasy of one optimal national plan. Eden Medina’s Cybernetic Revolutionaries shows the Chilean Cybersyn attempt at adaptive industrial control — and how politics, not wiring, determined survival.
Why I’m skeptical: swapping a static blueprint for a moving objective function doesn’t eliminate sacrifice — it relocates it into the loss function, the logging schema, and residual control over the stack.
Questions
- Does Dataism actually propose no telos — or a new telos (throughput) smuggled in as engineering neutrality?
- Can a system be simultaneously optimized (single reward) and self-correcting in Sen’s sense (expanding plural capabilities)? RLHF already shows Goodharting on proxy rewards — is that “correction” or metric entrenchment?
- Scott: readable variables prosper; unreadable ones die. Does a data-driven adaptive system amplify legibility bias faster than a dumb paper regulation?
- Algorithmic governance from the bottom up: when is “adaptive” governance democratic (contestable, exit-rich) vs. technocratic (transparent but unaccountable)?
- Wiener warned in The Human Use of Human Beings that cybernetic feedback without ethical commitment becomes amplified control, not liberation. Where does Wiener sit in this debate today?
Three traps when “adaptive” becomes a blueprint
Trap 1: Correction without contestability
A loop that updates weights is not a loop that admits political defeat. Popper needs institutions where losers live to argue again. If “self-correction” means the system nudges users toward compliance, you’ve built adaptive coercion — what Byung-Chul Han calls a transparency society stripping negativity from dissent.
Trap 2: Residual rights move to the stack
Incomplete contracting doesn’t vanish when contracts are generated by LLMs. Someone still owns the data, sets the benchmark, pays for compute, holds the kill switch. AEI’s agent-contract piece is optimistic about drafting; Baker et al. is clearer about control rights when the agent surprises you.
Trap 3: Sacrifice is deferred, not removed
Every correction presupposes an error — often paid for in lives, livelihoods, or capability loss before the loop triggers. Harari on AI: we may be running a historical experiment we cannot simulate; mistakes may be terminal before correction kicks in (WIRED on Nexus).
Questions
- What is the minimum viable contestability for an adaptive AI institution — analogous to amendment + judicial review + free press?
- Should corrigibility in alignment (Russell, Human Compatible) be treated as a machine analog of self-correction, or as a component inside a larger social correction loop?
- Can we name sacrifices ex ante (who is likely to lose when the metric updates) the way environmental impact statements name ecological tradeoffs?
- Is “no blueprint” compatible with constitutional AI — fixed principles, adaptive application — or is that just a blueprint with better marketing?
A sharper framing: design principles, not terminal states
If the Dataism + correction hypothesis fails as a perfect blueprint, what survives?
Candidate: meta-institutional principles — not “the society at the end of history,” but constraints on how rules may change:
- Polycentricity (Ostrom): multiple centers of decision, nested feedback
- Piecemeal falsification (Popper): reversible experiments, recorded failures
- Capability audit (Sen): did this reform expand real options for the worst-off?
- Convivial limits (Illich): tools must not grow past human scale of comprehension and refusal
- Residual human override: formal and real authority stay separable (Aghion & Tirole, 1997)
That is closer to governance research than to Harari’s Dataism — data as input to piecemeal inquiry, not as sacred output.
I sketched a related triangle elsewhere in my notes (Popper × Acemoglu × Gebru/Torres): resist historicism, steer technology direction without claiming a final society, and refuse salvation narratives that defer present harms. Dataism-as-blueprint fails all three tests unless heavily amended.
Questions
- What would an Ostrom design-principles checklist for foundation-model deployment look like — monitorable, local, graduated sanctions?
- Acemoglu’s mission orientation vs. Popper’s piecemeal: when does “steer AI toward labor-complementing uses” become the new utopian engineering Popper feared?
- Can data cooperatives and public infrastructure (algorithmic governance, bottom-up) supply distributed correction without Dataist metaphysics?
- What empirical cases best test adaptive vs. blueprint failure modes in AI — Cybersyn, China’s social credit, EU AI Act conformity assessment, RLHF iteration loops?
AI-specific pressure points
Alignment: Production stacks optimize scalar rewards; Gabriel’s taxonomy (2020) reminds us that choosing the target is politics. Self-correction inside the model (critique-and-revise, constitutional RLAIF) does not answer who writes the constitution.
Agent economy: Autonomous agents negotiating at machine speed recreate incomplete contracting + hold-up faster than courts adapt. Adaptive smart contracts may automate renegotiation while concentrating residual rights in protocol designers.
Speed vs. democracy: If information processing accelerates 1000× but electoral and deliberative cycles don’t, is the bottleneck institutional latency? Galbraith suggests redesigning organizations for uncertainty — but who participates in that redesign is the democratic question Harari keeps asking.
Questions
- If inference-time scaling is “self-correction,” do we owe users a right to the pre-correction draft — analogous to seeing legislative debate, not just final statute?
- When an agent economy protocol auto-updates terms, is that piecemeal engineering or rug-pull governance?
- Does a society with superhuman idea-generation (Harari on LLMs, Lex #390) lose the ability to resist ideas — Plato’s cave as engineering problem?
- What would falsify the claim that distributed AI inference is politically equivalent to Hayekian price discovery?
Research directions (if you want to pick this up)
| Track | Methods | Output |
|---|---|---|
| Institutional | Case studies: Cybersyn, OGAS fragments, platform moderation appeals, EU AI Act | Compare stated vs. real correction loops |
| Economic | Hart/Moore applied to model ownership, API terms, agent wallets | Map residual control in ML supply chains |
| Philosophical | Popper vs. Sen vs. Harari vs. Wiener | Typology of “correction” (social / scientific / algorithmic) |
| Empirical AI | Benchmark whether RLHF/RLAIF cycles expand or shrink user capability sets | Sen-inspired metrics, not just harm scores |
| Legal | Agent-drafted contracts + enforceability | Stress-test “perfect contract” optimism |
Closing: living with a good question
The intuition that organic feedback beats static blueprints is sound. The leap to Dataism + self-correction as the first sacrifice-free design is not — not because adaptation is worthless, but because adaptation without contested values is just faster optimization.
The productive version of the question might be:
Under what conditions can high-frequency feedback improve institutional learning without concentrating residual control, flattening mētis, or closing the political question of what counts as an error?
I don’t know the answer. I’d like to see more people treat that sentence as a research program — especially anyone building the systems that will define “error” for the rest of us.
If you work on this from any angle (legal, cybernetics, alignment, development economics), I’d like to hear what cases you’d add. This post will update as the literature and the deployments evolve.
Further reading
- James C. Scott, Seeing Like a State (1998)
- Karl Popper, The Open Society and Its Enemies (1945); “Science: Conjectures and Refutations” (1963)
- Amartya Sen, Development as Freedom (1999)
- Elinor Ostrom, Governing the Commons (1990)
- Yuval Noah Harari, Homo Deus (2016); Nexus (2024)
- Norbert Wiener, The Human Use of Human Beings (1950)
- Friedrich Hayek, “The Use of Knowledge in Society” (1945)
- Baker, Gill, Korinek, “Incomplete Contracting and AI Alignment” (AIES 2019)
- Veale & Borgesius, “Algorithmic Governance from the Bottom Up” (BYU Law Review, 2023)