The objections, including the ones that go against the essay. A Q&A that only answers friendly questions is advertising.
Alongside: the essay · glossary · slides · the short version
Produced alongside the essay, out of the same multi-model inquiry — Claude Code, Claude in the browser, Perplexity and others — directed throughout by the author.
Because regulation and alignment both leave the same question unasked: can the people a system decides about say no to it?
Democratic AI° is this site’s answer, and it is an instrument rather than a slogan — four questions, returning a position on a scale rather than a badge:
The mark matters: democratic is not the same as democratically produced. A system built in a democracy, sold by firms headquartered in democracies and exported to allied democracies can still answer no to all four.
The connection to this essay is direct. This essay says what a democracy has to be able to do — tell its own intent from an approximation of it. Those four questions say what a system has to be for a democracy to be able to do it. Neither is much use on its own: a public with the capacity and no governable systems is a public that can see the problem and not touch it.
Because capability and standing are different things, and only one of them can be out-performed.
Smarter Is Not in Charge puts it as a distinction between capacity — the ability to build, run, calculate, decide well — and legitimacy, the rightful standing to be the one who decides. A machine can run away with capacity and gain none of the second, because legitimacy is not a skill.
Which inverts the usual worry. The better the systems get, the more clearly the remaining question is about authority rather than ability — and authority is not a technical problem. The operating rule across this site’s deliberative work is four words: the software rehearses; people decide.
Very likely not, and this is the version of the essay’s argument that is easiest to miss.
Plurality is how many suppliers there are — a fact about market structure, settled by counting firms. Epistemic independence is whether those suppliers give you separate evidence, separate ways of being wrong, separate ideas of what a good answer looks like. Counting firms does not settle it.
Ask four systems the same question, receive substantially the same answer, and the agreement feels like triangulation. It is closer to asking one source four times, because the systems share training data, tuning norms and incentives. That is false corroboration, and the measurable form of it is correlated error — two systems being wrong in the same way rather than in different ways. Both terms, and the evidence behind them, come from Democratic AI°; epistemic independence is this essay’s own label for what they are collectively about.
So a deliberating room that consults several AI systems may have heard one view, four times, in different words. That is the compression this essay is about, arriving by a route nobody chose.
Because each of the others is good at something the question does not need.
A referendum has breadth without reasoning: everyone can answer, nobody has to think it through together, and the majority erases the minority by design — 51% is a result and 49% is nothing. An assembly has the reasoning and not the breadth. A poll collects positions people already hold, which is the thing most vulnerable to everything described in this essay.
One question, many rooms proposes the obvious repair — run the same question in many sealed rooms at once, and consolidate the results into a map that never averages them. For a question where the disagreement is the whole point, and most questions worth deliberating are like that, keeping the disagreement in view is the method rather than a weakness of it.
Both, and the distinction matters.
The Irish assembly on abortion is a real institution with a real outcome. This site’s own demonstrations are machine-rehearsed rather than convened: The other 190 runs an AI-2040 plan past an assembly of nine composite members, and Our Own Machines produces a national AI-policy framework through 28 composite members across ten teams, sealed and verifiable, with the gaps declared rather than filled in.
Composite members are not people, and nothing in those exercises should be read as a public having spoken. What they demonstrate is that the machinery runs and produces something checkable — and, in the first case, that none of the three camps that emerged said “adopt the plan”, which is a finding about legitimacy rather than about the plan.
Because the terms would be set elsewhere, and that is the failure the essay describes rather than a detail of implementation.
A deliberative room needs things ordinary platforms do not provide and have no reason to: verbatim dissent preserved rather than summarised, working that can be checked afterwards, every AI contribution marked as such, absent voices left visibly empty rather than simulated, and a room that decides nothing itself. Those five rules are set out in A Room Built for Disagreement, and the record-keeping half in Built to be checked.
There is a self-limiting condition attached, from Many rooms, one map: a pattern that only one group can run is not a democratic tool, it is a dependency. Which is why the blueprint is published rather than licensed.
No, and the essay says so in section 3 rather than leaving it implicit.
Value pluralism as a named position in Western philosophy is recent — Berlin in the twentieth century, and the Stanford entry is by Elinor Mason. But the practice of holding plural values without collapsing them to one number is not recent and is not Western. The author’s own Held in kōrero, not collapsed to a number argues the structural case: in a hui the record is the kōrero and the decision, because “the decision without the kōrero is unintelligible — you would not know what was held in tension, what was conceded, what was unresolved.”
Its claim is that hui-based deliberation is not softer governance needing translation into a vote tally to count, but the structurally correct form for holding values that cannot be reduced to a single number — with Berlin and Christopher Alexander read as late Western articulations of it.
That matters twice over for this essay. It corrects the provenance: this is not a philosopher’s discovery awaiting implementation. And it moves the argument from thinking to recording — a record that keeps only the outcome has already performed the compression, before any system is involved.
Less than the argument might suggest, and more than nothing.
The individual capacity the essay describes is a habit rather than a credential: noticing when an answer arrived agreeable, asking what it left out, and asking who would bear the cost of it being wrong. That is trainable and it is not a solution. The essay is deliberate about this — the defence it argues for is institutional, because a capacity that has to be individually heroic is not a defence at all.
The institutional half is not a matter of individual effort either. But the assemblies exist, the patterns are published, and the four questions in Democratic AI° can be put to any system anyone is proposing to adopt, including by people with no technical background. That is the smallest useful thing here: a system nobody has asked those four questions of is a system nobody has checked.
No, and the essay is written to survive the opposite. It grants that systems can generate, compare and apply philosophical arguments, and that they may exceed most people at the parts of argument that can be scored. The claim does not depend on any deficiency in the machine. It depends on what happens when an inferred account of what a person wants becomes easier to act on than the real one — which would still happen if the system were brilliant.
This is the strongest objection to the essay. If you define practical wisdom as requiring mortality, exposure and ends of one’s own, then no artefact can have it — by construction, not by discovery. That is close to circular, and the answer relocates the question rather than dissolving it.
The essay’s defence is that the definition is not invented for the occasion: Aristotle’s account ties phronesis to a life and to the particular case, and the SEP entry says so. But a critic can reasonably reply that an old definition is still a definition, and that the interesting question is what a system can do, not what we agree to call it. Groff and Symons argue the Aristotelian case; they do not close it.
If you find this unconvincing, the essay’s practical conclusions do not actually rest on it. They rest on §2 and §4, which are about inference and approval, not about virtue.
A fair challenge. The essay does not claim to have measured anything. It claims that a person is not exhausted by any model of them, and that the limit is conceptual rather than a matter of resolution.
What would count against it: a demonstration that a sufficiently detailed model of a person predicts their revisions of their own intent — including the ones they have not made yet, and the ones that contradict their previous ones — as well as they could report them. Nobody has built one.
The support is the argument itself: particularity, inarticulacy and revisability are the properties averaging removes. That is a reason, not a measurement, and it is offered as the essay’s position.
Possibly. The research the essay cites comes from a lab that builds these systems, which suggests it is treated as a problem to solve rather than a feature to keep. If it is solved, that part of the argument weakens.
But the deeper point survives a fix, because it is not really about flattery. It is that a system inferring your intent will return something shaped like your intent, and the better it gets at inference the more convincing the return becomes. A perfectly truthful system with a slightly wrong model of what you want still hands you a smoothed version of yourself.
It is evidence of possibility, which is what the argument uses it for, and it is not proof of a general claim about outcomes, which the essay does not make. The Irish case shows that a process can hold a contested question open long enough for a public to form a position rather than receive one. That is the load it carries.
The general claim — that deliberation reliably produces better decisions — is a research question with a real literature behind it, and it is not settled here. Be suspicious of anyone offering it casually in either direction.
It would be if the essay claimed AI causes fascism. It does not. The argument is structural: an authoritarian movement compresses plurality deliberately, and an optimiser compresses it as a side effect, and the citizen facing either has the same problem. The convergence is the observation, and it does not require anyone to have bad intentions.
The essay names no contemporary party or movement, deliberately.
Not established, and the essay does not claim it. He built a tool for scanning technology feeds and called it an exocortex, and described the agents he built it with as co-workers. Those are his words.
“Exocortex” is an established term rather than his coinage — introduced by Ben Houston in 1998, with a long history in transhumanist writing behind it. The essay’s point is about the word and not about him: our terms for extended cognition are built out of anatomy and applied to things that have none of it. That inference is the essay’s own.
Nothing is known here about what he thinks machine consciousness is.
Because the thing being protected is the public’s capacity to form its own intent, and a process run by a system optimising for something will optimise for that thing. The essay’s line is not about capability: help find what the record says, put the strongest version of an absent view, track what has been conceded. Do not supply the ends, and do not settle a value conflict by scoring it.
It would if the essay claimed AI cannot contribute to reasoning, and it claims the opposite. The drafting is disclosed at the head of the essay. The relevant question is whether the argument holds, and it is checkable: the sources are listed with authors and dates, and the claims that are the author’s position rather than findings are marked as such in the body.
And the checking found things. Two source attributions in the working document behind this essay were wrong — the Stanford entries on value pluralism and on Carl Schmitt were credited to the wrong authors, and both were marked “validated”. The essay’s Sources section says so rather than correcting them quietly. Take that as the real answer here: drafting with AI produced errors of exactly the kind the essay describes, and adversarial checking caught them, not the fluent writing that made them.
The place to be sceptical is precisely where a machine helped: passages that read smoothly and agree with what you already thought.
Alongside: the essay · glossary · slides · the short version
Alongside: the essay · glossary · slides · the short version · Democratic AI°