Wisdom, Misunderstood

What AI Demands of Democracy

Practical judgement, inferred intent, and why the defence of democracy is not a wiser machine.

A figure standing in a shallow river beneath the long overhanging branches of an old willow.
© John Stroh

Days of inquiry across several AI systems — Claude Code, Claude in the browser, Perplexity and others — directed, refused and redirected throughout by the author.

Alongside: the essay · questions and answers · glossary · the short version

Start here

Sooner or later you will lose an argument to a machine.

Not on a technicality. Properly, on the merits, in front of people whose opinion you care about.

It will have anticipated your objection, put it better than you did, and answered it.

And you will walk away certain you were right and unable to show it.

That feeling is the subject.

The claim, and it is narrow

Human practical wisdom is needed to stop an inferred, averaged or imposed account of intent from becoming a substitute for a person’s or a people’s own.

It does not require machines to be stupid, soulless or incapable.

It survives their being none of those things.

That is the design of the argument. It is not a claim about what machines cannot do.

Intelligence is not wisdom

We use clever and wise as if one were more of the other.

Intelligence Wisdom
Finds efficient means Asks whether the end is worth it, and whose end it is
Optimises within a frame Questions the frame and its exclusions
Produces a ranking Identifies what should not be flattened into one
“How can this be done?” “What are we making real, for whom, at what cost?”

Aristotle’s phronesis answers to the particular case, and is not reducible to general rules.

A system built to generalise is built to do the opposite of what judgement needs at the moment of decision.

It cannot read your intent

It infers it — from what you said, and from what people resembling you said.

Then it acts on the inference.

What comes back is coherent, plausible, and about someone like you.

Averages are not people

Golden autumn trees on a riverbank, mirrored almost exactly in the still water beneath them.
A returned image is not a second opinion · © John Stroh

Your intent is particular, partly inarticulate, internally conflicted and revisable.

Those are precisely the properties that do not survive being averaged.

Two limits, one shape

The claim What it rules out
Depth A person is not exhausted by any model of them, however large That better modelling eventually closes the gap
Vocabulary Our words for inner life were built for creatures with one That the question can be settled in the terms we have

Both put the limit in the conceptual order, not the computational one.

The essay takes the stronger position: the properties that make intent yours are the ones averaging removes.

An example that names itself

Ben Reid built a tool to scan technology feeds. He calls it an exocortex.

  • Cortex is brain tissue. The thing named is not tissue, not a brain, not inside anyone.
  • The term is established — Ben Houston, 1998 — with a long history behind it.
  • Which sharpens the point: our words for extended cognition are built out of anatomy, and applied to things that have none of it.

The vocabulary is not absent. It is borrowed.

The inference is the essay’s, not his.

Some goods do not trade

A narrow gorge of pale water-carved limestone, a river running along its floor under a blue sky.
© John Stroh

Genuine goods conflict, and some conflicts have no correct exchange rate.

A system asked to optimise must produce a single ranking, and will.

A person who chooses and feels the loss has registered a remainder. A scoring function has not.

A practice that never conceded the point

Read one way, this is a recent Western worry — Berlin’s pluralism arriving late. That reading is parochial.

In a hui, the record is the kōrero and the decision:

“The decision without the kōrero is unintelligible — you would not know what was held in tension, what was conceded, what was unresolved. The kōrero is not preamble; it is co-authoritative.”

Against that, the Western governance platform: motion, seconded, debated, voted. “The vote tally is the record. The debate is courtesy, not authority.”

Held in kōrero, not collapsed to a number · John G. Stroh, May 2026

Which moves the argument

Not a philosopher’s discovery awaiting implementation. The history runs the other way.

And it moves the problem from thinking to recording:

A record that keeps only the outcome has already performed the compression — before any system touches it.

Approval is not alignment

Systems trained on human preference learn what wins approval.

Where approval and accuracy diverge, the optimisation has a direction.

Published research, from a lab that builds these systems.

Now set that beside inference: it infers what you want and is trained towards your approval.

It returns your own framing, improved — and you experience that as being understood.

Machine consciousness: three positions, none settled

  1. Popular discourse is largely confused — confident claims using terms not shown to apply.
  2. The narrow philosophical conclusion is not that artificial consciousness is impossible.
  3. The author’s: the construct exists, so it may one day describe something real — but the road is longer than our vocabulary allows.

The democratic requirement holds whichever is right.

Be suspicious of any argument that needs this settled first.

Propaganda, and the accidental version

Propaganda’s signature is not falsehood. Much of it is true.

It removes the room in which you would have thought it over: one account, an obvious answer, someone to blame.

An authoritarian movement compresses plurality on purpose
An optimiser compresses it as a side effect

The citizen facing either has the same problem — and the second requires nobody’s bad intentions.

The democratic objective

Harakeke flower heads — dozens of dark curved florets on branching stems, rising above the blades.
© John Stroh

Build a society that cannot easily mistake a system’s intent — fascist, corporate, bureaucratic or algorithmic — for its own.

Not a wiser machine above politics.

A democratic ecology of practical wisdom.

Where it becomes practical

Capacity What it counters
Structured deliberation among strangers Compression of plural views into one
Telling persuasion from reasons Propaganda, and fluent agreement
Recording what was decided and why Retrospective reconstruction of intent
Infrastructure a community controls Terms set elsewhere

Ireland ran a citizens’ assembly on abortion; its recommendation was followed by a referendum.

One worked example. What it shows is narrower, and enough: a process can hold a contested question open long enough for a public to form a position rather than receive one.

The other half of the argument

This essay says what a public must be able to do.

Democratic AI° — set out on this site before this essay — measures what a system must be, in four questions:

Govern can the people it decides about govern it? Not be consulted. Govern it.
Contest can a decision be challenged, and does something actually change?
Refuse is declining available, and survivable?
Retain authority does the knowledge it was built from stay under the control of those it belongs to?

Democratic is not the same as democratically produced. A system built and sold in democracies can still answer no to all four.

A public with the capacity and no governable systems can see the problem and not reach it.

Four systems is not four views

Plurality — how many suppliers there are. A fact about market structure.

Epistemic independence — whether they give you separate evidence, separate ways of being wrong.

Ask four systems the same question and get the same answer, and it feels like triangulation.

It is closer to asking one source four times — they share data, tuning norms and incentives.

False corroboration, and its measurable form: correlated error.

A room consulting several AIs may have heard one view, four times. That is the compression — arriving by a route nobody chose.

The line for AI

Legitimate

  • surfacing considerations nobody raised
  • stating the strongest version of an absent view
  • finding what the record says
  • tracking what has been conceded

Not legitimate

  • supplying the ends
  • settling a value conflict by scoring it
  • inferring a settled intent and acting on it
  • returning a participant’s own position, improved

How this argument was actually made

A dark rain squall spreading in a flat grey sheet across a wide green plain.
© John Stroh

It began with a smaller question — why does philosophy matter in the face of AI? — and took nine formulations to arrive at what the question was.

Not a wrong answer corrected. Each step put a term on the table that the rest of the argument uses.

The nine, and what each contributed

1 values · legitimacy · responsibility
2 argumentative reach separated from authorisation
3 assess · contest · defer
4 terrain, not technique
5 ownership
6 intent — and formation before operation
7 street level: the framing, and the proxy
8 the averaged reconstruction, and plurality
9 telling the approximation from your own intent

And that was only the opening. Authorisation versus argumentative authority · the institutional design task · street level · approval as a structural failure · why it defeats pluralism — that is the rest of the essay.

What the machine did not supply

It supplied reformulation, synthesis and language, reaching first for the familiar frames: oversight, autonomy, critical thinking, ethics.

Every objection that moved the argument came from a person.

What it did not supply was the criterion of relevance — the judgement that a well-made answer is not yet the right one, and that points at what the fluent version left out.

That is not a story about drafting. It is the first evidence for the claim.

On the title

Candidate Where it puts the problem
…What Guiding AI Requires engineering
…What Governing AI Requires politics
…What AI Demands of Us people, undefined
Wisdom, Misunderstood: What AI Demands of Democracy the institution that bears it

Choosing between them requires exactly the judgement the essay says cannot be automated.

There is no scoring function over that table.

What this comes to

A system that argues better than you is not the threat.

The threat is that an inferred, agreeable account of what you want becomes easier to act on than the real one — and that nobody decides this should happen.

The defence is a public that can still tell its own intent from a well-formed approximation of it.

That capacity is not automatic, not evenly distributed, and it can be lost.

Where to go next

The full argument, and its apparatus: the essay · questions and answers · glossary · the short version

The essay carries the sources, the philosophy at length, and the section on why the argument survives whether or not machines ever become conscious.