James Madison

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI asks both the personal question—what should I become in the age of AI—and the structural one: how should power be arranged when it cannot be trusted to arrange itself? Madison is the cycle’s primary voice on the structural question. His method was empirical in the deepest sense: he studied the historical record of confederacies and republics, catalogued their failure modes, and designed against them—treating political architecture as a domain of progressive, cumulative knowledge rather than mere opinion. The same disposition, applied to the governance of artificial intelligence, is the disposition the cycle recommends.

His most immediate gift is the diagnosis of parchment barriers—the rules, principles, and voluntary commitments that declare limits without establishing any mechanism to enforce them. The landscape of AI governance is dense with such declarations: principles of responsible AI, ethics guidelines, safety commitments, statements of values. Madison watched their eighteenth-century equivalents fail. He would expect their modern descendants to fail the same way and for the same reason: a voluntary commitment that a company can abandon when competition demands is not a control at all. It is an angel’s promise, and Madison did not build for angels.

His theory of faction, articulated in Federalist No. 10, reads today like a diagnosis of algorithmic polarization written in advance. The recommendation systems of the past decade and the generative systems being built now are, by design and incentive, the most powerful engines for manufacturing faction ever devised: they collapse the friction that Madison’s extended republic relied upon, amplify outrage, and sort vast populations into intensely coordinated groups invisible to each other except as enemies. Madison’s response would not have been to ban the passions but to ask what new source of friction could perform the function his extended republic performed—making it hard for a single inflamed faction to capture the whole.

Where James Joyce diagnoses what the machine is, and James McClelland explains its mechanism, Madison diagnoses the governance failure that surrounds it. Together they form a trinity of honest seeing: mechanism, meaning, and structure—and Madison’s is the least comfortable, because it insists that structures matter more than the good intentions of those who build them.

Origin

James Madison was born in 1751 in the Virginia Piedmont, slight of build, bookish, and prone to what he called “epileptoid hysteria” under stress—a condition that kept him out of military service during the Revolution and left him free to think. He prepared for the Constitutional Convention of 1787 the way a scholar prepares for an examination he intends to dominate: he read everything he could find on the history of confederacies, ancient and modern, cataloguing their failure modes. He arrived in Philadelphia with a plan already drafted—the Virginia Plan that became the skeleton of the Constitution—spoke more than a hundred times over the summer, took the notes that remain the fullest record of the proceedings, and emerged having shaped the document more than any other single person.

With Alexander Hamilton and John Jay he wrote the Federalist Papers to win ratification; his contributions—especially No. 10 on faction and the extended republic, and No. 51 on checks and balances—are among the most penetrating things ever written about democracy’s dangers to itself. He drafted the Bill of Rights and guided it through a Congress that did not particularly want it, grounding the protections not in sentiment but in a theory of natural rights and the requirements of justice. As President he led the nation through the War of 1812. His enduring contribution was not policy but a structural theory of liberty: that freedom is secured not by virtue but by the careful arrangement of power against power. He died in 1836, the last Founding Father, aware that the republic he had designed was already being tested in ways he had not foreseen and could not have.

Key Ideas

Ambition must counteract ambition. The central design principle of Federalist No. 51—“Ambition must be made to counteract ambition”—is a principle about how to constrain any agent that optimizes for its own objectives without regard for the whole. It applies to human officials and, structurally, to the organizations building advanced AI. Madison did not trust actors to restrain themselves; he arranged the field so that the self-interest of each party would check the self-interest of every other. The hard engineering question for AI governance is what such an arrangement looks like: regulation that binds all parties at once, liability that makes recklessness expensive, independent oversight with real authority to compel and halt. Controls must be interior to the system’s incentives, not exterior exhortations laid on top of them.

James Joyce

The theory of faction. Madison defined a faction as a group “united and actuated by some common impulse of passion, or of interest, adverse to the rights of other citizens.” His insight was that even a majority can be a faction, and a majoritarian faction is the most dangerous kind. His mechanism for controlling factions was the extended republic: in a large territory with many diverse interests, the friction of distance and multiplicity would make it hard for a single passion to sweep the whole population. Technologies of attention have destroyed this friction entirely. A passion can now sweep a global population in hours. AI amplifies the danger by industrializing the production of factional content at scale.

Parchment barriers. A “mere demarcation on parchment,” Madison wrote, was not sufficient to guard against the concentration of power. A real control changes the payoff structure; it gives each guardian the means and the motive to defend its ground. Parchment barriers are rules with no force behind them—and the AI governance landscape is dense with them. His diagnosis applies precisely: a voluntary safety commitment a company can abandon under competitive pressure, an ethical guideline with no enforcement body, a principle carrying no consequence for its violation—all are demarcations on parchment.

The tyranny of the majority. The same critique that makes Madison suspicious of concentrated power makes him suspicious of the claim that aligning AI with aggregated human preferences is self-evidently good. A majority can prefer the oppression of a minority. Preference-based alignment, absent something like a bill of rights, faithfully reproduces whatever injustices the majority is prone to. He would recognize the missing piece in current AI alignment immediately: a feedback mechanism for preference, and no equivalent of inviolable protections placed beyond the reach of ordinary aggregation.

If machines were angels. Madison’s most famous sentence—“If men were angels, no government would be necessary”—is, read structurally, the statement of the alignment problem. An AI system that reliably wanted what we want would need no external constraint. The reason the field of AI safety exists is that we cannot verify such alignment. Madison’s mature counsel was not to reform the agents but to surround them with structures that constrain what they can do. If we cannot make AI systems reliably want what we want, we must make the system safe even if the agent inside it is not.