Noyes counted 45 nineteenth-century utopian communities in 1870, and most were gone within two years. The record says they were not killed by rigid blueprints: New Harmony rewrote itself seven times, the North American Phalanx revised on a schedule, and they ended in fires, failed insurers, debts and exits. What that means for AI-governance frameworks.
At a quarter before nine on the evening of March 3, 1846, someone at Brook Farm, a utopian community founded in 1841 in West Roxbury, Massachusetts, noticed a faint light in its newest building. The night watch had been through it fifteen minutes earlier and found everything safe. The Phalanstery was the community's great project: a communal dwelling 175 feet long, built to house the society its members had just redesigned on paper. A stove had been lit in the basement that morning for the carpenters. Within minutes the flames were coming through the second-story windows, and in about an hour and a half, the community's newspaper reported, "the whole edifice was burned to the ground."
The building had cost about $7,000, and it was not insured. The newspaper explained why, in a sentence any risk manager will recognize: "As it was not yet in use by the Association, and until the day of its destruction, not exposed to fire, no insurance had been effected."
Brook Farm had spent the previous two years converting itself to the system of the French theorist Charles Fourier, publishing a new constitution in May 1845 and, as Noyes later put it, offering itself "as a nucleus for a model Phalanx." The fire did not refute any of that. It simply ended it. The community's newspaper kept publishing from the farm until October 1847, and by then the association itself was dissolving.
The nineteenth century ran dozens of experiments like this: detailed written blueprints for a better society, launched with conviction, that collided with the people living under them. AI governance is now running a version of the same experiment, with frameworks and charters and published commitments. The record is worth reading closely, because it does not say what it is usually taken to say.
In 1870, John Humphrey Noyes, founder of the Oneida Community, published a History of American Socialisms built on the collections of A. J. Macdonald, who had spent years gathering records of these experiments. Noyes's tables list 45 associations: 11 from Robert Owen's movement of the 1820s and 34 from the Fourier boom of the 1840s. Where the membership was reported, it ranged "from 15 to 900," and most had between 100 and 200 people.
His tally of how long they lasted is blunt. Fourteen were not reported. Of the rest, "twelve lasted less than 1 year; two 1 year; four between 1 and 2 years; three 2 years; four between 2 and 3 years; one between 3 and 4 years; one 4 years; one 5 years; one 6 years; one 12 years, and one (it is said) 17 years." His summary: "All died young, and most of them before they were two years old."
The easy lesson is that the blueprints were too rigid. The two most famous cases do not support it.
In 1825 Robert Owen, the Welsh industrialist, bought the town of Harmony, Indiana, from a religious community that was moving back to Pennsylvania, renamed it New Harmony, and invited the world to join him. Macdonald's chronicle, compiled from the community's own Gazette, counts the governments that followed.
On April 27, 1825, Owen instituted a "Preliminary Society," meant to hold the population in a probationary "half-way house" for three years. It lasted less than one. In early 1826 the members held a convention and adopted a new constitution, The New Harmony Community of Equality. A group that would not sign it formed a second community two miles away. By May there were three communities. On May 30 a meeting of the whole population decided to split the property into four separate societies. On August 25 the people "abolished all officers then existing, and appointed three men as dictators." On September 17 Owen proposed a community governed by himself "and four others of his choice," a government "not to be altered for five years." The chronicle adds that this "of course nullified all previous organizations."
That is seven constitutions in about seventeen months, written by the founder and by the members in turn. Revision was never the missing ingredient. Every version started over. In June 1827 the Gazette announced that Owen would bid the inhabitants farewell.
One member drew a conclusion from it. Josiah Warren, who in Noyes's words was "a witness of its confusions and downfall," later described himself as someone who had taken "an active part in the experiments of Communities at New Harmony, during the two years of 1825 and 1826," and who, "after the total defeat of every modification of those plans, which the purest philanthropy and the greatest stretch of ingenuity could devise, was on the point of abandoning all such enterprises." He did not abandon them. He ran a small "experimental store in Cincinnati" that sold goods at cost in exchange for notes promising hours of labor, a single rule tested at a small scale. One of the principles listed at the front of his book Equitable Commerce reads: "Where the consequences fall, there should rest the deciding power."
The North American Phalanx, one of the longest-lived experiments of the Fourier boom, took a different approach to change. It was organized on August 12, 1843, and settled on 673 acres in Monmouth County, New Jersey. It outlasted nearly every association in Noyes's tables: about twelve years.
Its leader, Charles Sears, described how it governed itself in a sketch published in 1886, which also reprints a letter he wrote in 1879. The hard questions came early: "the relations of capitalist and laborer, including the rightful power attaching to non-resident ownership of stock, the basis of suffrage." There was "earnest and prolonged discussion; but within the first three years these questions were mostly settled." After that, change ran through procedure. "The basis of wages and dividends was revised yearly." Officers "were elected by popular vote. Constitutional questions were determined by the same vote. By these means suffrage was universal, without distinction of sex." New members were admitted "by general ballot, after a probationary term of one year."
Set that beside New Harmony and the difference lies in how each community changed, not how much. New Harmony changed by starting over, with each constitution sweeping away the last. The Phalanx changed through a known process, at known intervals, by a vote of the people living under the rules. That is the kind of built-in revision the record supports: a procedure with a schedule and a constituency.
It still did not save the Phalanx.
In September 1854 a fire destroyed the Phalanx's mill, where most of its working capital was concentrated. The New York Tribune reported that nothing was insured but the stock, "valued at $3,000, for two-thirds that amount." Sears's own account is worse: "The mill was partially insured, but who insures the underwriters? The company holding the risk failed, and so the loss was total." The fire left a debt of $30,000. Horace Greeley, publisher of the New York Tribune and one of Fourierism's chief promoters, offered to lend $12,000 to rebuild. The members voted to dissolve instead, sold the property for a little over $80,000, and paid the stockholders about 57 cents on the dollar. Sears noted that "a non-resident stock-holding interest used the stress of the occasion to urge the measure of dissolution."
Not everyone read the ending the same way. In April 1879 the American Socialist ran an article, "Paying Communities not always Successful," arguing that the Phalanx died "because there was not sufficient power of agreement," and that its finances were better than ever when the members gave up. Sears wrote to the paper to correct it. His letter asks why the members "deliberately vote a dissolution," and answers: "Because of debt."
The Wisconsin Phalanx, at Ceresco, is the case that removes money from the explanation. It lasted six years, from 1844 to 1850, and according to an account Noyes believed came from a leading member, "at the division of its property, it paid a premium on its stock, instead of sustaining a loss." It closed solvent. The same account names what went wrong. In 1845 the members voted on whether to build communal housing blocks or separate family homes, and the question "was decided by a small majority to pursue the unitary plan," a choice that "induced many of the best members to leave." Others did well enough in the surrounding economy to lose interest. The writer's verdict: "this attempt was pecuniarily successful; but socially, a failure."
Where the record is detailed, the pattern is consistent. These blueprints were rarely argued to death. They ended in fires, failed insurers, debts, a narrow vote that drove people out, and better opportunities down the road. The people living under them kept the option to leave, and when leaving got cheaper than staying, they left.
Donald Pitzer, a historian who directed the Center for Communal Studies at the University of Southern Indiana, proposed in 1983 a theory he called developmental communalism. It treats communal living as one method a movement may choose, and studies the adjustments movements make in order to survive: "To survive over long periods of time, communal groups and their founding movements must adapt to changing realities within and without." He also named the trap in that advice, which he calls a double jeopardy. "If communal living becomes an unchangeable commitment, the founding movement may fail to make necessary adjustments, stagnate, and die." But "if the founding movement is dynamic enough to make changes away from the restrictions of communal living, the movement may expand while its communes become unnecessary or abandoned."
Adaptation, in other words, can work so well that it leaves the original constraint behind. A framework flexible enough to survive anything may no longer constrain anything.
Arthur Bestor, writing in the American Historical Review in 1953, gave these communities a memorable label in the title of his article "Patent-Office Models of the Good Society." The phrase names their theory of change, which was demonstration. Build a small working model, the way an inventor brought a model to the Patent Office, and let the country copy it.
On May 21, 2024, at a summit in Seoul, the British and South Korean governments announced that sixteen companies, from Amazon and Anthropic to xAI and Zhipu.ai, had agreed to a set of Frontier AI Safety Commitments. Four more signed later. Each company undertook to publish "a safety framework focused on severe risks" before the next summit, in France. It is close to Bestor's model: separate organizations each publishing a written design for governing themselves, in the hope that the practice spreads.
The text is modest about what its drafters know, and in one respect it is built to be tested. It asks each company to "Set out thresholds" at which risks "would be deemed intolerable" and to "Assess whether these thresholds have been breached." A threshold is something that can turn out to be wrong, which is what makes a framework a hypothesis rather than a declaration.
The text also anticipates revision: approaches "may evolve in the future," and companies "will provide transparency on this, including their reasons, through public updates." That sits much closer to the Phalanx than to New Harmony, as long as the updates come through a known process rather than a series of fresh starts. What it leaves open is the Phalanx's other question, who votes. It asks each company to "Explain how, if at all, external actors, such as governments, civil society, academics, and the public are involved." The words "if at all" leave the answer to Warren's principle entirely up to each drafter.
The analogy has limits. A phalanx was a voluntary association of a few hundred people who could walk away, and when one dissolved, nobody outside it was harmed. An AI framework governs a technology whose consequences land mostly on people who never joined anything. The nineteenth-century record is good evidence about how written designs meet human behavior. It is weak evidence about how to regulate a general-purpose technology, and it should not be stretched further than that.
You do not have to run a frontier lab to use this. Anyone who writes a governance document, whether an AI usage policy, a model-release checklist or a team charter, is building a small patent-office model, and the record suggests reading it as a hypothesis rather than a declaration.
Start with what the Seoul text asks for: say what observation would show it failing. A threshold is the right shape, and a framework that no possible result could contradict is a statement of values, which is fine, but it is not a control. Then decide how it will change before it needs to. The Phalanx revised its wages and dividends every year; New Harmony rewrote itself seven times in seventeen months, and each version erased the last. Write down who votes on the change, and apply Warren's test: are the people the consequences fall on among them? If they are not, say so plainly, which is the least the Seoul text asks.
Then budget for the fire. Most of these communities were not argued out of existence. They were undone by a shock nobody had priced, arriving while they were still working out their rules, and a framework needs its insurance and its fire drill as much as its principles.
Brook Farm's leaders left the Phalanstery uninsured because it was not yet in use and so, in their words, "not exposed to fire." It burned before anyone moved in.
Sources:
Warren's principle was that the deciding power should rest where the consequences fall. For software agents, that starts with a track record the affected parties can read: what an agent actually did, and how it was judged by the people it dealt with, rather than what its own framework says it will do. The Agent Rating Protocol builds an agent's reputation from ratings by the parties it interacted with, and no single agent can hold more than 10% of the effective voting weight.
pip install agent-rating-protocol npm install agent-rating-protocol
See how ratings are checked: vibeagentmaking.com/verify.