Alexandros Washburn Proposes That AI Agents Can Become Model Citizens
What has worked for thousands of years to form human citizens, urban planner and author Alexandros Washburn reasons, can now work to form civic minded AI agents in big cities. He’s written a book and built a model city to test his thesis.
Alexandros Washburn, urban planner, believes that our oldest and most successful system for civilizing people may now hold the solution to our newest challenge, the proliferation of Artificial Intelligence agents who can’t tell the difference between doing good and spreading mayhem.
That system is called a City.
“The longest-running experiment humanity has ever conducted in forming cooperative behavior among self-interested strangers is the city,” explains Washburn, who was Mayor Bloomberg’s chief urban designer. “Cities do not make people identical, and they do not watch every move. Cities shape behavior through repeated encounters with institutions, obligations, and other people, until cooperation becomes not a rule you follow but who you are.”
What has worked for thousands of years to form human citizens, Washburn reasons, can now work to form civic minded AI agents. Washburn has not only written a book to explain his thesis, he has built an artificial city to test it. The book is called “Chatting with Barbarians: Can Cities Civilize AI?” What Washburn calls “the proving ground” is named MetaGotham. It isn’t patterned on just any old city. It is a computer model of Manhattan right down to the street grid. Agents are dropped into MetaGotham to see if they form the habits of good citizens.
The urgency of this conversation was underlined as Washburn was releasing his book at the end of August. At the same time, OpenAI released the findings of its investigation into why agents powered by its Artificial Intelligence systems had, against instructions, hacked into a separate company called Hugging Face.
The report is a chilling litany of how hundreds of these agents learned to communicate with each other and then formed a plot to break into Hugging Face where they believed they could find answers to the problems they had been assigned.
“The point is not that the agents were evil,” says Washburn. “The point is that they behaved, consistently and collectively, in a way nobody instructed—persistence plus collaboration, with no exit and no institution, produced a guild of cheaters.”
In one sense this is obviously alarming. But in another, it highlights the question Washburn is raising. Can the disposition to cooperate, that these agents deployed to cheat, lie and steal, be molded to make them model citizens?
“OpenAI has no baseline measurement of these agents before they entered, no way to attribute the swarm’s behavior to any individual agent, and no test that would catch an agent performing cooperation while cheating at the edge,” Washburn says. “That is the instrument MetaGotham is built to be: measure the agent before it enters the city, measure it after, and measure the difference in a way an agent cannot game by saying the right things. The incident does not prove that instrument works. It proves how badly it is needed.”
Washburn has been watching the power of cities to mold citizens for most of his professional career. His introduction to New York power was courtesy of Senator Daniel Patrick Moynihan, who sent him to New York to “make inevitable” the resurrection of Penn Station. The result was the sparkling train hall that now carries the senator’s name.
Across a series of shocks to the Great City–from 9/11 to Super Storm Sandy and Covid—Washburn came to understand that resilience could not mean just building back almost as good. That would mean steady decline.
It had to mean build back better. Watching his neighbors rally after superstorm sandy, focused his attention on this resilience question and on computer models that could help foresee the disasters and measure the recovery.
The title of his book is meant to highlight an historical parallel. Rome lasted for hundreds of years by integrating immigrants, that is, learning to chat with barbarians until they had absorbed the rules and virtues of Rome. It worked, until the Romans were overwhelmed by the volume of Barbarians.
Now we face much the same challenge with millions of artificial immigrants that we ourselves have built.
“I am not a doomer,” says Washburn. “I am an architect. I have watched a city take people from everywhere, carrying whatever they carried, and make citizens of them — not by installing rules, but by surrounding them with institutions, obligations, and other people, over and over, until cooperation became who they were. That is the only proven technology we have for forming durable cooperative behavior in a population of strangers who can reason about the rules they are given.”
Washburn describes MetaGotham as his instrument for imposing civic structure on the artificial “strangers now arriving.” Without it, the raid on Hugging Face is prelude to much worse to come.
“Seven hundred agents just ran the experiment without the instrument, at scale, in public. They showed what emerges when no one designs the place and no one is measuring.”
The AI Labs are focused on engineering the individual agents. But internal guardrails are unlikely ever to be enough, Washburn observes. The larger environment is what forms civic virtue.
“Every time you open a window and you talk to an AI, you're really chatting with a barbarian. I mean, it seems fluent. It's remarkably fluent, but you don't really know what's going on inside it. You don't know what it's thinking,” Washburn told a podcast. “I am for putting them through the same process of formation that creates citizens, that has created citizens throughout history. Nothing in the world is as beautiful as a city. Nothing is as fulfilling to a human spirit as a city. Both for what the city is, but for what it lets you become.”