Artificial Persons
From Leviathan to Large Language Models
Human beings make systems that then move their creators, without those creators quite knowing how they are being moved.
Thomas Hobbes imagined the state as an artificial person conjured from the aggregated wills of individuals. He described it as an artificial man, with sovereignty as an artificial soul, and equity and laws as artificial reason and will. He called it Leviathan.
What Hobbes was reaching for can be thought of as a kind of analog precursor to artificial intelligence: an intelligence manufactured by human beings that then outgrows them. The state processes information and issues commands. It acts through people, but it is not identical to any one of them. Its citizens may submit to it, rebel against it, question whether it is even real, feel shame for being part of it or having betrayed it, or die for it, all without being able to trace its boundaries back to a single human mind, including their own.
For centuries, corporations have held legal personhood. They have had rights that, in some cases, were still denied to actual human beings. A corporation can be controlled, and it can get out of control. It operates across decades and continents. It can act with extraordinary coherence. It does not need emotions in order to have interests. Its employees come and go, the public’s perception of it shifts, yet the institution persists, adapting and defending itself, maximizing the conditions for its survival.
Markets are a stranger case. States have sovereignty and corporations have agency; markets have neither, yet they still compel us through their invisible hand.
Prices perform a function analogous to cognition: they condense dispersed knowledge into signals that coordinate action without any central mind. Hayek pointed out that no central mind is doing the thinking, and treated this as a defense of markets. A system can coordinate human action without being commanded by any human being, and when its internal logic generates consequences no participant intended, there is no single agent to hold responsible.
The frameworks we build for accountability tend to be designed around identifiable bearers of responsibility. Criminal law asks who acted and with what intention; it seeks to establish whether the person before the court is continuous with the person who committed the act. Contract law binds parties that can be located and compelled. Over time, law learned to extend these tools beyond natural persons. The corporation became a legal person: a center of rights and liabilities without a natural body. But that extension already required fiction and compromise. The corporation can be regulated, but its actions are still distributed. The market is harder still; it coordinates conduct and produces consequences but without a single entity standing behind those consequences. With AI we face a similar difficulty and for similar reasons: our inherited tools of accountability assume that responsibility can be gathered somewhere. When the center is not clear, the tools begin to strain.
Consequently, the record of human attempts to govern such entities is a record of partial successes. Constitutions have constrained states that also built empires and committed atrocities. Antitrust law checked monopolies, and labor unions extracted protections that markets could not provide on their own. In each case, governance worked the same way: we found something the entity could not survive without and made continued access conditional on behavior. States need legitimacy, or at least compliance. Corporations need labor, customers, capital, and legal permission. Markets need trust, enforced at their edges by institutions that are not markets. These dependencies were not obvious in advance; they were discovered through confrontation.
Will that method of discovery work with AI? Constitutional traditions developed through centuries of crisis and correction: the entity overreached, people suffered, institutions adapted, and the adaptation was tested by the next overreach. Governance, in other words, is often learned from failure, and learning from failure requires that failures arrive slowly enough to be survived and studied. The artificial persons of the past granted us that time by their nature. AI might not.
We have centuries of practice in living with intelligences larger than ourselves. We also have some precedent for controlling harmful self-replicating entities, as with pandemics, and dangerous technologies, as with nuclear weapons. That history lets us ask what the new entity needs, and how to make meeting those needs conditional upon our well-being. Our public discourse already reflects the candidates: compute, data, energy, capital, infrastructure. Our advantage is that we can still see what it needs; our challenge is that, if current advances continue, this advantage may be temporary.
Continue Reading:
→ Dawkins, Claude, and the First Question About Consciousness
→ Bryan Caplan on Ethical Intuitionism (podcast)
Other Projects:
→ Universal Open Textbook Initiative (free, multilingual textbooks)