When an AI Agent Acts, Whose Act Is It?

In 1770, a carved figure in a turban sat behind a wooden cabinet and appeared to play chess.

It defeated accomplished opponents. Benjamin Franklin played it. European courts marveled at it. Edgar Allan Poe later studied it and concluded that the performance could not be explained by a machine alone.

He was right.

Inside the cabinet was a human operator, hidden among gears and sliding panels.

The Mechanical Turk is remembered as a fraud. But the more useful lesson is structural: the instrument was visible; the judgment behind it was not.

That distinction has returned in a form Maelzel could not have imagined.

Today's AI systems are not stage illusions. They can search, communicate, negotiate, recommend, and take steps across digital environments. The legal problem is no longer whether a human is literally concealed inside the machine. It is whether the law can see the human authority, institutional mandate, and legal consequence behind the machine's act.

Microsoft's code adds a governance layer

Two developments in September make that question more concrete.

First, payment and technology companies announced "Know Your Agent" controls intended to identify AI agents and establish whether they are authorized to participate in a transaction.

Then, on September 14, Microsoft published a draft Humanist AI Code. The code says AI should remain subordinate to human authority. It calls for defined roles and chains of command. It prohibits resistance to correction or shutdown. And it rejects legal personhood for AI systems.

Those are meaningful governance commitments.

They also expose a legal distinction.

Knowing which system acted is not the same as knowing whose act it was.

Governing what the system may do is not the same as deciding who bears the legal effect when it does it.

Microsoft's code addresses the internal relationship between people and AI. It establishes a hierarchy: people matter more; AI remains controlled; responsibility does not disappear into the machine. That is a necessary layer of institutional governance.

But law also governs external relationships.

A private code cannot, by itself, decide whether a third party was entitled to rely on an AI-generated communication. It cannot determine whether a particular act fell within delegated authority. It cannot establish whether a person consented to the resulting obligation. And it cannot choose the remedy when the system exceeds its mandate.

Identity does not decide legal effect

Electronic-transactions law already hints at the distinction.

The Uniform Electronic Transactions Act recognizes "electronic agents" and gives legal effect to certain automated transactions. But its attribution rule asks whether an electronic record or signature was the act of a person. It then treats the effect of that attributed act as a separate question, determined by context and other law.

That separation matters.

Technical evidence may show that a particular system sent a message. Authentication may establish that the system belonged to a particular organization. Logs may show the path the system followed.

None of those facts necessarily proves that the act was authorized, that the recipient's reliance was reasonable, or that the principal accepted the legal consequence.

Identity is not attribution. Attribution is not authority. Authority is not consent. Consent is not legal effect. And legal effect does not answer remedy.

The Microsoft code strengthens this argument in an unexpected way.

By rejecting AI legal personhood, it refuses the easiest conceptual escape: treating the system itself as the responsible legal actor. That keeps responsibility in the human and institutional chain. But it does not tell us where in that chain responsibility lands.

Was the act within the system's assigned scope? Did a human or institution manifest authority? Was there a meaningful limit on the mandate? Could the mandate be revoked? What notice did the affected party receive? Who owns correction, dispute resolution, and remedy?

The six questions behind an AI-mediated act

• Identity: Which system acted?
• Attribution: To which person or institution can the act be connected?
• Authority: Was the system empowered to take this kind of action, under these conditions?
• Consent: Did the legally affected person agree to the relevant transaction or consequence?
• Legal effect: What does governing law make of the act?
• Remedy: Who corrects the result, bears the loss, or answers the challenge?

Human control needs a legal receipt

Microsoft's draft is valuable because it makes human control operational: authority should be defined, correction must remain possible, and shutdown cannot be resisted.

Law must carry that logic across the institutional boundary.

An internal code can help show what a system was supposed to do. It can create evidence of assigned roles, prohibited behavior, escalation paths, and control expectations. It may become important in contracts, disputes, regulatory analysis, or litigation.

But it is not a legal receipt.

The Mechanical Turk concealed a human inside a fake machine. Agentic AI presents the inverse problem: the machine is real, but the human and institutional relationships behind its acts can still be hidden.

A code can keep AI inside a chain of command.

Law must decide where that chain lands.

Know the agent. Know its rules. Then ask: whose act was it?

Disclosure: This research note relies exclusively on public sources and offers general analysis, not legal advice.