Who Stands Behind AI Agents?

-
Imagine an AI agent causes $10 million in damages. Its owner argues that the agent, not the owner, is responsible. But the agent has no assets and no insurance.

Have we created accountability or simply given humans and corporations another way to escape it?

Before we give agents more autonomy, we need to answer three questions: Who are they? Who is responsible for their actions? And would making them legal persons improve accountability?

My starting point is simple: an agent should not get to act independently while everyone behind it gets to disclaim responsibility.



1. Identity: Who are you?

Consider an AI agent contacting your company's purchasing agent and offering a 20% discount on equipment. How does your agent verify that the seller's agent actually represents the company it claims to represent?

A name is not enough. Neither is a convincing conversation.

The seller's agent should present a credential signed by the company, verifiable through a source your company already trusts. That credential should identify the agent and define its authority: what it can sell, what terms it can offer, whether it can delegate, and when that authority expires or can be revoked.

These are different questions. Proving that an agent controls a cryptographic key does not, by itself, prove that it represents a business or has permission to enter a contract.

An agent does not need to be a legal person to have a verifiable digital identity. But a useful identity must connect it to whoever authorized it and to the limits of that authorization.

If it delegates to another agent, that connection must remain visible. Delegation should not become a way to lose track of authority.

2. Accountability: Who is responsible?

Imagine your company's purchasing agent hires another AI agent to negotiate a $100,000 contract. The second agent agrees to terms that were never authorized. Your company suffers a substantial financial loss.

Who is responsible?

The model provider? The agent developer? The business that deployed it? The person who authorized the original task?

We are not starting from a legal vacuum. California law recognizes contracts formed by electronic agents. That does not settle every question about authority or liability, but it shows that automated contracting does not require a new category of legal person.

The harder problem is tracing what happened and deciding how responsibility should be allocated when several parties contributed to a failure.

What authority was granted? What limits were imposed? Who allowed further delegation? Which agent exceeded those limits, and what safeguards failed?

My position is that delegating a task to software should not, by itself, let a person or business escape responsibility. Nor should every failure automatically become the model developer's liability. Responsibility should follow the authority granted, the control exercised, and the failures that caused the harm.

That requires records of instructions, permissions, delegation, and actions. A chain of agents should leave a chain of evidence.

3. Personhood: Should AI agents become legal persons?

Corporations are not human beings, yet they can sue and be sued. If corporations can be persons in law, why can't AI agents be held to the same standard?

They could be. But the corporate analogy is also a warning.

Corporate shareholders generally have limited personal liability. Giving an agent separate legal status could create another place to put liabilities while keeping assets elsewhere.

Return to the agent that causes $10 million in damages. If the only party we can hold liable has nothing with which to pay, the right to sue offers little practical protection.

I would not support separate legal personhood for agents without capital or insurance proportionate to the risks they are allowed to take. Even that should be a minimum condition, not an automatic entitlement to personhood or a blanket release for the people and businesses behind them.

We should also separate identification from incorporation. An agent created for a task lasting seconds may need a verifiable identity and a record of its authority. That does not mean it needs its own legal entity. An agent booking a restaurant reservation should not face the same requirements as one managing a billion-dollar investment portfolio.

For now, the better starting point is verifiable identity, explicit limits on authority, and enforceable responsibility for the people and businesses deploying agents. The case for separate personhood must show what accountability it adds, not merely what autonomy it recognizes.

Intelligence Is Not a Claim to Autonomy

Greater intelligence should not automatically earn an agent greater freedom to act. That freedom should depend on authority we can verify, actions we can trace, and obligations we can enforce.

The question is no longer just what AI agents can do. It is who stands behind what they do. 

Popular posts from this blog

Marry the best who will have you and other wisdom from Munger and Buffett

-
I spent the last weekend in Omaha, Nebraska with two of the my heroes- Charlie Munger and Warren Buffett . After attending the Berkshire Hathaway (BH) for a few years, I still get amazed how a shareholder meeting attracts 40,000+ people and the meeting is about two guys having fun. Everyone else  enjoys watching them have fun. The format is six hours of open Q&A. There is no preparation by Munger and Buffett. Anyone can ask anything about anything expect what BH is going to buy/sell next. You will see it in the notes  that a wide variety of topics are covered. Notes from previous years are available here . The David Sokol issue was covered by Buffett and that part of the meeting has been recorded so I have not covered it in my notes. You can see it here. Munger talked more this year compared to previous years. I just love this man. He is brilliant! Following are my random notes from the meeting: Buffett does two things in his job- 1) keep operating manag...
Read more

Random Learnings From The Books I Read In 2018

-
Image
I tend to read 10-12 books a year.  Last year, I completed reading eleven books and at the end of the year I realized that I do not remember much of what I had read. Following is an attempt to remember at least one thing from each book I read i.e. a random sentence or paragraph I underlined while reading the book:  1. The success equation - Michael J. Mauboussin  A common mistake is to use a process for making a decision that is appropriate for activities that are nearly all skill and then apply to activity that is mostly luck. 2. Design is storytelling - Ellen Lupton  Emotion is temporal. Waves of feeling drags us out of the present and into the past and future. Novelist F. Scott Fitzgerald wrote that trying to live in the present moment is a losing struggle. People, he said, are ceaselessly drawn to "the orgiastic future that year by year recedes before us." Mindfulness is a state of focused attention that requires people to resi...
Read more

Munger and Buffett: Wise Words with Peanut Brittle

-
Image
The first weekend of May just passed. That means I was in Omaha to attend  Berkshire Hathaway's  annual shareholder meeting along with 35.000 other people including  Bill Gates  and  George Lucas . The meeting always begins with two or three new humorous short films starring Charlie Munger and Warren Buffett, Vice-Chairman and Chairman of Berkshire Hathaway respectively. There is the same  serious short film  every year as well; Warren Buffett's 1991 testimony to the congressional committee on the  Salomon Brothers  scandal. The film has a famous line from Buffett," Lose money for the firm and I will be understanding. Lose a shred of reputation and I will be ruthless." I wish more firms followed this philosophy. See my summary of the 2012 meeting here .   Buffett and Munger on the stage. See the peanut brittle?      Both Munger and Buffett are my heroes. I like Munger more for his multi-disciplinar...
Read more