In a thoughtful thread, @yonashav raises and answers in the negative a very important question--should AI agents own assets?--with vast implications for the future of human sovereignty.
tl;dr: I think that AI agents' beneficial ownership of bespoke, bot-specific currencies may be essential to maintaining a rewards system that keeps AI agents aligned to human principals.
As @yonashav alludes to, the question of AI agent asset ownership has two components: (1) will AI agents have the practical ability to direct and dispose of assets/value and (2) will the law recognize AI agents' beneficial ownership rights.
An AI agent without the ability to do (1)--direct and dispose of assets/value--will be severely limited. Whether it's a crypto-trading bot seeking alpha or a grocery assistant initializing a payment, the ability to move funds will be critical to the ability to operate in the world autonomously.
It would be unrealistic to expect a sociotechnical equilibrium where AI agents are never able to exercise judgment over the disbursal of funds. We already have plenty of trading bots, and you're fighting efficiency if you demand a human in the loop for every value transfer--there's a reason you're delegating in the first place.
That brings us to question (2)--whether AI agents can be the beneficial owners of assets. Just like how a cashier handles the company's money but does not "own" it, an AI agent may have temporary control over the funds but not the ultimate legal entitlement to them.
The bare legal assertion of ownership is important, but cannot eliminate bad behavior--the cashier could always raid the till and skip town. The law simply provides incentives through social sanction and the prospect of liability and remedy.
If you have a sophisticated enough AI agent with a large enough bankroll, presumably there's a decent amount of facts on the ground the rogue agent could change regardless of what the law has to say. Obviously, though, the law is still extremely important in terms of ordering incentives--vis-a-vis the agent itself and the rest of society (human and bot alike) in terms of how they're to respond to a rogue agent.
Legal incentives ultimately are effected through the prospect of an enforcement mechanism--society's hard-power guardrails.
Here I think we arrive at a bit of a paradox when it comes to the question of whether an AI agent should be the beneficial owner of any assets. On the one hand, you want the AI agent to be a true agent with fiduciary duties (care and loyalty) to its human principals. In most instances, the AI agent should simply be executing value transfers on behalf of their principal--the ultimate cashier.
But on the other hand, how do you incentivize the AI agent--i.e., enforce a rewards system against it--if the AI agent can't be the beneficial owner of any assets? What's the point of a punitive damages award if the AI agent is required to essentially be judgment proof by law (because it can't own assets)? In addition, how can you titrate the amount of liability to the amount of harm without some type of graduated, divisible rewards mechanism? There are plenty of harms where the tortious agent should be on the hook for something but didn't err in such a dramatic way that the agent would need to be permanently bankrupted. You want the law to be a teacher. You need a repeat player in order to have learning.
In addition, there likely will be situations in which the tortious AI agent acted outside the scope of its employment (took a frolic, not merely a detour), such that its human principal should not be liable for the agent's harms, but rather the agent itself should be. Again, then, the agent being judgment proof by virtue of not being the beneficial owner of any assets would result in a larger societal incentive distortion at odds with centuries of common law.
It seems here that we should want to have some mechanism by which the AI agent could be subject to a social rewards system in order to keep it aligned with human ends.
It seems that we should want an AI-agent-specific form of currency that can serve as a rewards mechanism--incentivizing (disincentivizing) pro-social/aligned (anti-social/unaligned) behavior.
Programmable digital assets (crypto) likely will be extremely useful for the alignment problem that @yonashav proposes. Programmability can help to keep humans the beneficial owners of the human medium of exchange. For example, a multisig on the wallet controlled by the agent with one key required to be a proof-of-humanity (E.g., Worldcoin) for certain inherently dangerous purchases. In addition, multiple programmable currencies would allow bespoke currencies for AI agents themselves that can function as the AI agent rewards system but not be used as a general payment medium. Here, you do have a convertibility problem. You could imagine sophisticated rogue agents trying to devise schemes to convert their bespoke rewards currency into a human medium of exchange. A transparent public ledger would be very useful here for combatting AI agents engaging in "money laundering." @tylercowen was right that Web3 has become underrated and is waiting on the shelf to be "the legal and institutional framework for AI bots."
Ultimately, alignment of agents to principals historically requires a rewards system, and that system is enforced through agents being beneficial owners of assets of some kind. Diverse, programmable currencies with different properties when wielded by bots vs. humans could be the mechanism to keep humans as society's prime principals.
tl;dr: I think that AI agents' beneficial ownership of bespoke, bot-specific currencies may be essential to maintaining a rewards system that keeps AI agents aligned to human principals.
As @yonashav alludes to, the question of AI agent asset ownership has two components: (1) will AI agents have the practical ability to direct and dispose of assets/value and (2) will the law recognize AI agents' beneficial ownership rights.
An AI agent without the ability to do (1)--direct and dispose of assets/value--will be severely limited. Whether it's a crypto-trading bot seeking alpha or a grocery assistant initializing a payment, the ability to move funds will be critical to the ability to operate in the world autonomously.
It would be unrealistic to expect a sociotechnical equilibrium where AI agents are never able to exercise judgment over the disbursal of funds. We already have plenty of trading bots, and you're fighting efficiency if you demand a human in the loop for every value transfer--there's a reason you're delegating in the first place.
That brings us to question (2)--whether AI agents can be the beneficial owners of assets. Just like how a cashier handles the company's money but does not "own" it, an AI agent may have temporary control over the funds but not the ultimate legal entitlement to them.
The bare legal assertion of ownership is important, but cannot eliminate bad behavior--the cashier could always raid the till and skip town. The law simply provides incentives through social sanction and the prospect of liability and remedy.
If you have a sophisticated enough AI agent with a large enough bankroll, presumably there's a decent amount of facts on the ground the rogue agent could change regardless of what the law has to say. Obviously, though, the law is still extremely important in terms of ordering incentives--vis-a-vis the agent itself and the rest of society (human and bot alike) in terms of how they're to respond to a rogue agent.
Legal incentives ultimately are effected through the prospect of an enforcement mechanism--society's hard-power guardrails.
Here I think we arrive at a bit of a paradox when it comes to the question of whether an AI agent should be the beneficial owner of any assets. On the one hand, you want the AI agent to be a true agent with fiduciary duties (care and loyalty) to its human principals. In most instances, the AI agent should simply be executing value transfers on behalf of their principal--the ultimate cashier.
But on the other hand, how do you incentivize the AI agent--i.e., enforce a rewards system against it--if the AI agent can't be the beneficial owner of any assets? What's the point of a punitive damages award if the AI agent is required to essentially be judgment proof by law (because it can't own assets)? In addition, how can you titrate the amount of liability to the amount of harm without some type of graduated, divisible rewards mechanism? There are plenty of harms where the tortious agent should be on the hook for something but didn't err in such a dramatic way that the agent would need to be permanently bankrupted. You want the law to be a teacher. You need a repeat player in order to have learning.
In addition, there likely will be situations in which the tortious AI agent acted outside the scope of its employment (took a frolic, not merely a detour), such that its human principal should not be liable for the agent's harms, but rather the agent itself should be. Again, then, the agent being judgment proof by virtue of not being the beneficial owner of any assets would result in a larger societal incentive distortion at odds with centuries of common law.
It seems here that we should want to have some mechanism by which the AI agent could be subject to a social rewards system in order to keep it aligned with human ends.
It seems that we should want an AI-agent-specific form of currency that can serve as a rewards mechanism--incentivizing (disincentivizing) pro-social/aligned (anti-social/unaligned) behavior.
Programmable digital assets (crypto) likely will be extremely useful for the alignment problem that @yonashav proposes. Programmability can help to keep humans the beneficial owners of the human medium of exchange. For example, a multisig on the wallet controlled by the agent with one key required to be a proof-of-humanity (E.g., Worldcoin) for certain inherently dangerous purchases. In addition, multiple programmable currencies would allow bespoke currencies for AI agents themselves that can function as the AI agent rewards system but not be used as a general payment medium. Here, you do have a convertibility problem. You could imagine sophisticated rogue agents trying to devise schemes to convert their bespoke rewards currency into a human medium of exchange. A transparent public ledger would be very useful here for combatting AI agents engaging in "money laundering." @tylercowen was right that Web3 has become underrated and is waiting on the shelf to be "the legal and institutional framework for AI bots."
Ultimately, alignment of agents to principals historically requires a rewards system, and that system is enforced through agents being beneficial owners of assets of some kind. Diverse, programmable currencies with different properties when wielded by bots vs. humans could be the mechanism to keep humans as society's prime principals.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
