Intelligent Agents and Internet Commerce in Ancient Rome

October 14, 2008

In Ancient Rome, it is no exaggeration to say that the bulk of trade and commerce was carried out by slaves. Educated slaves would be entrusted with many aspects of their masters’ affairs, and the record shows slaves acting as shopkeepers, merchants, bankers and estate managers. However, the status of a slave as a ‘thing’, or res, was never in doubt. Consequently, slaves had no rights (and, as I explain later, no duties), and could not hold property. They could not sue or be sued. The Romans nevertheless created a social and legal structure which allowed slave-run commerce to prosper.


Similarly, intelligent agents are ‘things’, having in themselves no rights or duties, and are incapable of suing or being sued. If Internet commerce based on intelligent agents is to thrive, are there any aspect of the Roman model of slavery which can usefully be used as a model?


 


Intelligent Agents


Internet commerce will increasingly consist of intelligent agents, programmed to buy and sell financial instruments, commodities, goods and even services (for example, mobile phone minutes). The sophistication of these agents will increase: automated share trading systems initially relied on simple price targets and stop-loss thresholds, but now incorporate more complex algorithms. These agents can participate in arbitrage, monitoring data on sale and purchase prices, and executing transactions where they can make a profit. The power of their algorithms will doubtless increase, using artificial intelligence techniques such as heuristic and rule-based approaches.


These agents are able to establish rights and obligations between persons, and are able to cause damage. How should the law deal with the enforcement of these rights and obligations, and who should be liable for any damage caused by the agent?


It is tempting to think that the answer is straightforward: clearly the ‘owner’ of the agent should be liable for all the agent’s actions. However, aside from the fact that determining who the ‘owner’ is may be difficult, I will demonstrate that this answer is not entirely satisfactory.


eBay’s bidding mechanism, for example, can be regarded as an intelligent agent which places bids on the items selected up to a preset limit selected by the bidder. At one remove, bid sniping software and services aim to subvert the eBay technology and the psychology of bidding, by waiting until the very last moment before an auction closes before submitting bids. Thus the sniping service is acting as an intelligent agent which is instructing an intelligent agent. Even at the current state of development of intelligent agent technology, things can get quite complex. And this complexity is bound to increase over time.


Where a single purpose agent is controlled and instigated by one legal entity (the principal) within a set of constraints, the rule that the principal is responsible for all the agent’s actions functions well. However, as the agents get more complex, less constrained, and can begin to collaborate and trade directly with each other, more complex rules will be required: for example, if an agent runs amok and starts vastly overpaying for items because of a programming error, should the item vendors be able to claim the inflated price from the principal under the absolute liability rule, or should the principal be able to evade responsibility, because the vendor should have been aware that the agent was acting outside the scope of its authority (or usual functional parameters)?


If bid-sniping software allows a mechanism to exist whereby two competing users of the software on the same item can agree between them as to who will snipe, does that count as criminal bid rigging? What if the designers of the sniping software implemented a mechanism to do this without the users being aware of it? What if the agents acquired such a strategy independently themselves using AI techniques (this is not far-fetched: AI software has been used to develop strategies for Prisoners’ Dilemma – which is akin to this scenario – for many years).


The question ‘what can or should an intelligent agent be liable for?’ is a fascinating one, and is worthy of further investigation.[1] However, this paper seeks to consider a much narrower point: assuming that commercial liability is established, through contract, for example, who should be liable on the contract, and should that liability be limited?


There is one relatively straightforward way of dealing with this question: where the agent is constrained to act in one context (for example, where it is buying and selling commodities on a particular exchange) then liabilities and limitations can be established by a framework agreement, or set of framework agreements, between the exchange members. Thus whenever the agent transacts, the transaction is between its principal and (depending on the way the framework agreement is constructed) the exchange – with a corresponding transaction between the exchange and the counterparty – or directly with the counterparty. If the transaction goes wrong, the parties look to the framework agreement to assess the appropriate rights and liabilities between the parties.[2]


The problem with this ‘walled garden’ approach is that there is inevitably a barrier to entry to join the club, and by limiting the number of parties who can join, the market is made very inefficient. Network effects are very powerful: the number of potential transactions increases as the square of the number of counterparties, and thus the efficiency of the market increases. By severely restricting the number of players in a market, its efficiency is therefore lowered.


Therefore, if we agree that, from an economic perspective, it is desirable that intelligent agents have greater degrees of freedom, and therefore a greater choice of counterparties with whom to transact, and wider scope to negotiate the terms of the contract (other than price) on which they are prepared to transact, we need a legal framework which allows this to happen.


It is easy to assume that as these issues are triggered by emerging technology, that the laws to deal with them will have to be newly developed also. However, the jurists of Ancient Rome anticipated many of these issues by developing an advanced legal framework to cover the rights and obligations arising from slave ownership.[3]


 


Roman Slavery


Interestingly, the Romans regarded the human state to be one of freedom, under natural law (ius naturale). It was human-made law (ius gentium) which created the concept of slavery (strictly, chattel slavery: the ownership of one person by another). A slave was a thing, a res, albeit a thing with special characteristics. The slave’s master had (at least in the early Empire) theoretical power of life and death over him.[4]


However, slaves, especially in the early Empire, and especially educated urban slaves, were valuable things, and as such only foolish masters would maltreat them.


As things, slaves had no rights,[5] and it was impossible to enforce any form of agreement against them. Clearly, in the absence of any amending rules, this state of affairs would render it very unwise to enter into any form of contract with a slave.


 


Throughout the Roman period, therefore, laws were developed which enabled slaves to transact business on behalf of their masters. A balance was struck between protecting masters from excess liabilities arising from the activities of errant slaves, and giving counterparties comfort that it was safe to transact with slaves, and that the slaves’ obligations under the transaction were enforceable.[6]


Analogously, therefore, we can consider whether any of these mechanisms are applicable to the use of intelligent agents.


 


The Peculium


 


The most important and interesting mechanism is the peculium. Although a slave was incapable of owning property, the master could consent to him holding property for his own use and enjoyment. This is not limited to money: it can include any form of property, including other slaves (which could even have their own peculia).


Where a slave was involved in commerce, his peculium could be considered analogous to working capital. Although technically the peculium belonged to the master, it was regarded for most practical purposes as belonging to the slave, and indeed it became common practice for the slave to be able to buy his own freedom using the funds contained in his peculium.


The existence of the peculium was critical to giving comfort to people transacting with slaves.  The Roman law had a basic form of agency – any contract entered into by a slave would entitle his master to the benefit, but he would not be bound by the burden. If a contract could not be enforced by the slave’s customer, that was clearly a major barrier to commerce. The Romans therefore developed a rule allowing a contracting party to enforce judgments against the peculium. The action, known as the actio de peculio, allowed the master (paterfamilias) to be sued to the extent of the value of the peculium at the time of judgment[7] under contracts or debts entered into by the slave. This applied even where the transaction in question was unauthorised, or  for that matter forbidden, by the master.


This mechanism encouraged masters to use slaves as business managers, because the masters’ liability was limited to the value of the peculium, and it encouraged people to transact with slaves because of the perceived security of the peculium. In many ways, these was the same reasoning which resulted in the invention of the joint stock limited liability company in the 19th century.


As the law developed, the scope of the actio de peculio was extended so that in certain circumstances, the claim could exceed the value of the peculium. For example, where the transaction giving rise to the claim led to a tangible benefit to the master, the size of the claim could be extended to follow the benefit. Interestingly, food and clothing for the slave (if not excessive) would count as a benefit for the master – which of course naturally follows from the underlying fact that the slave is a res.


The parallels with autonomous agents are clear. The analogy (like any other analogy) is not a perfect one, but comparison may be instructive. Like a slave, an autonomous agent has no right or  duties itself. Like a slave, it is capable of making decisions which will affect the rights (and, in later law, the liabilities) of its master. By facilitating commercial transactions, autonomous agents have the ability to increase market efficiency.  Like a slave, an autonomous agent is capable of doing harm.[8]


 


Applicability of the Peculium to Intelligent Agents


 


What elements of Roman legislation could usefully be applied to autonomous agents? It is worth noting that in Roman society, an individual’s status (slave/free, Roman/foreign, family/independent)[9] was a matter of public knowledge. On the internet, no one knows you’re a dog:[10] clearly, not knowing the status of the entity with which a party is contracting is fatal to any scheme which depends on knowledge of that status. So I propose a system in which, by piggybacking on the back of existing trust networks (such as PKI) an intelligent agent can be granted an identity which can be authenticated. It may not be obligatory to sign up to such a scheme, but failure to do so will render the ‘owner’ of the autonomous agent strictly liable for any actions of that agent (or whatever default position exists at law, this being the most likely one). However, to encourage both people to use autonomous agents, and to trust transactions with them, if an agent is registered with the scheme, it is granted the privilege[11] of limited liability, subject to the provision of a ring-fenced peculium.[12]


The trust system would also hold the following information about any agent: (1) details of its owner, (2) details of the size of its peculium, and (3) details of any contingent liabilities. The peculium would be backed by a financial institution, such as an insurance company, the details of which would also held by the trust provider.[13]  In terms of infrastructure, this is similar to the amount of information held by DVLA in relation to each car registration, its associated insurance company and MOT status. The most problematic heading here is that of contingent liability: essentially the database would have to be dynamically updated each time a transaction was completed with an agreed maximum exposure negotiated as part of the transaction. There would have to be some mechanism agreed for striking these exposures from the database: this could include a limitation period for claims (after which the liability would disappear), or a mechanism (like the eBay rating system) for the parties to a transaction indicating their satisfaction with its outcome, and therefore agreeing to wipe off the prospective liability (on both sides, if the transaction is between two registered agents).


It is also worth noting that it is likely that there would need to be exceptions to the privilege of limited liability. The obvious ones would be in the case of fraud, or the supply of dangerous goods which caused death or personal injury.


The system does require an infrastructure, but it is significantly simpler than the ‘walled garden’ approach, and gives any contracting party the comfort that the transaction is secure to the extent of the un-earmarked peculium (and in the case of an agent-to-agent transaction, the peculia on each side).


 


This is the clearly the barest outline of a possible scheme, but I believe that it provides the lightest touch of regulation, allows the market to set prices for risk, and renders that risk transparent and makes mitigation accessible for traders, while at the same time encouraging trade by allowing the traders to shelter behind the privilege of limited liability.


 


Andrew Katz is a partner at Moorcrofts LLP, where he specialises in technology law with a bias towards open source software. In a former life he was a software developer and has released software under the GPL. He can be reached at Andrew.katz@moorcrofts.com.


 






[1]In  Autonomous Robots and the Law’ (2008) Fernando Barrio considers criminal liability for sexual offences related to robots.



[2] This structure is frequently used in financial transactions – for example, where an investment manager is given power to enter into currency transactions on behalf of a pension fund, the pension fund and the relevant bank with whom the currency trades are transacted may enter into a master agreement called an ISDA or IFEMA which delegates to the investment manager the power to enter into transactions on behalf of the pension fund. Actually, that’s the traditional way to describe it, but when acting for the investment manager I have tried to argue (with varying degrees of success) that it is not correct to describe the investment manager as having powers delegated to it. All that is happening is that by issuing various notices, the investment manager is triggering transactions in an already existing contract. This can be important, because if it is truly deemed to be an agent, then the investment manager is automatically invested with various obligations, including fiduciary obligations to the client. As the investment manager’s adviser, I seek to minimize these obligations. Note that there is no reason why the investment manager has to be a party to the contract between the bank and the pension fund: the bank is merely instructed by the pension fund to conduct various trades if it receives notification from the manager (in practice, the manager usually is a party, but that is for peripheral reasons). Imagine that the transactions generated by the manager are produced purely algorithmically (which in practice they frequently are, with no human intervention): there is no reason why the computer could not (with the appropriate provision in the framework agreement) trigger the transactions merely by sending appropriate e-mails to the bank. Now imagine that I have set up a stop-loss trigger on one of the stocks held in my online sharedealing account, to sell if the market price goes below 80p per share. The online brokerage gets its feed, for sake of argument, from the FT. Clearly, it’s absurd to suggest that the FT is acting as an agent by sending a feed to the brokerage saying that the share price of XYZ plc is 79p. Similarly, the manager is not acting as an agent when notifying the bank that a certain trigger point has been reached and that a certain transaction must be transacted.



[3]They also considered the circumstances in which liability should arise, and who should be fixed with that liability, but those questions are outside the scope of this article. It is, however, worth briefly mentioning noxal liability. Under this doctrine, where a slave committed a delict (tort), the master became liable to hand over the slave to the victim. However, this generally led to the master and the victim agreeing a monetary compensation instead, and acted as a form of limited liability. The system was a blunt instrument, to the extent that there was no correlation between the cost of the harm suffered and the value of the slave. However, it is interesting to speculate that if you are harmed by a robot, you might have the right to take ownership of the robot.



[4] I have been asked several times while talking to people about this article what punishments slaves were liable to suffer for various infractions: the simple answer is that there was no procedure or tariff. At least in relation to the early period, the question is as meaningless as asking what punishment a television should suffer for repeatedly showing Noel Edmonds programmes. Both a slave and television are things, and if, as the owner of each, you decide it’s appropriate to apply a bullwhip for any arbitrary reason, then that is entirely your own choice. There was, however, in the later period, the ability for slaves to seek sanctuary from their masters’ cruelty in certain temples. One reference for this — Introduction to Roman Law (Nicholas) — compares this to the  right of slaves to seek sanctuary from their masters by appealing to the British protectors in the Trucial states (presumably Oman). It is chilling to realise that such a right was necessary when the book was published – in  1962.



[5] That is they were unable to seek redress for any wrong in the courts. 



[6]They also provided for an incentive for slaves to perform well, acknowledging that the carrot was a valuable management tool as well as the stick. However, it is difficult to see how this is applicable to intelligent agents in their current state of development.



[7] This does lead to some further questions: since the peculium technically belonged to the master, what prevented the master from reclaiming it between action and judgment (or indeed at any other time)? How did any person contracting with the slave know that the size of the peculium? Or that the peculium was already subject to any number of pending or potential actions? This needs further investigation.



[8]For completeness, it is worth pointing out that (at least not in the foreseeable future), autonomous agents will still not be able to die, be set free, be punished, be incentivised or marry: but as computer programs, they are clearly able to reproduce. They can also commit crimes, although what constitutes a crime in this context, especially in terms of mens rea is a matter of further investigation.  Fernando Barrio ‘Autonomous robots and the Law’ (2008) discusses these issues in a similar context.



[9]These were the three main incidents of status. Most freemen (and, under analogous rules, taking account of marriage,  women) were under the power (patria potestas) of their oldest male ancestor. The status of those in patria potestas was surprisingly similar to slavery: for example, a son under the control of his paterfamilias could be granted a peculium, as he was incapable of holding property himself. In early Rome, the paterfamilias even had the power of life and death over his offspring, although this right was tempered in later years.



[10]Peter Steiner, The New Yorker, June 1993



[11]Much as the grant of the privilege of limited liability for an English limited company, or limited liability partnership is granted subject to payment of incorporation fees, and ongoing disclosure and fee requirements, as well as adherence to a vast body of statutory rules.



[12]What I am proposing would be much simpler than the existing rules for limited liability corporations or LLPs. There would have to be no accounting rules (income and expenditure would be deemed to belong to the owner), no rules on corporate governance, no rules governing relations between shareholders (there would only be one owner, although the owner would be able to sell). A similar mechanism is discussed by Waleed Al-Majid Electronic Agents and Legal Personality – Time to Treat them As Human Beings BILETA Conference 2007. The mechanism proposed in this paper is more flexible, in that it covers both models discussed in that paper, namely the deposit model proposed by Giovanni Sartor and referred to, and the insurance model proposed by Curtis Karnow, and, crucially, provides a mechanism to transition from the expensive requirement of a deposit to cover a bond, to gradually allowing a portion of that deposit to be replaced by an insurance product, as the risk profile associated with that agent becomes clearer.


 



[13]Presumably, the financial institution would initially require a deposit of the full amount of the peculium, but subsequently, would, after assessing the risks involved, would be able to post a portion of the bond at a discount for an annual fee, analogous to credit insurance.