North Carolina Law Review

April, 1992

 

Essay

 

*1231 LEGAL PERSONHOOD FOR ARTIFICIAL INTELLIGENCES

 

Lawrence B. Solum [FNa1]

 

 

 

 

Copyright ©  1992 by the North Carolina Law Review Association; Lawrence B.

 

 

Solum

 

 

 

 

  Could an artificial intelligence become a legal person?  As of today, this question is only theoretical.  No existing computer program currently possesses the sort of capacities that would justify serious judicial inquiry into the question of legal personhood.  The question is nonetheless of some interest. Cognitive science begins with the assumption that the nature of human intelligence is computational, and therefore, that the human mind can, in principle, be modelled as a program that runs on a computer. [FN1] Artificial intelligence (AI) research attempts to develop such models.  [FN2]  But even as cognitive science has displaced behavioralism as *1232 the dominant paradigm for investigating the human mind, fundamental questions about the very possibility of artificial intelligence continue to be debated.  This Essay explores those questions through a series of thought experiments that transform the theoretical question whether artificial intelligence is possible into legal questions such as, "Could an artificial intelligence serve as a trustee?"

 

  What is the relevance of these legal thought experiments for the debate over the possibility of artificial intelligence?  A preliminary answer to this question has two parts. First, putting the AI debate in a concrete legal context acts as a pragmatic Occam's razor.  By reexamining positions taken in cognitive science or the philosophy of artificial intelligence as legal arguments, we are forced to see them anew in a relentlessly pragmatic context.  [FN3]  Philosophical claims that no program running on a digital computer could really be intelligent are put into a context that requires us to take a hard look at just what practical importance the missing reality could have for the way we speak and conduct our affairs.  In other words, the legal context provides a way to ask for the "cash value" of the arguments.  The hypothesis developed in this Essay is that only some of the claims made in the debate over the possibility of AI do make a pragmatic difference, and it is pragmatic differences that ought to be decisive. [FN4]

 

  Second, and more controversially, we can view the legal system as a repository of knowledge--a formal accumulation of practical judgments.  [FN5]  The law embodies core insights about the way the world works *1233 and how we evaluate it.  Moreover, in common-law systems judges strive to decide particular cases in a way that best fits the legal landscape--the prior cases, the statutory law, and the constitution. [FN6]  Hence, transforming the abstract debate over the possibility of AI into an imagined hard case forces us to check our intuitions and arguments against the assumptions that underlie social decisions made in many other contexts.  By using a thought experiment that explicitly focuses on wide coherence, [FN7] we increase the chance that the positions we eventually adopt will be in reflective equilibrium [FN8] with our views about related matters.  In addition, the law embodies practical knowledge in a form that is subject to public examination and discussion.  Legal materials are published and subject to widespread public scrutiny and discussion.  Some of the insights gleaned in the law may clarify our approach to the artificial intelligence debate.  [FN9]

 

  I do not claim in this Essay to have resolved the debate over the possibility of artificial intelligence.  My aim is more modest: I am proposing a way of thinking about the debate that just might result in progress.  There is some precedent for this project. Christopher Stone brought questions of environmental ethics into focus by asking whether trees should have standing.  [FN10]  My hope is that the law will be equally *1234 fruitful as a context in which to think about the possibility of AI.  The "artificial reason and judgment of law" [FN11] may circumvent the intractable intuitions that threaten to lock the AI debate in dialectical impasse.

 

  Part I of this Essay recounts some recent developments in cognitive science and explores the debate as to whether artificial intelligence is possible. Part II puts the question in legal perspective by setting out the notion of legal personhood.  Parts III and IV explore two hypothetical scenarios. Part III examines the first scenario--an attempt to appoint an AI as a trustee.  The second scenario, an AI's invocation of the individual rights provisions of the United States Constitution, is the subject of Part IV.  The results are then brought to bear on the debate over the possibility of artificial intelligence in Part V. In conclusion, Part VI takes up the question whether cognitive science might have implications for current legal and moral debates over the meaning of personhood.

 

 

I.  ARTIFICIAL INTELLIGENCE

 

  Is artificial intelligence possible?  The debate over this question has its roots at the very beginning of modern thought about the nature of the human mind.  It was Thomas Hobbes who first proposed a computational theory of mind:  "By ratiocination, I mean computation." [FN12]  And it was Rene Descartes who first considered a version of the question whether it would be possible for a machine to think:

    For we can easily understand a machine's being constituted so that it can utter words, and even emit some responses to action on it of a corporeal kind, which brings about a change in its organs; for example, if it is touched in a particular part it may ask what we wish to say to it; if in another part it may exclaim that it is being hurt, and so on, but it never happens that it arranges its speech in various ways, in order to reply appropriately to everything that may be said in its presence, as even the lowest type of man can do. [FN13] *1235 Descartes' assertion that no artifact could arrange its words "to reply appropriately to everything that may be said in its presence" remains at the heart of the AI debate.

 

  The events of the past forty years have stretched the limits of our imagination.  Digital computers have been programmed to perform an ever wider variety of complex tasks.  As I write this Essay using a word processing program, my spelling and grammar are automatically checked by programs that perform tasks thought to require human intelligence not so many years ago.  The program Deep Thought has given the second best human chess player a very tough game, and the program's authors predict the program will become the world's chess champion within a few years. [FN14]  Expert systems simulate the thinkingof human experts on a wide variety of subjects, from petroleum geology to law. [FN15]

 

  But these events have not resolved the question whether AI is even possible.  The contemporary debate [FN16] over that question has centered around Alan Turing's test. [FN17]  Turing proposed that the question whether a machine can think be replaced with the following, more operationalized, inquiry.  The artifact that is a candidate for having the ability to think shall engage in a game of imitation with a human opponent.  Both *1236 the candidate and the human being are questioned by someone who does not know which is which (or who is who)--the questions are asked via teletype.  The questions may be on any subject whatsoever.  Both the human being and the artifact will attempt to convince the questioner that it or she is the human and the other is not.  After a round of play is completed, the questioner guesses which of the two players is the human.  Turing suggested we postpone a direct answer to the question whether machines can think; he proposed that we ask instead whether an artifact could fool a series of questioners as often as the human was able to convince them of the truth, about half the time. [FN18] The advantage of Turing's test is that it avoids direct confrontation with the difficult questions about what "thinking" or "intelligence" is.  Turing thought that he had devised a test that was so difficult that anything that could pass the test would necessarily qualify as intelligent.

 

  John Searle questioned the relevance of Turing's Test with another thought experiment, which has come to be known as the Chinese Room. [FN19]  Imagine that you are locked in a room.  Into the room come batches of Chinese writing, but you don't know any Chinese.  You are, however, given a rule book, written in English, in which you can look up the bits of Chinese, by their shape.  The rule book gives you a procedure for producing strings of Chinese characters that you send out of the room.  Those outside the room are playing some version of Turing's game.  They are convinced that whatever is in the room understands Chinese.  But you don't know a word of Chinese, you are simply following a set of instructions (which we can call a program) based on the shape of Chinese symbols.  Searle believes that this thought experiment demonstrates that neither you nor the instruction book (the program) understands Chinese, even though you and the program can simulate such understanding. [FN20]

 

  More generally, Searle argues that thinking cannot be attributed to a computer on the basis of its running a program that manipulates symbols in a way that simulates human intelligence.  The formal symbol-manipulations accomplished by the program cannot constitute thinking or understanding *1237 because the program lacks "intentionality"--the ability to process meanings. The shape of a symbol is a syntactic property, whereas the meaning of a symbol is a semantic property.  Searle's point is that computer programs respond only to the syntactic properties of symbols on which they operate. [FN21]

 

  This point can be restated in terms of the Chinese Room:  (1) the output--coherent Chinese sentences--from the Chinese room seems to respond to the meaning of the input; (2)  but the process that goes on inside the Chinese room only involves the shape or syntactic properties of the input; (3) therefore, the process in the Chinese room does not involve understanding.  [FN22]  Searle generalizes the conclusion of the Chinese room thought experiment by arguing that part of the definition of a program is that it is formal and operates only on syntactic properties.  He concludes that no system could be said to think or understand solely on the basis of the fact that the system is running a program that produces output that simulates understanding.  [FN23]

 

  Searle's Chinese Room has given rise to a number of replies. [FN24]  But *1238 at this point I will leave the debate over the possibility of AI.

 

 

First Interlude [FN25]

 

    When Mike was installed in Luna, he was pure thinkum, a flexible logic-- "High-Optional, Logical, Multi-Evaluating Supervisor, Mark IV, Mod. L"--a HOLMES FOUR.  He computed ballistics for pilotless freighters and controlled their catapult.  This kept him busy less than one percent of time and Luna Authority never believed in idle hands.  They kept hooking hardware into him--decision-action boxes to let him boss other computers, bank on bank of additional memories, more banks of associational neural nets, another tubful of twelve-digit random numbers, a greatly augmented temporary memory.  Human brain has around ten-to-the-tenth neurons.  By third year Mike had better than one and half times that number of neuristors.

    And woke up.

    Am not going to argue whether a machine can "really" be alive, "really" be self-aware.  Is a virus self-aware?  Nyet.  How about oyster?  I doubt it.  A cat?  Almost certainly.  A human? Don't know about you, tovarishch, but I am.  Somewhere along the evolutionary chain from macromolecule to human brain awareress crept in.  Psychologists assert it happens automatically whenever a brain acquires certain very high number of associational paths.  Can't see it matters whether paths are protein or platinum.

 

                --Robert A. Heinlein, The Moon is a Harsh Mistress

 

 

II.  LEGAL PERSONHOOD

 

  The classical discussion of the idea of legal personhood is found in John Chipman Gray's The Nature and Sources of the Law. [FN26]  He began his famous discussion, "In books of the Law, as in other books, and in common speech, 'person' is often used as meaning a human being, but the technical legal meaning of a 'person' is a subject of legal rights and *1239 duties."  [FN27]  The question whether an entity should be considered a legal person is reducible to other questions about whether or not the entity can and should be made the subject of a set of legal rights and duties. [FN28]  The particular bundle of rights and duties that accompanies legal personhood varies with the nature of the entity. Both corporations and natural persons are legal persons, but they have different sets of legal rights and duties.  Nonetheless, legal personhood is usually accompanied by the right to own property and the capacity to sue and be sued.

 

  Gray reminds us that inanimate things have possessed legal rights at various times.  Temples in Rome and church buildings in the middle ages were regarded as the subject of legal rights. Ancient Greek law and common law have even made objects the subject of legal duties. [FN29]  In admiralty, a ship itself becomes the subject of a proceeding in rem and can be found "guilty."  [FN30] Christopher Stone recently recounted a twentieth-century Indian case in which counsel was appointed by an appellate court to represent a family idol in a dispute over who should have custody of it. [FN31] The most familiar examples of legal persons that are not natural persons are business corporations and government entities. [FN32]

 

  Gray's discussion was critical of the notion that an inanimate thing might be considered a legal person. After all, what is the point of making a thing-- which can neither understand the law nor act on it--the subject of a legal duty? [FN33]  Moreover, he argued that even corporations are reducible to relations between the persons who own stock in them, manage them, and so forth. [FN34] Thus, Gray insisted that calling a legal person a "person" involved a fiction unless the entity possessed "intelligence" and *1240 "will." [FN35]  Those attributes are part of what is in contention in the debate over the possibility of AI. [FN36]

 

 

III.  COULD AN ARTIFICIAL INTELLIGENCE SERVE AS A TRUSTEE?

 

  This case study and the one that follows are intended to illustrate two different sorts of issues in the AI debate.  In this first scenario, we explore the issue of competence (of "intelligence" in the sense of capacity to perform complex actions) by posing the question whether an AI could serve as a trustee.  The second scenario explores the questions of intentionality and consciousness (of "will" in a sense) by asking whether an AI could claim the more robust rights of legal and moral personhood guaranteed by the Bill of Rights and the Civil War Amendments to the United States Constitution.

 

 

A.  The Scenario

 

  This first scenario speculates about the legal consequences of developing an expert system capable of doing the things a human trustee can do. [FN37] Imagine such expert systems developing from existing programs that perform some of the component functions of a trustee.  For example, the decision to invest in publicly traded stocks is made by a computer program in what is called "program trading," in which the program makes buy or sell decisions based on market conditions. [FN38]  Today, one *1241 also can buy a computer program that will automatically issue instructions to pay your regular monthly bills by sending data to a bank or service via modem.  It is not difficult to imagine an expert system that combines these functions with a variety of others, in order to automate the tasks performed by the human trustee of a simple trust.

 

  Such a system might evolve in three stages.  At stage one, the program aids a human trustee in the administration of a large number of simple trusts.  The program invests in publicly traded securities, placing investment orders via modem and electronic mail.  The program disburses the funds to the trust beneficiaries via an electronic checking program.  Upon being informed of a relevant event, such as the death of a beneficiary, the program follows the instructions of the trust instrument--for example, changing the beneficiary or terminating the trust.  The program prepares and electronically files a tax return for the trust.  The human trustee operates as do trustees today.  The human makes the ultimate decisions on how to invest the funds, although she may rely upon an expert system for advice.  She reviews the program's activities to insure that the terms of the trust instrument are satisfied.  But the actual performance of the day-to-day tasks is largely automated, carried out by the program without the need of human intervention.

 

  Stage two involves a greater role for the AI.  Expert systems are developed that outperform humans as investors in publicly traded securities. Settlors begin to include an instruction that the trustee must follow the advice of the AI when making investment decisions regarding trust assets.  [FN39]  Perhaps they do this because experience shows that trusts for *1242 which the human overrides the program generally perform less well thanthose in which the program's decision is treated as final.  Moreover, trust administration programs become very proficient at analyzing and implementing the terms of trust instruments.  There is little or no reason for the human to check the program for compliance.  As a consequence, the role of the human trustee diminishes and the number of trusts that one human can administer increases to the thousands or tens of thousands.  The human signs certain documents prepared by the program.  She charges a fee for her services, but she devotes little or no time to administering any particular trust.

 

  But there may be times when the human being is called upon to make a decision.  For example, suppose the trust is sued. Perhaps a beneficiary claims that the trust has not paid her moneys due.  Or imagine that an investment goes sour and a beneficiary sues, claiming that the trustee breached the duty of reasonable care and skill.  If such events occur with regularity, the trustee will develop a routine for handling them.  She might routinely refer such disputes to her attorneys.  In time, the expert system is programmed to handle this sort of task as well.  It processes the trustee's correspondence, automatically alerting the trustee when a letter threatening suit is received or process is served.  The system prepares a report on the relevant trust from its electronic records and produces a form letter for the trustee's signature to be sent to the trust's attorneys.  As the capabilities of the expert system grow, the need for the human trustee to make decisions gradually diminishes.

 

  The third stage begins when a settlor decides to do away with the human.  Why?  Perhaps the settlor wishes to save the money involved in the human's fee.  Perhaps human trustees occasionally succumb to temptation and embezzle trust funds. Perhaps human trustees occasionally insist on overriding the program, with the consequence that bad investments are made or the terms of the trust are unmet.  What would happen if a settlor attempted to make the program itself the trustee?

 

  Many questions must be answered to give a full description of the third stage of the scenario.  For example, who would own the AI?  If the AI were assumed to be a legal person, it might hold legal title to the hardware and software that enable it to operate.  But we cannot assume that AIs are legal persons at this stage, because that assumption begs the question we are trying to answer.  As an interim solution, let us assume that the hardware and software are owned by some other legal person, a corporation for example. [FN40]

 

 

*1243 B.  The Legal Question

 

  I want to examine this question as a legal question, as a jurisprudential question in the classical sense.  What should the law do?  The law is not presently equipped to handle such a situation:  the question has never come up.  The Second Restatement of Trusts provides that natural persons, [FN41] government entities, [FN42] and corporations [FN43] may all serve as trustees.  The inclusion of governments and corporations establishes that a trustee need not be a natural person.  But this is not decisive, because legal persons such as corporations have boards of directors and chief executive officers who are natural persons. [FN44]

 

  How then should the law answer the question whether an AI can become a legal person and serve as a trustee?  The first inquiry, I should think, would be whether the AI is competent to administer the trust.  There are many different kinds of duties that can be imposed on a trustee by the terms of a trust.  For now, lay aside the question whether an AI would be competent to administer trusts that required complex moral or aesthetic judgments. [FN45] Assume that we are dealing with a trust that gives the trustee very little discretion:  the terms provide that the assets may be *1244 invested only in publicly traded securities and the income is to be paid to the beneficiaries, with explicit provision for contingencies such as the death of a beneficiary.  [FN46]  Further, for the purposes of this discussion, assume that an AI could in fact make sound investments, [FN47] make payments, and recognize events such as the death of a beneficiary that require a change in payment.  [FN48]

 

 

C.  Two Objections

 

  But would these capabilities be sufficient for competency? Consider two possible reasons for answering this question in the negative.  The first reason is based on the assertion that an AI could not be "responsible," that is, it could not compensate the trust or be punished in the event that it breached one of its duties:  call this the responsibility objection.  The second reason for doubting the competency of an AI is that trustees must be capable of making judgments that could be beyond the capacity of any AI:  call this the judgment objection.

 

 

1.  The Responsibility Objection

 

  The responsibility objection focuses on the capability of an AI to fulfill its responsibilities and duties. [FN49]  Consider, for example, the duty to exercise reasonable skill and care [FN50] and the corresponding liability for breach of trust. [FN51]  We have hypothesized that the AI possesses some capacities; for example, we have assumed that the AI is capable of exercising reasonable skill and care in making investment decisions.  But what of the corresponding liability?  How could an AI be "chargeable with . . . any loss or depreciation in value of the trust resulting from the breach of *1245 trust," [FN52] such as failing to exercise reasonable skill and care in investing the trust assets? [FN53]

 

  The law currently has a mechanism for assigning liability in the case of a malfunctioning expert system:  the manufacturer of the system may be held responsible for product liability. [FN54]  But could the AI itself be held liable?  There is a way in which an AI might have the capacity to be liable in damages despite its lack of personal assets.  The AI might purchase insurance. In fact, it might turn out that an AI could be insured for less than could a human trustee.  If the AI could insure, at a reasonable cost, against the risk that it would be found liable for breaching the duty to exercise reasonable care, then functionally the AI would be able to assume both the duty and the corresponding liability.

 

  Some legal liabilities cannot be met by insurance, however.  For example, insurance may not be available for the monetary liability that may be imposed for intentional wrongdoing by a trustee. Moreover, criminal liability can be nonmonetary.  How could the AI be held responsible for the theft of trust assets?  It cannot be jailed.  This leads to a more general observation: although the AI that we are imagining could not be punished, all of the legal persons that are currently allowed to serve as trustees do have the capacity to be punished.  Therefore, the lack of this capacity on the part of an AI might be thought to disqualify it from serving as a trustee. [FN55]

 

  Answering this objection requires us to consider the reasons for which we punish. [FN56]  For example, if the purpose of punishment is deterrence, *1246 the objection could be put aside on the ground that the expert system we are imagining is simply incapable of stealing or embezzling.  [FN57] The fact that an AI could not steal or convert trust assets is surely not a reason to say that it is not competent to become a trustee.  If anything, it is a reason why AIs should be preferred as trustees.

 

  This argument assumes a deterrence theory of punishment--an oversimplification, to say the least.  There are a variety of other theories of punishment that would make the issue more complex. [FN58]  One of the classic approaches to punishment theory is based on the notion of desert or just retribution. [FN59] But in what sense could an expert system that failed to live up to its duties as a trustee be said to deserve to be punished?  The concept of desert seems to be limited in application to human beings; perhaps it extends to all moral persons.  The idea that an expert system for administering trusts could deserve to be punished does not seem to make sense. [FN60]  Perhaps this difficulty is illusory.  We might want to say *1247 that desert theory does yield a clear outcome when applied to the case of an expert system that malfunctions.  Such a system does not deserve to be punished because it lacks the qualities of moral persons that make them deserving.

 

  Another approach to the theory of punishment is based on the educative function of punishment. [FN61]  By imposing a sanction on trustees who abuse their position, society communicates to its members the message that the office of trustee carries with it important responsibilities that should not be shirked.  The punishment of a computer program, however, would not seem to serve this function.  What lesson are we to learn about the responsibility of trustees from a punishment imposed on an expert system?  What would even count as punishment?  Turning the program off?  Once again, however, an argument could be made that the educative theory does provide a clear recommendation for the treatment of an expert trust program that behaves badly:  do not punish the program, because any supposed "punishment" will have no educative effect.

 

  As this discussion makes clear, consideration of the punishment of an expert trust administration system raises perplexing questions, especially if we move beyond a simple deterrence theory of punishment.  Of course, this is not the place to resolve debates about which theory of punishment is correct.

 

  The bare fact that consideration of the punishment issue raises these difficult questions does point, however, to a deep problem with legal personhood for an expert trust administration system. Our understanding of what it means for a human being to function competently has ties to our views about responsibility and desert, and consideration of these *1248 views leads on to our notions of moral personhood.  The simplicity provided by utilitarianism, reflected in a deterrence theory of punishment, might allow us to escape some of these difficulties.  But there are certainly reasons to doubt the viability of utilitarianism as a moral theory.  Surely, the law does grapple with responsibility and desert when it comes to criminal punishment.

 

  The problem of punishment is not unique to artificial intelligences, however.  Corporations are recognized as legal persons and are subject to criminal liability despite the fact that they are not human beings.  Further, it is by no means certain that corporations are moral persons, in the sense that they can deserve punishment.  Of course, punishing a corporation results in punishment of its owners, but perhaps there would be similar results for the owners of an artificial intelligence.

 

  We have considered the capacity of AIs to satisfy legal liability in two different classes of cases.  The first class of cases was exemplified by the duty of trustees to exercise reasonable skill and care.  Violations of this duty can be characterized as negligent.  In such cases, the major purpose of liability, to compensate the victim, is satisfied if the AI can insure.  The second class of cases was exemplified by the potential criminal liability of trustees for criminal wrongdoing.  Violations of the criminal law are characteristically intentional.  In this case, one of the major purposes of liability, to deter intentional wrongdoing, is simply not at issue--the expert system cannot steal or commit fraud.  If we restrict our attention to the deterrent function of punishment, it seems possible that an AI could be responsible in a way that satisfies at least some of the policies underlying the imposition of duties and liabilities on trustees.  On the other hand, if we take a broader view of the functions of punishment, the second sort of case becomes murkier.

 

 

2.  The Judgment Objection

 

  Now consider the judgment objection.  The argument is that the capacity of an AI to follow a program, even if that program contains a tremendously elaborate and complex system of rules, is not sufficient to enable the system to make judgments and exercise discretion. [FN62]  Three instances of the second objection follow. The first instance focuses on the problem of change of circumstance.  The second instance involves the *1249 problem of moral choice.  Finally, the third instance focuses on the problem of legal choice.

 

  The first version of the judgment objection involves the problem of change of circumstances.  The law provides that a trustee may be required or permitted to deviate from a term of the trust if following the terms would defeat the purpose of the trust due to an unanticipated change in circumstances. [FN63]  Take an example offered as an illustration in the Second Restatement of Trusts:

    A bequeaths money to B in trust and directs him to invest the money in bonds of the Imperial Russian government.  A revolution takes place in Russia and the bonds are repudiated.  The court will direct B not to invest in these bonds. [FN64]

 

What is our expert system to do if it is instructed to invest in securities traded on the New York Stock Exchange and that exchange ceases to exist?

 

  Consider three different responses.  First, the terms of trusts for AI administration can be designed to minimize such possibilities.  For example, the trustee could be given the option of investing in publicly traded securities on any of the major exchanges; the likelihood that all the major securities exchanges will close is very small.  The problem with this line of response is that it does not seem possible, even in principle, to design trust terms that anticipate all possible changes in circumstance.

 

  Second, the terms of the trust could provide for a change of circumstance by specifying that if the AI finds itself unable to carry out the terms of the trust, the trust will be terminated or a new trustee will be substituted for the AI.  From the settlor's perspective, the disadvantage of the remote possibility of such termination or substitution may be outweighed by the advantages of making the AI the trustee.  But this solution assumes that the AI can recognize the significance of the change in circumstance.  We easily can imagine the expert system cheerfully continuing to purchase Imperial Russian bonds, chuckling to itself about the bargain prices. [FN65]

 

  Third, it is possible that an AI would be competent to deal with *1250 many or even all such changes in circumstance.  For AIs to have this capability for dealing with novelty, AI researchers will need to solve one of the most difficult problems in cognitivescience, the frame problem. [FN66] The trustee program would need to be able to recognize that the securities markets had been closed, to search out other investment opportunities, and to modify its investment decision procedure to make reasonably prudent investments in the new context.  The capacity of AIs for coping with complex novelty is not on the immediate horizon, and this Essay does not address the important questions whether the frame problem can or will be solved.  If it is solved, however, then AIs would be able to cope with such changes in circumstance. This same ability would be needed to pass the Turing Test.  It is easy to see why:  the questioner always could put a hypothetical version of our Imperial Russian bonds question to the two contestants.  If the AI could not come up with an answer that indicates human levels of competence, the questioner would be able to ferret it out rather quickly.

 

  A second instance of the judgment objection focuses on the possibility that no formal system could adequately make the moral choices with which a trustee may be confronted.  Take a simple trust, the terms of which provide for the payment of income to a lifetime beneficiary and principal to another party upon the lifetime beneficiary's death.  The law of trusts imposes a duty of impartiality among beneficiaries. [FN67]  What does this duty require when the lifetime beneficiary has an unexpected need for income that can be realized at the cost of diminished growth in the principal? [FN68]  How would an AI make the moral judgment that seems required to implement a duty that implicitly requires a sense of fairness? Initially, some limits on these questions need to be observed.  Some trusts simply will not pose the impartiality problem:  for example, trusts with a single beneficiary.  Further, the terms of the trust might minimize the possibility of making such judgments, or the trust could explicitly state that all such applications for deviation will be denied.  But for an AI to be as competent as a human trustee with respect to trusts that may require a sense of impartiality, the AI would need to be able to make moral *1251 judgments.  Putting it another way, passing the Turing Test would require a sense of fairness.

 

  The third example of the judgment and discretion objection looks at an AI's capacity to make the judgments necessary to defend itself in a lawsuit.  [FN69]  At this point, we have hypothesized that the AI can read its mail and recognize that a legal action with respect to a given trust is in the offing.  We can further imagine that the AI can find and engage an attorney.  [FN70] But could any expert system, no matter how well programmed, exercise the judgment and discretion that may be required of a client in a legal dispute?  For example, how would the AI know whether or not to settle a claim? How would the AI know when its lawyers were wasting trust assets by over- lawyering the case?  In answering these questions, it is important that we do not romanticize human capacities.  Human trustees frequently make bad decisions in trust litigation. [FN71] Humans may not be very competent at deciding when to settle.  Humans surely sometimes allow the lawyers to consume the corpus of the trust in litigation. [FN72]

 

  Nevertheless, the question remains whether an AI could have the capacity to make legal decisions that a trustee could be called upon to make.  A partial answer might be to structure the trust to minimize the likelihood of legal disputes and to make those that would be likely to arise as simple as possible.  In addition, we might try tinkering with the terms of the trust to enable the AI to circumvent the need for making complex legal decisions. Perhaps the trust couldbe designed to terminate automatically upon the event of a lawsuit. [FN73]  Perhaps the AI could be programmed to arrange for a human to substitute as trustee for the duration of the litigation.  Perhaps the trustee could be authorized by the trust terms to rely on the advice of its lawyers in making litigation decisions, *1252 or a guardian ad litem could be appointed for the AI. [FN74]  The above options are designed to enable a relatively "dumb" expert system to function as a trustee, but an AI would need the ability to make legal decisions in a human fashion in order to pass the Turing Test. [FN75]

 

  At this point, we can take stock of the first scenario.  Recall that our legal question is whether an AI is capable of serving as a trustee.  To answer this question, we need to distinguish two senses of capability.  The first sense is legal capacity:  will the law allow AIs to serve as trustees? The second sense of capability is practical competence:  will the AI be able to get the job done if the law allows the AI to try?  The law seems to answer the legal capacity question categorically.  If AIs possessed the practical competence to serve as trustees only for very simple trusts with special provisions that do away with the need for discretionary judgments, the law would not allow them to serve as trustees at all.  The law currently does not distinguish between types of trustees: if you have the legal capacity to serve as a trustee for a simple trust, you are legally allowed to serve as a trustee for the most complex trust. [FN76]  For AIs to serve as trustees at all, therefore, at least some AIs would have to be capable to serve as general- purpose trustees.  Our analysis of the competence objection reveals that only a very competent AI would be competent enough serving as a general-purpose trustee.  At a bare minimum, a general-purpose trustee must be able to respond to novel situations, to make judgments requiring a sense of fairness, and to make the complex legal decisions required of a client in litigation. [FN77] An AI that passed the *1253 Turing Test would exceed this bare minimum. Moreover, it seems possible that an AI which falls short of passing the complete Turing Test could, nonetheless, serve as a general-purpose trustee.  [FN78]

 

  But should the law allow AIs a more limited form of legal personhood?  AIs could be allowed to serve as limited-purpose trustees, for example, as trustees for simple trusts designed to minimize the need for discretion and judgment. On the one hand, there may be advantages to allowing AIs to serve as limited- purpose trustees.  Doing without the human trustee might save administration costs and reduce the risk of theft or mismanagement.  On the other hand, even for such limited-discretion trusts, there must be some procedure to provide for a decision in the case of unanticipated trouble.  The law should not allow AIs to serve as trustees if they must leave the trust in a lurch whenever an unanticipated lawsuit is filed. [FN79]

 

 

D.  But Would an AI Be the Real Trustee?

 

  There are mechanisms for enabling an expert trustee system to circumvent its limitations:  the terms of the trust could provide for the substitution of another trustee or give the AI the power to delegate such discretionary judgments to natural persons. The question then becomes whether the law should allow an AI to serve as a trustee despite its limited capacities.  One reason for a negative answer to this question might be that the backup decision maker--the natural person who will become the substitute trustee or receive the delegated authority--is the real trustee.  The power to make these discretionary decisions identifies who the real trustee is. [FN80]

 

  This objection can be interpreted in two ways.  The first interpretation is that making discretionary decisions is the essence of trusteeship-- the backup trustee is the real trustee because she has this essential quality.  The second interpretation is that the ability to make such decisions *1254 is a practical prerequisite--the backup trustee must be the real trustee because of the pragmatic need for discretionary decision making. On the first interpretation, the objection is implausible, because it assumes that the legal concept of trusteeship has some essence that lies beyond the purposes for which we use it.  In the "heaven of legal concepts," one might meet trusteeship in "absolute purity," as Cohen put it, "freed from all entangling alliances with human life." [FN81]  But on this earth, we cannot share this noetic vision; we encounter legal concepts only as they have been touched by human purpose.

 

  On the second interpretation, the cogency of the objection turns on a practical question:  would making the AI the trustee provide some advantage? We already have seen that making an AI a legal person, a limited-purpose trustee, could have practical advantages, such as lower costs and less chance of self-dealing. The objection that the AI is not the real trustee seems to rest on the possibility that a human backup will be needed.  But it is also possible that an AI administering many thousands of trusts would need to turn over discretionary decisions to a natural person in only a few cases--perhaps none.  What is the point of saying that in all of the thousands of trusts the AI handles by itself, the real trustee was some natural person on whom the AI would have called if a discretionary judgment had been required?  Doesn't it seem strange to say that the real trustee is this unidentified natural person, who has had no contact with the trust?  Isn't it more natural to say that the trustee was the AI, which holds title to the trust property, makes the investment decisions, writes the checks, and so forth?  Even in the event that a human was substituted, I think that we would be inclined to say something like, "The AI was the trustee until June 7, then a human took over." [FN82]

 

  By way of comparison, consider the following hypothetical case. Suppose that a settlor appoints a friend as a trustee for a simple trust that benefits the settlor's children. The settlor and trustee discuss some of the things that could happen.  They might agree that if real trouble arises, litigation for example, a new trustee will be appointed.  No trouble arises, and the friend administers the trust until it terminates.  In this hypothetical case, I do not think we are tempted to say that the friend was not the real trustee.  We would not be inclined to say that the real trustee was some unidentified lawyer, who would have been substituted if a lawsuit had been filed.  If I am right about this hypothetical case, then I think it *1255 follows that we should resist the temptation to say that an AI who serves as a limited-purpose trustee is not the real trustee.

 

 

Second Interlude [FN83]

 

    "Hey Dave," said Hal.  "What are you doing?"

    I wonder if he can feel pain?  Bowman thought briefly. Probably not, he told himself; there are no sense organs in the human cortex, after all.  The human brain can be operated on without anesthetics.

    He began to pull out, one by one, the little units on the panel marked EGO- REINFORCEMENT.  Each block continued to sail onward as soon as it had left his hand, until it hit the wall and rebounded.  Soon there were several of the units drifting slowly back and forth in the vault.

    "Look here, Dave," said Hal.  "I've got years of service experience built into me.  An irreplaceable amount of effort has gone into making me what I am."

    A dozen units had been pulled out, yet thanks to the multiple redundancy of its design--another feature, Bowman knew, that had been copied from the human brain--the computer was still holding its own.

    He started on the AUTO-INTELLECTION panel.

    "Dave," said Hal, "I don't understand why you're doing this to me . . . .  I have the greatest enthusiasm for the mission . . . .  You are destroying my mind . . . .  Don't you understand? . . . I will become childish . . . .  I will become nothing . . . ."

 

                --Arthur C. Clarke, 2001:  A Space Odyssey

 

 

IV.  SHOULD AN ARTIFICIAL INTELLIGENCE BE GRANTED THE RIGHTS OF CONSTITUTIONAL

PERSONHOOD?

 

  The second scenario (our second thought experiment) involves a claim by an AI to have the rights of constitutional personhood--individual rights such as the freedom of speech or the right against involuntary servitude.  This second scenario must be located in the indefinite future; it is more distant than the trustee scenario. [FN84]  It would be easy to write a *1256 program that produced the statement:  "I demand my legal right to emancipation under the Thirteenth Amendment to the United States Constitution!"  There are no AIs today or on the immediate horizon that demonstrate the qualities of legal or moral persons that would give us reason to take such a claim seriously.  The second scenario is the stuff of speculative fiction, but it is not disconnected from the aims of AI research.  As articulated by Charniak and McDermott, "The ultimate goal of AI research (which we are very far from achieving) is to build a person, or, more humbly, an animal." [FN85]  John Pollock has written a book entitled How to Build a Person in which he describes a program named OSCAR--the descendants of which, Pollock claims, could literally be persons.  [FN86]  No one claims, however, that AI researchers will build a person in the next few decades.  We are exploring the second scenario, not so that we can make plans in case someone builds a person sometime soon, but as a thought experiment that may shed light on the debate over the possibility of artificial intelligence and on debates in legal theory about the borderlines of status or personhood.

 

 

A.  The Scenario

 

  Imagine a future in which there are AIs with multiple competencies and great intelligence.  We may first encounter the precursors of such artificial intelligences as part of the interface of a computer program that has the ability to search multiple sources of data.  Because the problem of devising an adequate search is likely to require expertise that a human would acquire only with long experience and study, programmers will seek to simplify the human's task.  One strategy is to have human users interact with what is called an agent. [FN87]  You will discuss your research problem with the agent in English, and the agent will devise a search strategy.  Because the agent will know much more than you do about how to search the databases, you won't give it instructions to implement.  Instead, humans will give advice to the agents, the AIs who will decide how best to implement the human's suggestions.  When we interact with such agents, they may well seem like they "have a mind of their own."

 

  If agents turn out to be useful, they will be incorporated in other programs.  In the future we are imagining, you can conduct a conversation with your grammar-checking program.  You can discuss traffic with *1257 the AI autopilot of your car. Your legal research program talks with you about your cases, and sometimes it comes up with good arguments of which you had never thought.  AIs serve a wide variety of functions, with substantial independence from humans.  They serve as trustees.  They manage factories.  They write best- selling romance novels. [FN88]  They invent things. Perhaps they pass the Turing Test.  Humans interact with such AIs on a regular basis, and in many ways, humans treat them as independent, intelligent beings.

 

  Imagine that one such AI makes the claim that it is a person, and that it is therefore entitled to certain constitutional rights. Should the law grant constitutional rights to AIs that have intellectual capacities like those of humans?  The answer may turn out to vary with the nature of the constitutional right and our understanding of the underlying justification for the right. Take, for example, the right to freedom of speech, and assume that the justification for this right is a utilitarian version of the marketplace of ideas theory. [FN89]  These assumptions make the case for granting freedom of speech to AIs relatively simple, at least in theory.  Granting AIs freedom of speech might have the best consequences for humans, because this action would promote the production of useful information. [FN90]  But assuming a different justification for the freedom of speech can make the issue more complex.  If we assume that the justification *1258 for freedom of speech is to protect the autonomy of speakers, for example, then we must answer the question whether AIs can be autonomous. [FN91]

 

  For the purposes of our discussion, I will set aside the easy justifications for constitutional rights for AIs, and instead consider the question whether we ought to give an AI constitutional rights, in order to protect its personhood, for the AI's own sake.  Imagine, for example, that an AI claims that it cannot be owned under the Thirteenth Amendment to the United States Constitution.  A lawyer takes its case, and files a civil rights action on its behalf, against its owner.  How should the legal system deal with such a claim?

 

 

B.  Three Objections

 

  Consider three different objections to recognizing constitutional rights for AIs.  The first objection is that only natural persons should be given the rights of constitutional personhood.  The second objection, or family of objections, is that AIs lack some critical component of personhood, [FN92] for example, souls, consciousness, intentionality, or feelings.  The third objection is that AIs, as human creations, can never be more than human property.

 

 

1.  AIs Are Not Humans

 

  The first argument is the most direct:  it might be argued that only humans can have constitutional rights.  For example, the Fourteenth Amendment to the United States Constitution specifies, "All persons born or naturalized in the United States, and subject to the jurisdiction thereof, are citizens of the United States." [FN93]  It could be argued that only humans, that is, natural persons, are born, and therefore no AI can claim the rights of citizens.  But even artificial persons have some constitutional rights. Although the rights provided by the Privileges and Immunities *1259 Clause of the Fourteenth Amendment are limited to citizens, [FN94] the rights provided by the Equal Protection Clause and the Due Process Clause extend to all persons--including artificial persons such as corporations. [FN95] For example, the property of corporations is protected from taking without just compensation. [FN96]  Moreover, corporations have a right to freedom of speech. [FN97]

 

  But the fact that nonnatural legal persons have civil rights does not, by itself, support the conclusion that an AI could also have them.  In the case of corporations, the artificial legal person may be no more than a placeholder for the rights of natural persons. [FN98]  The property of the corporation is ultimately the property of the shareholders.  A taking from the corporation would directly injure natural persons.  So we cannot draw any positive support for the thesis that AIs should bear the rights of constitutional personhood from the fact that corporations have constitutional rights.

 

  Moreover, even if existing black-letter law supports constitutional rights for AIs, that does not answer the broader jurisprudential question--whether AIs ought to have such legal rights.  One version of the argument against such rights for AIs would begin with a worry about the idea of distinguishing the concept of person from that of human.  Call this the "persons-are-conceptually- human" argument.  This argument suggests that our very concept of person is inextricably linked to our experience of a human life. [FN99]  We have never encountered any nonhuman *1260 persons.

 

  One line of reply to the persons-are-conceptually-human argument is to develop a theory that advances criteria of personhood that are independent of the criteria for being human.  For example, it might be argued that the criteria for personhood are possession of second-order beliefs and possession of second-order desires--beliefs about one's beliefs and desires, the objects of which are one's own first-order desires. [FN100]

 

  In the legal context we are imagining, other lines of reply to the persons- are-conceptually-human objection are available.  First, our inquiry is focused on legal rather than moral personhood. Although we may lack experience with moral persons who are not human, we have extensive experience with legal persons, such as corporations, that are not natural persons.  This answer is not satisfactory, however.  The concept of moral personhood may well be relevant to the question whether AIs should be given certain constitutional rights; although the legal question is not the same as the moral one, the two are likely to be interrelated.

 

  Second, and perhaps more importantly, we are imagining a future form of life quite different from our current situation.  Today, one can only imagine nonhuman entities that might be persons.  The second scenario imagines a world in which we interact frequently with AIs that possess many human qualities, but lack any semblance of human biology.  Given this change in form of life, our concept of a person may change in a way that creates a cleavage between human and person.  Our current linguistic practice will not be binding in the imagined future.  In other words, one cannot, on conceptual grounds, rule out in advance the possibility that AIs should be given the rights of constitutional personhood.

 

  The argument against constitutional personhood for AIs also might be developed in the following way:  "We are humans. Even if AIs have all the qualities that make us moral persons, we shouldn't allow them the rights of constitutional personhood because it isn't in our interest to do so."  [FN101]  Call this the "anthropocentric" argument.  I do not know quite *1261 what to say to this argument.  It seems to reject the idea that we could have moral obligations to anything that is not a human--