Artificial Intelligence Legal Research
A critical question concerning Artificial Intelligence is whether a legal system should view AI as a legal person. Previous year students identified that the Supreme Court in a series of abortion cases determined that a fetus is not a person with rights under the constitution until it reaches the stage of viability, that is, the stage where the fetus may survive outside the womb. http://www.daviddfriedman.com/Academic/Course_Pages/21st_century_issues/21st_century_law/ai_law_alive_04.htm; http://en.wikipedia.org/wiki/Fetus#Viability
This standard, however, presumes that a person with protectable rights is a natural person. Issues arise where AI develops to the point where it may be considered conscious. Constitutional jurisprudence evolves with societal norms and standards. For example, there was a point in our country’s history where statutes banning interracial marriage may have been acceptable. After Love v. Virginia, however, that is no longer the case. Consider also, the abolition of “separate but equal” in Brown v. Board of Education. Prior to that case, the
Koops confronted the issue of whether computer agents would qualify for a claim to what he deemed posthuman rights and liberties. That is, human rights like privacy, due process, and bodily integrity claimed by and attributed to non-human agents such as non-biological machines, cyborgs, or synthetic biological entities. Koops perceives three objections. First, only natural persons should qualify for constitutional rights of personhood. Second, AIs lack some critical aspect of personhood. Third, since AI is human creation, it can never be more than human property. Bert-Jaap Koops , BRIDGING THE ACCOUNTABILITY GAP: RIGHTS FOR NEW ENTITIES IN THE INFORMATION SOCIETY? 11 Minn. J.L. Sci. & Tech. 497 (2010)
Under the natural persons objection, only natural persons, and thus, not AI, may qualify for constitutional rights. Importantly, specific constitutional rights already apply to non-human legal persons such as corporations. Indeed, corporations also have a right to freedom of expression. The objection, however, maintains that, in the case of corporations, an AI simply acts as a place-holder for the rights of natural persons, specifically, the executives of the corporation, the shareholders, and board of directors. This objection weakens where AI is perceived as achieving consciousness, and therefore capable of holding moral viewpoints and basing its decisions on such. Id.
Under the “missing something” objection, AI may not qualify for constitutional rights natural persons enjoy because it is missing some critical characteristic natural persons possess. This argument is based on an intuitive sense that AI is missing a “soul” or “feelings.” Id. While this objection seems vague and based more on fear of the unknown, in Future Imperfect, Professor Friedman argues that one of the critical limitations of AI is that it likely not capable of initiative. That is, AI is merely reacting to its environment, based on a set of rules, albeit a complex and evolved set of rules, just as Searle’s Chinese box (see below) demonstrates. If this is true, that “missing something” is initiative, and though an AI may appear conscious and thus, worthy of constitutional protection, in reality it is merely a complex program reacting to its environment. is likely incapable of formulating intent and thus, cannot be held criminally liable for its actions.
II. Criminal Law
Applicability of criminal laws to artificial intelligence creates an interesting problem whose answer likely depends on a given society’s perception of artificial intelligence. In the case where the criminal law perceives an AI as only a computer program – a product – the criminal law would likely view AI as a tool used to perpetrate a crime, and in the courtroom, evidence. Typically, when a criminal act is committed by an innocent agent, like when a person causes a child, a person who is mentally incompetent or otherwise incapable of formulating the requisite intent, the person who caused the agent to commit the act is criminally liable as a perpetrator-by another. The law perceives the intermediary, in this case the AI, as an instrument and the party orchestrating the offense as the real perpetrator as a principal in the first degree. Where an AI is not perceived as capable of formulating intent, whoever instructs the program to perpetrate an unlawful deed, would be held liable. Gabriel Hallevy, I, ROBOT - I, CRIMINAL--WHEN SCIENCE FICTION BECOMES REALITY: LEGAL LIABILITY OF AI ROBOTS COMMITTING CRIMINAL OFFENSES, 22 Syracuse Sci. & Tech. L. Rep. 1
This set of rules, however, likely does not apply where the AI that perpetrated an unlawful act may be conscious. A person cannot be tried or sentenced while mentally incompetent. California Penal Code §1367(a); Godinez v Moran (1993) 509 US 389. Thus, a threshold question arises – is this AI is conscious and thus competent to stand trial?
Generally, to prove an alleged perpetrator committed a criminal act, the State must prove both the actus reus and the mens rea. The actus is the factual element, i.e., criminal conduct, the mens rea is the mental element, i.e., knowledge or general intent in relation to the conduct element. The actus reus requirement is expressed mainly by acts or omissions. Sometimes, other factual elements are required in addition to conduct, such as the specific results of that conduct and the specific circumstances underlying the conduct. The mens rea requirement has various levels of mental elements. The highest level is expressed by knowledge, while sometimes it is accompanied by a requirement of intent or specific intention. Lower levels are expressed by negligence (a reasonable person should have known), or by strict liability offenses. Hallevy, supra, note 1
i. Actus Reus
In the case of a natural person allegedly committing a crime, the actus reus is relatively simple to prove - show that the criminal act occurred, show the person committed that act. There exist multitude means to prove the act occurred, and this should not be any different when applied to AI. Indeed, the American legal system permits entities that are not natural persons – corporations – to be held criminally liable for its acts. Thus, just as the State can show that a corporation committed an act, so too can it prove an AI did. Id.
ii. Mens Rea
Proving the mental element of a specific, or even general intent crime, may prove considerably more difficult. Indeed, even where technology exists to create AI at or above human capacity for intelligence, does not mean that all or any AI possess the capacity to formulate intent. Thus, should an AI allegedly commit a criminal act, courts will likely struggle with proof of intent. Before an AI may be held criminally liable, courts will likely require proof the AI possesses sufficient consciousness to formulate the requisite intent for any crime. Once the State proves the ability to formulate intent, it must prove intent to commit the crime.
1. Turing Test
Courts may apply the Turing test to determine whether an AI possesses consciousness sufficient to formulate intent in its actions. The Turing test is relatively simple – a game of imitation with a human opponent. A person unaware of which competitor is which, questions the human and the potentially conscious AI via type, on any subject whatsoever. After a series of questions, the questioner guesses which is human. Essentially, Turing proposes an indirect means of ascertaining whether an AI is conscious by determining whether it can fool a series of questioners. If the AI can convince the questioner it is the human, at least as often as the questioner believes the human, it passes Turing’s test. Solum LEGAL PERSONHOOD FOR ARTIFICIAL INTELLIGENCES, 70 N.C. L. Rev. 1231 (1992)
Courts applying the Turing test may find it cumbersome, especially in a world in which a large AI population exists and a court is regularly confronted with the issue of whether the AI is, indeed conscious and thus able to formulate criminal intent. Thus, efficiency concerns may prompt courts to consider other tests. For example, the agencies regularly have lists of conforming products, which the agency has tested and deemed compliant with its particular standards. In a world in which AI is common, a government could require that any company releasing AI to the public must first subject the AI to a number of tests, one of which could ascertain whether that particular AI program is capable of formulating intent. Thus, where a State seeks to prosecute a criminal act perpetrated by an AI, it need only look to the agency list as to whether the particular AI qualifies as capable of formulating criminal intent. If so, the AI itself is charged and prosecuted. If not, the person who instructed the AI to commit the act is.
a. Searle’s Chinese Box Thought Experiment
John Searle offers a clever criticism of the Turing test in the form of “the Chinese box” though experiment. Imagine that you are locked in a room that periodically receives batches of Chinese writing you must decipher, but you don’t know Chinese. Persons outside the room are playing Turing’s game. You are given a rule book, in which you can look up the Chinese symbols by their shape. Outside the room, the people are convinced that whatever is in the room understands Chinese. But you don’t, you are following a set of instructions (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. Thus, 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.
A clear shortcoming of the Chinese box hypothetical, however, is that the “program” assisting the person decipher Chinese serves only that purpose. That is, the Chinese manual does not help the person, or even a potentially conscious AI, formulate its response to the question. It only serves to translate. Thus, a human or potentially conscious AI must still formulate its response to questions to convince the questioner of its consciousness regardless of the language in which the question is posed. Considering the breadth of the questions to be posed – any subject whatsoever – the Turing test seems like a legitimate test, at least regarding intelligence.
As discussed above, Professor Friedman argues that one of the critical limitations of AI is that it likely not capable of initiative. Searle’s Chinese box demonstrates this notion. If this is true, an AI is likely incapable of formulating intent and thus, cannot be held criminally liable for its actions.
How AI may be punished is a critical to a discussion on criminal liability with AI. Can an AI be punished in the same manner as its natural person counterparts? There are certainly issues where corporations cannot be held liable in the same manner as natural counterparts. The death penalty is legal for humans, only for the most heinous of crimes. Can an AI be put down for a misdemeanor? Courts are permitted to require, as a condition of parole or probation that people take prescribed medication, does the same rule apply to fix “bugs” in an AI’s program? Can a court order this type of fixing?
i. Punishment for Criminally Liable AI
argues that most common punishments
are applicable to AI robots. The imposition of specific penalties on AI robots
does not negate the nature of these penalties in comparison with their
imposition on humans. Of course, some general punishment adjustment
considerations are necessary in order to apply these penalties, but still, the
nature of these penalties remains the same relative to humans and to AI robots.
“I, ROBOT - I, CRIMINAL”--WHEN SCIENCE FICTION BECOMES REALITY: LEGAL LIABILITY OF AI ROBOTS COMMITTING CRIMINAL OFFENSES, 22 Syracuse Sci. & Tech. L. Rep. 1.
Hallevy fails to address a significant issue, namely, that an AI likely does not have a life span. To potentially send an AI to prison then, may not serve the same goals as it would for human beings. Indeed, when a lifetime may last forever, a couple years incapacitated by incarceration does not seem nearly as harsh as when a lifespan is relatively short. A court, however, may legally require that a convict take medically necessary medication to ensure he is not a danger to himself or others. This likely also applies to AI where the “medication” may be a software fix and the “medical condition” is some defect in its programming causing it to commit criminal acts.
Isaac Asimov set down three fundamental laws of robotics in “I, Robot”: (1) a robot may not injure a human being or, through inaction, allow a human being to come to harm; (2) a robot must obey the orders given to it by human beings, except where such orders would conflict with the First Law; and (3) a robot must protect its own existence, as long as such protection does not conflict with the First or Second Laws. If a country adopts such a standard as necessary in any AI, Is it enough to simply “remind” a culpable robot of these laws
III. Tort Law
i. Products Liability
Restatement (Second) of Torts § 402A (1965)
(1) One who sells any product in a defective condition unreasonably dangerous to the user or consumer or to his property is subject to liability for physical harm thereby caused to the ultimate user or consumer, or to his property, if(a) the seller is engaged in the business of selling such a product, and(b) it is expected to and does reach the user or consumer without substantial change in the condition in which it is sold.(2) The rule stated in Subsection (1) applies although(a) the seller has exercised all possible care in the preparation and sale of his product, and(b) the user or consumer has not bought the product from or entered into any contractual relation with the seller.
See Michael D. Scott, Tort Liability for Vendors of Insecure Software: Has the Time Finally Come?, 67 Md. L. Rev. 425, 457 (2008)
Scott argues that strict tort liability should be extended to insecure software. While courts have been reluctant to do so, largely because there is often no physical damage, generally only economic damages such as employee time spent recovering lost data, etc. that can be remedied by contract law. Courts, however, are increasingly recognizing data as property and where damage occurs software vendors may be liable, if not strictly, for their products that damage consumers through lost data and other issues.
With the emergence of AI in many industries, including finances and other industries where it may make critical decision regarding process and transactions, courts may extend strict liability for products not specifically designed for a single customer. Additionally, considering that AI may be present in robots, the damage could be much more than economic. An AI with faulty software could physically harm people or property. It is likely if AI advances to this point, courts may recognize manufacturers responsibility and apply strict liability principles.
As with criminal liability, however, the question of whether an AI is conscious and culpable for its actions, is central to tort liability as well. Additionally, considerations of whether an AI could pay any judgment against it must be addressed. If the
Professional Malpractice Liability for AI creators
Presently, courts are reluctant to extend professional malpractice standards to software engineers. This is true despite the fact that software engineers often undergo extensive education and training, and many companies require certifications. Michael D. Scott, Tort Liability for Vendors of Insecure Software: Has the Time Finally Come?, 67 Md. L. Rev. 425, 472 (2008). For example, in Hosp. Computer Sys., Inc. v. Staten Island Hosp., a federal district court opined that a profession differs from mere business through:
“the requirement of extensive formal training and learning, admission to practice by a qualifying licensure, a code of ethics imposing standards qualitatively and extensively beyond those that prevail or are tolerated in the marketplace, a system for discipline of its members for violation of the code of ethics, a duty to subordinate financial reward to social responsibility, and, notably, an obligation on its members, even in non-professional matters, to conduct themselves as members of a learned, disciplined, and honorable occupation. Hosp. Computer Sys., Inc. v. Staten Island Hosp., 788 F. Supp. 1351, 1361 (D.N.J. 1992)
While a software engineer designing today’s programs may not be a “profession” for purposes of malpractice, the law may evolve in the future to accommodate advanced AI. In the future, software engineers may act as creators of consciousness and free will in the programs they construct. They will set the initial parameters of that consciousness – presumably including the basic morals and rules the AI will apply to its experiences. It is likely the law will require those creating such consciousness do so responsibly, that training will be extensive, that licensing boards will be assembled, who will apply a code of ethics to professionals under their supervision with disciplinary consequences for violations. Should these professional creators violate these professional duties, presumably, the law will permit professional malpractice suits for those damaged by their negligence or malice.
IV. Trusts and Other Financial Instruments
In his article, Legal Personhood for Artificial IntelligencesI, Lawrence Solum explores the possibility of AI acting as trustee for a trust. The program would likely be similar to that Kurzweil described that determines buy and sell decisions for investment funds. Lawrence B. Solum, Legal Personhood for Artificial Intelligences (1992) 70 N.C. L. Rev. 1231, 1240. An AI may serve as a more reliable trustee since a programmer could easily instruct it not to embezzle or steal funds (alternatively, a greedy programmer may install a program similar to that in the movie, Office Space http://en.wikipedia.org/wiki/Office_Space, to steal undetected from the income of a trust by taking fractions of a penny for each transaction.) Solum argues that the inclusion of corporations and governments as trustees establishes that a trustee need not be a natural person, but points out two critical objections to AI acting as a trustee.
First, the responsibility objection, centers on the idea that AI could not be responsible if it breached one of its duties as trustee. As noted above, a software designer may be held liable for its defective product, but Solum explores how an AI itself may be held liable for breach of its duty as trustee. A possible solution to this is to require the AI maintain insurance against such a breach. An insured AI, Solum argues, would serve the ends of liability, to compensate the victim.
Second, the judgment objection, doubts an AI trustee’s competency to make complex judgments. The thrust of the judgment objection is that an expert system consists only of a complex system of rules, which leaves no room for system to make judgments in the sense of exercising discretion. The objection is played out in three versions. First, it is argued that an AI cannot cope with a change of legally relevant circumstances; second, it is argued that an AI cannot make the moral choices it may encounter; and third, it is argued that an AI cannot make some of the legal choices it will face. In all three versions, the problem is that, even in the case of parallel distributed algorithms, an expert system cannot do anything but follow rules. As to the first argument, expert systems seem to lack the kind of common sense needed to solve unexpected problems. As to the second argument, expert systems seem to lack the sense of fairness that is warranted when unexpected circumstances require overruling the letter of a rule in order to serve its purpose. As to the third argument, expert systems seem to lack the ability to take the necessary action if called to account in a court of law.
Solum concludes that AIs presently do not have the capacity to perform the duties of a trustee, especially in the case of unexpected circumstances affecting the trust. He raises the question whether a more limited form of legal personhood could be designed, allowing an AI to serve as a limited purpose trustee and/or for simple trusts whose operation can be fully automatic.82 In that case, the terms of the trust will need to specify a human take-over whenever unanticipated circumstances rule out automatic behavior.83 We note that Solum seems to restrict himself here to automatic devices. Where autonomic computing is concerned, it seems that responsiveness to changed circumstances is part of its definition: even if the system cannot but follow rules, it is supposed to be capable of adjusting the rules that determine its performance. The first objection may thus fail in the case of autonomic devices. As to the third objection, this also applies to corporations and funds to which legal personhood has been attributed. This leaves the second objection as the only real objection with regard to autonomic computer agents. Bert-Jaap Koops, Mireille Hildebrandt, David-Olivier Jaquet-Chiffelle, Bridging the Accountability Gap: Rights for New Entities in the Information Society?, 11 Minn. J.L. Sci. & Tech. 497, 500 (2010)
Where a nation has a significant population of robots operating on AI, the implications of warfare targeted at that population are interesting. Consider a scenario where a nation attacks another’s AI population, infecting it with a virus instructing the AI population to turn against its nation and Asimov’s basic laws. If the virus is effective, the country releasing it essentially has a standing AI army ready to blindly follow orders. The AI population is used as a weapon against its host population.
Where the AI robots are considered conscious, is this biological warfare? Do international conventions exist that this type of warfare? Where the robot population is not considered conscious, but rather as property, is this covered under hacking statutes?
a. AI as Property/Infrastructure
In 2009, the US government acknowledged its computer network controlling its national power grid had been hacked into. Allegedly both Russian and Chinese agents hacked the system in an attempt to locate weaknesses in the system. The visitors refrained from damaging the system, but left behind hidden software capable of bringing the system down. http://digitaljournal.com/article/303531#ixzz1s2R7rHDb
This type of hacking is a serious concern for any computer network, and is particularly disconcerting for a country with a large AI population presumably open to viral and potential hacking attacks. A viral or hacking attack on a country with a sizeable AI population could be devastating. Suppose another country released a contagious virus on a country’s AI population, causing the AI to turn against the human residents of the country. . Presently, there are no international conventions preventing hacking, however, in 2010, the Munich Conference on Security Policy addressed the issue. http://www.securityconference.de/TOP-NEWS.425+M581b9350a37.0.html?&L=1&no_cache=1&sword_list=cyberwar The conference recognized the growing concern of hacking in international conflict and acknowledged the importance of digital infrastructure to a country’s continued well-being. In particular, one commentator noted that any successful hacking policy would require the harmonization of hacking laws in the international community and the need for strong interaction between private and public sectors. Id.
In the United States, the federal computer fraud and abuse statute, 18 U.S.C. 1030, outlaws conduct that victimizes computer systems. It is a cyber security law. It protects federal computers, bank computers, and computers connected to the Internet. It shields them from trespassing, threats, damage, espionage, and from being corruptly used as instruments of fraud. http://www.fas.org/sgp/crs/misc/97-1025.pdf Hacking laws are not only relevant in international law, but in the criminal law as well. Where a person installs an unwanted program on an AI and uses that AI to commit an unlawful act, the person could be liable, as discussed above,
b. AI as Beings
i. Biological warfare?
If a country releases a viral attack on another country’s AI population, whether it be in robot form or otherwise, this may be considered biological warfare, especially where AI is perceived as conscious and retains the rights of a natural person. The Biological Weapons Convention (BWC) effectively prohibits the development, production, acquisition, transfer, stockpiling and use of biological and toxin weapons and is a key element in the international community’s efforts to address the proliferation of weapons of mass destruction. The BWC opened for signatures in 1972 and came into force in 1975. Presently, the BWC has 165 States Parties and 12 signatories. There are 19 states which have neither signed nor ratified the Convention. http://www.unog.ch/80256EE600585943/%28httpPages%29/04FBBDD6315AC720C1257180004B1B2F?OpenDocument ; http://www.unog.ch/80256EDD006B8954/%28httpAssets%29/699B3CA8C061D490C1257188003B9FEE/$file/BWC-Background_Inf.pdf Thus, most countries have agreed biological warfare will not be tolerated. Could this same convention be applied to an AI population?
It is likely that if a country, or more importantly, a consensus of world states perceives AI as worthy of protections similar to natural persons, then the BWC may be extended to AI, or a similar convention may be drafted specifically for AI.