Artificial
Intelligence: Legal Research
Contents
III. Defining Artificial Intelligence
C. Objectives of
Artificial Intelligence
i. The
Perpetration-by-Another Liability Model: AI Robots as Innocent Agents
ii. The
Natural-Probable-Consequence Liability Model: Foreseeable Offenses Committed
by AI Robots
iii. The Direct Liability Model: AI Robots as
Direct Subject of Criminal Liability
This research assumes that eventually a form of artificial
intelligence (AI) will exist that is on par with human intelligence. Further,
we should assume that:
i.
We will employ
intelligent agents specifically for their creative intelligence and
corresponding judgment;
ii.
These agents will make
decisions, and
iii.
Some of the resultant
decisions will be ÒpathologicalÓ; that is, not only unpredictable and
surprising, which may be desirable, but also having unintended and destructive
consequences both for networks and for the substantial portions of our
infrastructure connected to those networks.[1]
This research contemplates the anticipated legal issues that
will arise if AI does reach such a level, and then attempts to develop theories
of legal liability for offenses committed by an AI.
á
If an entity is aware of
life enough to assert its rights by protesting its dissolution, is it entitled
to the protection of the law?[2]
á
Should a self-conscious
computer facing the prospect of imminent unplugging have standing to bring a
claim of battery?[3]
á
If we agree that a
machine could potentially be a candidate for rights, we still must answer:
Which machines and which rights? What would a computer have to do to deserve
legal or moral personhood?[4]
á
With complex computer
systems consisting of a combination of overlapping programs created by different
coders, it is often difficult to know who should bear moral blame or legal
liability for a computer action that produces an injury. [5]
á
Computers often play
major roles in writing their own software; what if one created a virus and sent
it around the world?[6]
á
Computers now help
operate on us and help us handle our investments; should we hold them as
accountable as we do our surgeons and financial analysts when they screw up?[7]
á
Can society impose
criminal liability upon robots? If so, how do you punish an AI robot?[8]
á
Does the growing
intelligence of AI robots subject them to legal social control, just as any
other legal entity?[9]
á
How can AI entities
fulfill the two requirements of criminal liability (i.e., actus reus and mens rea)?[10]
á
If an AI entity can be held
criminally liable for an offense, can it raise a defense of insanity in
relation to a malfunctioning AI algorithm, when its analytical capabilities
have become corrupted as a result of that malfunction?[11]
May it assert a defense of being under the influence of an intoxicating
substance (similar to humans being under the influence of alcohol or drugs) if
its operating system is infected by an electronic virus?[12]
There are five attributes that one would expect an
intelligent entity to have:[13]
1.
Communication.
á
The easier it is to
communicate with an entity, the more intelligent the entity seems.
2.
Mental knowledge.
á
An intelligent entity is
expected to have some knowledge about itself.
3.
External knowledge.
á
An intelligent entity is
expected to know about the outside world, to learn about it, and utilize that
information.
4.
Goal-driven behavior.
á
An intelligent entity is
expected to take action in order to achieve its goals.
5.
Creativity.
á
An intelligent entity is
expected to have some degree of creativity. In this context creativity means
the ability to take alternate action when the initial action fails.
Software agents are rules-based software products which may
aid with internet searches, filter incoming electronic mail, find the
appropriate area of a help program in online documentation, watch for news on
topics you have specified, and suggest changes to your stock portfolio.[14]
Advances in the area of artificial intelligence are
producing Òlearning algorithmsÓ that will soon produce software agent products
that are based on neural networks. These products, the Òsecond generation of
intelligent software agents,Ó will be able to learn from their experiences and
adapt their behavior accordingly. Software programs with this ability to learn
will, consequently, be capable of decision-making, resulting in software that
may take actions that neither the licensor nor the licensee anticipated.[15]
The objectives of AI scientists may be grouped into two
broad categories: the development of computer-based models of intelligent
behavior, and the pursuit of machines capable of solving problems normally
thought to require human intelligence.[16]
Through empathy or arrogance, or perhaps through an instinct
for ethical consistency, we tend to seek rights for things that appear to be
like us and to deny rights to things that donÕt.[17]
The list of threshold characteristics proposed for a
computer to have legal or moral personhood is exhaustive: the ability to
experience pain or suffering, to have intentions or memories, and to possess
moral agency or self-awareness. None of these characteristics is well-defined,
though, and this is especially the case with the most oft-cited of the lot:
consciousness.[18]
It is most likely that a machine that has the ability to
interact with humans in the world will be the first candidate for rights.[19]
One possibility would be to treat A.I. machines as valuable
cultural artifacts, to accord them landmark status with stipulations about
their preservation and disassembly.[20]
Intelligent software agents will be capable of causing harm.
Unlike earlier software agents, they will be capable of finding their own
sources of information and making commitments – possible unauthorized
commitments.[21]
Where software is capable of unplanned – rather than predictable – behavior, the law
will therefore need to be applied without factual precedent to these
situations.[22]
Agency law provides a suitable framework in which to find a
solution for harms committed by the next generation of intelligent software; an
agency relationship is formed when the software licensee installs and then
executes the software program. Accordingly, intelligent software agents should
be regulated under agency law.[23]
Agency law provides an adequate framework because, in
general, a software licensee will be activating software for some purpose. The
intelligent software agent will then use its learning, mobility and autonomous
properties to accomplish specific tasks for the licensee. Thus, we see the
software agent in the legal role of the Òagent,Ó and the software licensee in
the legal role of the Òprincipal.Ó[24]
This relationship of agent/principal has been formed whether or not the parties
themselves intended to create an agency or even think of themselves as agent
and principal.[25]
Once intelligent software agents are viewed as legal agents
within an agency relationship, it follows that liability can be attributed to
the actions of the software agents, binding the software licensee (principal)
to legal duties.[26]
It appears that the courts are willing to consider machines
as participants in ordinary consumer transactions.[27]
By extending this previous court treatment of transactions
conducted by autonomous machines, or by using the liability theories available
under contract (where the action of an agent bind the principal to third
parties), or tort (where the principal may be vicariously liable for the
actions of the agent), intelligent software agents can be treated as Òlegal
agents.Ó[28]
Once intelligent software agents are viewed as having legal
status with formation of the agency relationship, it follows that liability can
be attributed to the actions of the software agents and the licensees will
consequently be held responsible. Certain legal duties vest in the licensees as
principals in the newly formed agency relationship. Therefore, any harm
committed by the intelligent software agent that breaches the licenseeÕs legal
duties will cause the licensee to be liable.[29]
In order to impose criminal liability upon a person, two
main elements must exist. The first is the factual element, i.e., criminal
conduct (actus reus), while the other
is the mental element, i.e., knowledge or general intent in relation to the
conduct element (mens rea). If one of
them is missing, no criminal liability can be imposed.[30]
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 that conduct.[31]
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.[32]
Gabriel Hallevy has proposed that AI entities can fulfill
the two requirements of criminal liability under three possible models of
criminal liability: (i) the Perpetration-by-Another liability model; (ii) the
Natural-Probable Consequence liability model; and (iii) the Direct liability
model.
This model does not consider the AI robot as possessing any
human attributes. Instead, the model theorizes that AI entities are akin to
mentally limited persons, such as a children, and therefore do not have the
criminal state of mind to commit an offense.[33]
The AI robot is viewed as an intermediary that is used as an instrument, while
the party orchestrating the offense is the real perpetrator (hence the name,
perpetration-by-another).[34]
The person controlling the AI, or the perpetrator, is regarded as a principal
in the first degree and is held accountable for the conduct of the innocent
agent (the AI).[35]
The perpetratorÕs liability is determined on the basis of that conduct and his
own mental state.[36]
The AI robot is an innocent agent.[37]
This model would likely be implemented in scenarios where
programmers have programmed an AI to commit an offense, or where a person
controlling the AI has commanded it to commit an offense.[38]
This model would not be suitable when the AI robot decides to commit an offense
based on its own accumulated experience or knowledge.
This model of criminal liability assumed deep involvement of
the programmers or users in the AI robotÕs daily activities, but without any
intention of committing an offense via the AI robot.[39]
For instance, one scenario would be when an AI robot commits an offense during
the execution of its daily tasks. This model is based upon the ability of the
programmers or users to foresee the potential commission of offenses; a person
might be held accountable for an offense if that offense is a natural and
probable consequence of that personÕs conduct.[40]
Natural-probable-consequence liability seems to be legally
suitable for situations where an AI robot committed an offense, but the
programmer or user had no knowledge of it, had not intended it and had not
participated in it. [41]
The natural-probable-consequence liability model only requires the programmer
or user to be in a mental state of negligence, not more.[42]
Programmers or users are not required to know about any forthcoming commission
of an offense as a result of their activity, but are required to know that such
an offense is a natural, probable consequence of their actions.[43]
Liability may be predicated on negligence and would be
appropriate in a situation where a reasonable programmer or user should have
foreseen the offense and prevented it from being committed by the AI robot.[44]
The third model does not assume any dependence of the AI
robot on a specific programmer or user. The third model focuses on the AI robot
itself.[45]
This model essentially treats AI robots like human actors, and accordingly, if
an AI robot fulfills the factual element (actus
reus) and mental element (mens rea),
it will be held criminally liable on its own.[46]
The premise is that if an AI robot is capable of fulfilling the requirements of
both the factual element and the mental element, and in fact, it actually
fulfills them, there is presumptively nothing to prevent criminal liability
from being imposed on that AI robot.[47]
The criminal liability of an AI robot does not replace the
criminal liability of the programmers or the users, if criminal liability is
imposed on the programmers and users by any other legal path.[48]
Criminal liability is not to be divided, but rather, added; the criminal
liability of the AI robot is imposed in addition to the criminal liability of
the human programmer or user.[49]
Not only positive factual and mental elements may be
attributed to AI robots; rather, all negative fault elements should be
attributable to AI robots. Most of these elements are expressed by the general
defenses in criminal law (e.g., self-defense, necessity, duress, or
intoxication).[50]
The most obvious theory of tort liability that seems
applicable to injuries caused by artificially intelligent entities is products
liability.[51]
Products liability is the area of law in which manufacturers, distributors,
suppliers, retailers, and others who make products available to the public are
held responsible for the injuries those products cause.[52]
Artificially intelligent entities will presumably be manufactured by a company,
and accordingly the company may be held liable when an AI goes awry.
A manufacturer may be held liable under a negligence cause
of action when an AI causes an injury that was reasonably foreseeable to the
manufacturer. The typical prima facie negligence claim requires that an injured
plaintiff must show (i) that the manufacturer owed a duty to the plaintiff,
(ii) the manufacturer breached that duty, (iii) the breach was the cause in
fact of the plaintiff's injury (actual cause), (iv) the breach proximately
caused the plaintiff's injury, (v) and the plaintiff suffered actual
quantifiable injury (damages).
Alternatively, a manufacturer may be strictly liable for
injuries caused by its product. Strict liability does not require a showing of
negligence, and accordingly, a manufacturer may be liable even if it exercised
reasonable care. Accordingly, the focus of strict liability will primarily be
on whether the a defect in the manufacturerÕs product was a cause of the plaintiffÕs injury.
Injuries caused by an AI raise unique issues under both
theories of products liability. Specifically, we may not be able to determine
if the harm was caused by human or natural agencies; rather, the congruence of
human, natural and technical agencies caused the ultimate harm.[53] Which one do we pick out as the legally
responsible cause? Or should we blame all the causal vectors?
[1] Curtis E.A. Karnow. Liability for Distributed Artificial
Intelligences, 11 Berkeley Tech. L.J.
147, 173 (1996).
[2] Benjamin Soskis. Man and the Machines: ItÕs Time to Start
Thinking about How We Might Grant Legal Rights to Computers, 36 Legal Aff. 37, 37 (2005).
[3] Id. at 38.
[4] Id. at 39.
[5] Id. at 38.
[6] Id.
[7] Id.
[8] 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, 1 (2010).
[9] Id. at 2.
[10] Id. at 9.
[11] Id. at 24.
[12] Id. at 26.
[13] Id. at 4-5; see also
Roger C. Schank, What is an AI, Anyway?,
The Foundations of Artificial
Intelligence 3 (Derek Partridge & Yorick Wilks eds., 2006).
[14] Suzanne Smed. Intelligent Software and Agency Law, 14 Santa Clara Computer & High Tech. L.J.
503, 503 (1998).
[15] Id.
[16] Steven J. Franka. Tort Adjudication and the Emergence of
Artificial Intelligence Software, Suffolk
U. L. Review 623, 625 (1987).
[17] Soskis, supra note 2, at 41.
[18] Id. at 39.
[19] Id. at 40.
[20] Id. at 41.
[21] Smed, supra note 14, at 504.
[22] Id. at 505.
[23] Id. at 504.
[24] Id. at 505; see also Restatement (Second) of Agency ¤ 1
(1958).
[25] Smed, supra note 14, at 505; see also Restatement
(Second) of Agency ¤ 1, cmt. b (1958).
[26] Smed, supra note 14, at 505.
[27] See McEvens v Citibank, 408 NYS 2d 870 (N.Y. County Civ. Ct. 1978)
(bank responsible for money lost by automated teller machine); see also State Farm Mutual Automobile Ins. Co. v. Bockhorst, 453 F.2d 533,
535-536 (10th Cir. 1972) (insurance company was forced to pay a claim that
occurred during a lapse period of customerÕs policy due to computer error); see also Allen v. Beneficial Fin. Co., 531 F.2d 797 (7th Cir. 1976) (bank
did not comply with truth in lending regulations because its computer-generated
explanation regarding load terms was not clear enough for an ordinary borrower
to understand.
[28] Smed, supra note 14, at 506.
[29] Id. at 506-507.
[30] Hallevy, supra note 8, at 7.
[31] Id. at 7.
[32] Id. at 7-8.
[33] Id. at 9-10.
[34] Id. at 10.
[35] Id. at 10.
[36] Hallevy, supra note 8, at 10.
[37] Id. at 11.
[38] Id. at 11.
[39] Id. at 13.
[40] Id. at 14.
[41] Id. at 14.
[42] Hallevy, supra note 8, at 14.
[43] Id. at 14-15.
[44] Id. at 15-16.
[45] Id. at 18.
[46] Id. at 19.
[47] Id. at 19.
[48] Hallevy, supra note 8, at 25.
[49] Id.
[50] Id. at 26.
[51] See Restatement (Third) of
Torts: Products Liability ¤ 1, et. seq. (2011).
[52] Products Liability, Wikipedia
(April 8, 2012), http://en.wikipedia.org/wiki/Product_liability.
[53] Karnow, supra note 1, at 175-176.