Digital Analogues (part 5): Lessons from Animal Law, Continued


The last post in this series on “Digital Analogues”–which explores the various areas of law that courts could use as a model for liability when AI systems cause harm–examined animal liability law.  Under traditional animal liability law, the owner of a “wild” animal is strictly liable for any injury or damage caused by that animal.  For domesticated animals, however, an owner is only liable if that particular animal had shown dangerous tendencies and the owner failed to take adequate precautions.

So what lessons might animal liability law offer for AI? Well, if we believe that AI systems are inherently risky (or if we just want to be extra cautious), we could treat all AI systems like “wild” animals and hold their owners strictly liable for harms that they cause. That would certainly encourage safety precautions, but it might also stifle innovation.  Such a blanket rule would seem particularly unfair for AI systems whose functions are so narrow that they do not present much risk to anyone. It would seem somewhat silly to impose a blanket rule that treats AlphaGo as if it is just as dangerous as an autonomous weapon system.

» Read more

Digital Analogues (Part 4): Is AI a Different Kind of Animal?

Source: David Shankbone


The last two entries in this series focused on the possibility of treating AI systems like “persons” in their own right.  As with corporations, these posts suggested, legal systems could develop a doctrine of artificial “personhood” for AI, through which AI systems would be given some of the legal rights and responsibilities that human beings have.  Of course, treating AI systems like people in the eyes of the law will be a bridge too far for many people both inside the legal world and in the public at large.  (If you doubt that, consider that corporate personhood is a concept that goes back to the Roman Empire’s legal system, and it still is highly controversial)

In the short-to-medium term, it is far more likely that instead of focusing on what rights and responsibilities an AI system should have, legal systems will instead focus on the responsibilities of the humans who have possession or control of such systems. From that perspective, the legal treatment of animals provides an interesting model.

» Read more

Digital Analogues (Part 3): If AI systems can be “persons,” what rights should they have?


The last segment in this series noted that corporations came into existence and were granted certain rights because society believed it would be economically and socially beneficial to do so.  There has, of course, been much push-back on that front.  Many people both inside and outside of the legal world ask if we have given corporations too many rights and treat them a little too much like people.  So what rights and responsibilities should we grant to AI systems if we decide to treat them as legal “persons” in some sense?

Uniquely in this series, this post will provide more questions than answers.  This is in part because the concept of “corporate personhood” has proven to be so malleable over the years.  Even though corporations are the oldest example of artificial “persons” in the legal world, we still have not decided with any firmness what rights and responsibilities a corporation should have.  Really, I can think of only one ground rule for legal “personhood”: “personhood” in a legal sense requires, at a minimum, the right to sue and the ability to be sued.  Beyond that, the meaning of “personhood” has proven to be pretty flexible.  That means that for the most part, we should be able decide the rights and responsibilities included within the concept of AI personhood on a right-by-right and responsibility-by-responsibility basis.

» Read more

Digital Analogues (Part 2): Would corporate personhood be a good model for “AI personhood”?

Source: Biotwist (via Deviant Art)


This post is part of the Digital Analogues series, which examines the various types of persons or entities to which legal systems might analogize artificial intelligence (AI) systems. This post is the first of two that examines corporate personhood as a potential model for “AI personhood.”  It is cross-posted on the website of the Future of Life Institute.  Future posts will examine how AI could also be analogized to pets, wild animals, employees, children, and prisoners.


Could the legal concept of “corporate personhood” serve as a model for how legal systems treat AI?  Ever since the US Supreme Court’s Citizens United decision, corporate personhood has been a controversial topic in American political and legal discourse.  Count me in the group that thinks that Citizens United was a horrible decision and that the law treats corporations a little too much like ‘real’ people.  But I think the fundamental concept of corporate personhood is still sound.  Moreover, the historical reasons that led to the creation of “corporate personhood”–namely, the desire to encourage ambitious investments and the new technologies that come with them–holds lessons for how we may eventually decide to treat AI.

An Overview of Corporate Personhood

For the uninitiated, here is a brief and oversimplified review of how and why corporations came to be treated like “persons” in the eyes of the law.  During late antiquity and the Middle Ages, a company generally had no separate legal existence apart from its owner (or, in the case of partnerships, owners).  Because a company was essentially an extension of its owners, owners were personally liable for companies’ debts and other liabilities.  In the legal system, this meant that a plaintiff who successfully sued a company would be able to go after all of an owner’s personal assets.

This unlimited liability exposure meant that entrepreneurs were unlikely to invest in a company unless they could have a great deal of control over how that company would operate.  That, in turn, meant that companies rarely had more than a handful of owners, which made it very difficult to raise enough money for capital-intensive ventures.  When the rise of colonial empires and (especially) the Industrial Revolution created a need for larger companies capable of taking on more ambitious projects, the fact that companies had no separate legal existence and that their owners were subject to unlimited liability proved frustrating obstacles to economic growth.

The modern corporation was created to resolve these problems, primarily through two key features: legal personhood and limited liability.  “Personhood” means that under the law, corporations are treated like artificial persons, with a legal existence separate from their owners (shareholders).  Like natural persons (i.e., humans), corporations have the right to enter into contracts, own and dispose of assets, and file lawsuits–all in their own name.  “Limited liability” means that the owners of a corporation only stand to lose the amount of money, or capital, that they have invested in the corporation.  Plaintiffs cannot go after a corporate shareholder’s personal assets unless the shareholder engaged in unusual misconduct. Together, these features give a corporation a legal existence that is largely separate from its creators and owners.

» Read more

Digital Analogues (Intro): Artificial Intelligence Systems Should Be Treated Like…

This piece was originally published on Medium in Imaginary Papers, an online publication of Arizona State University’s Center for Science and the Imagination.  It is also cross-posted on the website of the Future of Life Institute.  Full credit to Corey Pressman for the title.


Artificial intelligence (A.I.) systems are becoming increasingly ubiquitous in our economy and society, and are being designed with an ever-increasing ability to operate free of direct human supervision. Algorithmic trading systems account for a huge and still-growing share of stock market transactions, and autonomous vehicles with A.I. “drivers” are already being tested on the roads. Because they operate with less human supervision and control than earlier technologies, the rising prevalence of autonomous A.I. raises the question of how legal systems can ensure that victims receive compensation if (read: when) an A.I. system causes physical or economic harm during the course of its operations.

An increasingly hot topic in the still-small world of people interested in the legal issues surrounding A.I. is whether an autonomous A.I. system should be treated like a “person” in the eyes of the law. In other words, should we give A.I. systems some of the rights and responsibilities normally associated with natural persons (i.e., humans)? If so, precisely what rights should be granted to A.I. systems and what responsibilities should be imposed on them? Should human actors be assigned certain responsibilities in terms of directing and supervising the actions of autonomous systems? How should legal responsibility for an A.I. system’s behavior be allocated between the system itself and its human owner, operator, or supervisor?

» Read more

Tay the Racist Chatbot: Who is responsible when a machine learns to be evil?

Warning


By far the most entertaining AI news of the past week was the rise and rapid fall of Microsoft’s teen-girl-imitation Twitter chatbot, Tay, whose Twitter tagline described her as “Microsoft’s AI fam* from the internet that’s got zero chill.”

(* Btw, I’m officially old–I had to consult Urban Dictionary to confirm that I was correctly understanding what “fam” and “zero chill” meant. “Fam” means “someone you consider family” and “no chill” means “being particularly reckless,” in case you were wondering.)

The remainder of the tagline declared: “The more you talk the smarter Tay gets.”

Or not.  Within 24 hours of going online, Tay started saying some weird stuff.  And then some offensive stuff.  And then some really offensive stuff.  Like calling Zoe Quinn a “stupid whore.”  And saying that the Holocaust was “made up.”  And saying that black people (she used a far more offensive term) should be put in concentration camps.  And that she supports a Mexican genocide.  The list goes on.

So what happened?  How could a chatbot go full Goebbels within a day of being switched on?  Basically, Tay was designed to develop its conversational skills by using machine learning, most notably by analyzing and incorporating the language of tweets sent to her by human social media users. What Microsoft apparently did not anticipate is that Twitter trolls would intentionally try to get Tay to say offensive or otherwise inappropriate things.  At first, Tay simply repeated the inappropriate things that the trolls said to her.  But before too long, Tay had “learned” to say inappropriate things without a human goading her to do so.  This was all but inevitable given that, as Tay’s tagline suggests, Microsoft designed her to have no chill.

Now, anyone who is familiar with the social media cyberworld should not be surprised that this happened–of course a chatbot designed with “zero chill” would learn to be racist and inappropriate because the Twitterverse is filled with people who say racist and inappropriate things.  But fascinatingly, the media has overwhelmingly focused on the people who interacted with Tay rather than on the people who designed Tay when examining why the Degradation of Tay happened.

» Read more

Analysis of the USDOT’s Regulatory Review for Self-Driving Cars (Part 1): References to “drivers” in the federal regulations

Editor’s Note: Apologies for the unannounced gap between posts.  I have been on parental leave for the past two weeks bonding with my newborn daughter.  In lieu of the traditional cartoon, I will be spamming you today with a photo of Julia (see bottom of post).  Now, back to AI.


The U.S. Department of Transportation recently released a report “identifying potential barriers and challenges for the certification of automated vehicles” under the current Federal Motor Vehicle Safety Standards (FMVSS).  Identifying such barriers is essential to the development and deployment of autonomous vehicles because the manufacturer of a new motor vehicle must certify that it complies with the FMVSS.

The FMVSS require American cars and trucks to include numerous operational and safety features, ranging from brake pedals to warning lights to airbags.  It also specifies test procedures designed to assess new vehicles’ safety and whether they comply with the FMVSS.

The new USDOT report consists of two components: (1) a review of the FMVSS “to identify which standards include an implicit or explicit reference to a human driver,” which the report’s authors call a driver reference scan; and (2) a review that evaluates the FMVSS against “13 different automated vehicle concepts, ranging from limited levels of automation . . . to highly automated, driverless concepts with innovative vehicle designs,” termed an automated vehicle concepts scan.  This post will address the driver reference scan, which dovetails nicely from my previous post on automated vehicles.

As noted in that post, the FMVSS defines a “driver” as “the occupant of a motor vehicle seated immediately behind the steering control system.”  It is clear both from this definition and from other regulations that “driver” thus refers to a human driver.  (And again, as explained in my previous post, the NHTSA’s recent letter to Google did not change this regulation or redefine “driver” under the FMVSS, media reports to the contrary notwithstanding.)  Any FMVSS reference to a “driver” thus presents a regulatory compliance challenge for makers of truly self-driving cars, since such vehicles may not have a human driver–or, in some cases, even a human occupant.

» Read more

Will technology send us stumbling into negligence?

Two stories that broke this week illustrate the hazards that can come from our ever-increasing reliance on technology.  The first story is about an experiment conducted at Georgia Tech where a majority of students disregarded their common sense and followed the path indicated by a robot wearing a sign that read “EMERGENCY GUIDE ROBOT”:

A university student is holed up in a small office with a robot, completing an academic survey. Suddenly, an alarm rings and smoke fills the hall outside the door. The student is forced to make a quick choice: escape via the clearly marked exit that they entered through, or head in the direction the robot is pointing, along an unknown path and through an obscure door.

The vast majority of students–26 out of the 30 included in the experiment–went where the robot was pointing.  As it turned out, there was no exit in that direction.  The remaining four students either stayed in the room or were unable to complete the experiment.  No student, it seems, simply went out the way they came in.

Many of the students attributed their decision to disregard the correct exit to the “Emergency Guide Robot” sign, which suggested that the robot was specifically designed to tell them where to go in emergency situations.  According to the Georgia Tech researchers, these results suggest that people will “automatically trust” a robot that “is designed to do a particular task.”  The lead researcher analogized this trust “to the way in which drivers sometimes follow the odd routes mapped by their GPS devices,” saying that “[a]s long as a robot can communicate its intentions in some way, people will probably trust it in most situations.”

As if on cue, this happened the very same day that the study was released:

» Read more

Who’s to Blame (Part 5): A Deeper Look at Predicting the Actions of Autonomous Weapons

Dilbert

Source: Dilbert Comic Strip on 2011-03-06 | Dilbert by Scott Adams


An autonomous weapon system (AWS) is designed and manufactured in a collaborative project between American and Indian defense contractors. It is sold to numerous countries around the world. This model of AWS is successfully deployed in conflicts in Latin America, the Caucuses, and Polynesia without violating the laws of war. An American Lt. General then orders that 50 of these units be deployed during a conflict in the Persian Gulf for use in ongoing urban combat in several cities. One of those units had previously seen action in urban combat in the Caucuses and desert combat during the same Persian Gulf conflict, all without incident. A Major makes the decision to deploy that AWS unit to assist a platoon engaged in block-to-block urban combat in Sana’a. Once the AWS unit is on the ground, a Lieutenant is responsible for telling the AWS where to go. The Lt. General, the Major, and the Lieutenant all had previous experience using this model of AWS and had given similar orders to these in prior combat situations without incident.

The Lieutenant has lost several men to enemy snipers over the past several weeks. He orders the AWS to accompany one of the squads under his command and preemptively strike any enemy sniper nests it detects–again, an order he had given to other AWS units before without incident. This time, the AWS unit misidentifies a nearby civilian house as containing a sniper nest, based on the fact that houses with similar features had frequently been used as sniper nests in the Caucuses conflict. It launches a RPG at the house. There are no snipers inside, but there are 10 civilians–all of whom are killed by the RPG. Human soldiers who had been fighting in the area would have known that that particular house likely did not contain a sniper’s nest because the glare from the sun off a nearby glass building reduces visibility on that side of the street at the times of day that American soldiers typically patrol the area–a fact that the human soldiers knew well from prior combat in the area, but a variable that the AWS had not been programmed to take into consideration.

In my most recent post for FLI on autonomous weapons, I noted that it would be difficult for humans to predict the actions of autonomous weapon systems (AWSs) programmed with machine learning capabilities.  If the military commanders responsible for deploying AWSs were unable to reliably foresee how the AWS would operate on the battlefield, it would be difficult to hold those commanders responsible if the AWS violates the law of armed conflict (LOAC).  And in the absence of command responsibility, it is not clear whether any human could be held responsible under the existing LOAC framework.

A side comment from a lawyer on Reddit made me realize that my reference to “foreseeability” requires a bit more explanation.  “Foreseeability” is one of those terms that makes lawyers’ ears perk up when they hear it because it’s a concept that every American law student encounters when learning the principles of negligence in their first-year class on Tort Law.

» Read more

Who’s to Blame (Part 4): Who’s to Blame if an Autonomous Weapon Breaks the Law?

accountability-joke


The previous entry in this series examined why it would be very difficult to ensure that autonomous weapon systems (AWSs) consistently comply with the laws of war.  So what would happen if an attack by an AWS resulted in the needless death of civilians or otherwise constituted a violation of the laws of war?  Who would be held legally responsible?

In that regard, AWSs’ ability to operate free of human direction, monitoring, and control would raise legal concerns not shared by drones and other earlier generations of military technology.  It is not clear who, if anyone, could be held accountable if and when AWS attacks result in illegal harm to civilians and their property.  This “accountability gap” was the focus of a 2015 Human Rights Watch report.  The HRW report ultimately concluded that there was no plausible way to resolve the accountability issue and therefore called for a complete ban on fully autonomous weapons.

Although some commentators have taken issue with this prescription, the diagnosis seems to be correct—it simply is not clear who could be held responsible if an AWS commits an illegal act.  This accountability gap exists because AWSs incorporate AI technology could collect information and determine courses of action based on the conditions in which they operate.  It is unlikely that even the most careful human programmers could predict the nearly infinite on-the-ground circumstances that an AWS could face.  It would therefore be difficult for an AWS designer–to say nothing of its military operators–to foresee how the AWS would react in the fluid, fast-changing world of combat operations.  The inability to foresee an AWS’s actions would complicate the assignment of legal responsibility.

» Read more

1 2