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

Who’s to Blame (Part 6): Potential Legal Solutions to the AWS Accountability Problem

The law abhors a vacuum.  So it is all but certain that, sooner or later, international law will come up with mechanisms for fixing the autonomous weapon system (AWS) accountability problem.  How might the current AWS accountability gap be filled?

The simplest solution—and the one advanced by Human Rights Watch (HRW) and the not-so-subtly-named Campaign to Stop Killer Robots (CSKR)—is to ban “fully autonomous” weapon systems completely.  As noted in the second entry in this series, the HRW defines such an AWS as one that can select and engage targets without specific orders from a human commander (that is, without human direction) and operate without real-time human supervision (that is, monitoring and control). One route to such a ban would be adding an AWS-specific protocol to the Convention on Certain Conventional Weapons (CCW), which covers incendiary weapons, landmines, and a few other categories of conventional (i.e., not nuclear, biological, or chemical) weapons. The signatories to the CCW held informal meetings on AWSs in May 2014 and April 2015, but it does not appear that the addition of an AWS protocol to the CCW is under formal consideration.

In any event, there is ample reason to question whether the CCW would be an effective vehicle for regulating AWSs. The current CCW contains few outright bans on the weapons it covers (the CCW protocol on incendiary weapons does not bar the napalming of enemy forces) and has no mechanisms whatsoever for verification or enforcement.  The CCW’s limited impact on landmines is illustrated by the fact that the International Campaign to Ban Landmines (which, incidentally, seriously needs to hire someone to design a new logo) was created nine years after the CCW’s protocol covering landmines went into effect.

Moreover, even an outright ban on “fully” autonomous weapons does not adequately account for the fact that weapon systems can have varying types and degrees of autonomy.  Serious legal risks would still accompany the deployment of AWSs with only limited autonomy, but those risks would not be covered by a ban on fully autonomous weapons.

A more balanced solution might require continuous human monitoring and adequate means of control whenever an AWS is deployed in combat, with a presumption of negligence (and therefore command responsibility) attaching to the commander responsible for monitoring and controlling an AWS that commits an illegal act.  That presumption could only be overcome if the human being shows   This would ensure that at least one human being would always have a strong legal incentive to supervise an AWS that is engaged in combat operations.

An even stronger form of command responsibility based on strict liability might seem tempting at first, but applying a strict liability standard to command responsibility for AWSs would be problematic because, as noted in the previous entry in this series, multiple officers in the chain of command may play a role in deciding whether, when, where, and how to deploy an AWS during a particular operation (to say nothing of the personnel responsible for designing and programming the AWS).  It would be difficult to fairly determine how far up (or down) the chain of command and how far back in time criminal responsibility should attach.


Much, much more can and will be said about each of the above topics in the coming weeks and months.  For now, here are a few recommendations for deeper discussions on the legal accountability issues surrounding AWSs:

  • Human Rights Watch, Mind the Gap: The Lack of Accountability for Killer Robots (2015)
  • International Committee of the Red Cross, Autonomous weapon systems technical, military, legal and humanitarian aspects (2014)
  • Michael N. Schmitt & Jeffrey S. Thurnher, “Out of the Loop”: Autonomous Weapon Systems and the Law of Armed Conflict, 4 Harv. Nat’l Sec. J. 231 (2013)
  • Gary D. Solis, The Law of Armed Conflict: International Humanitarian Law in War (2015), chapters 10 (“Command Responsibility and Respondeat Superior“) and 16 (“The 1980 Certain Conventional Weapons Convention”)
  • U.S. Department of Defense Directive No. 3000.09 (“Autonomy in Weapon Systems”), issued Nov. 21, 2012
  • Wendell Wallach & Colin Allen, Framing Robot Arms Control, 15 Ethics and Information Technology 125 (2013)

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