Is Medical Artificial Intelligence Ethically Neutral?

Will Knight has written several articles over this past year in MIT’s flagship journal Technology Review that have discussed growing concerns in the field of Artificial Intelligence (AI) that may be of concern for bioethicists. The first concern is in the area of bias. In an article entitled “Forget Killer Robots – Bias is the Real AI Danger”, Knight provides real world examples of this hidden bias affecting people negatively. One example is an AI system called COMPASS, which is used by judges to determine the likelihood of reoffending by inmates who are up for parole. An independent review claims that algorithm may be biased against minorities. In a separate article, Knight identified additional examples in other AI algorithms that introduced gender or minority bias in software used to rank teachers, approve bank loans and interpret natural language processing. None of these examples argued that this bias was introduced intentionally or maliciously (though that certainly could happen).

This is where Knight’s the second concern becomes apparent. The problem may be that the algorithms are too complex for even their programmers to retroactively examine for bias. To understand the complexity issue, one must have an introductory idea of how the current AI programs work. Previously, computer programs had their algorithms “hard-wired” so to speak. The programs were essentially complex “if this, then do this” sequences. A programmer could look at the code and generally understand how the program would react to a given input. Beginning in the 1980’s, programmers started experimenting with code written to behave like a brain neuron might behave. The goal of the program was to model a human neuron, including the ability of the neuron to change its output behavior in real time. A neurobiologist would recognize the programming pattern as modeling the many layers of neurons in the human brain. A biofeedback expert would recognize the programming pattern as including feedback to change the input sensitivities based upon certain output goals – “teaching” the program to recognize a face or image in a larger picture is one such example. If you want to dive deep here, begin with this link.

This type of programming had limited use in the 1980s because the computers were too simple and could only model simple neurons and only a limited number at one time. Fast forward to the 21st century and 30 years of Moore’s Law of exponential growth in computing power and complexity, and suddenly, these neural networks are modeling multiple layers with millions of neurons. The programs are starting to be useful in analyzing complex big data and finding patterns (literally, a needle in a haystack) and this is becoming useful in many fields, including medical diagnosis and patient management. The problem is that even the programmers cannot simply look at these programs and explain how the programs came to their conclusions.

Why is this important to consider from a bioethics standpoint? Historically, arguments in bioethics could generally be categorized as consequentialist, deontological, virtue, hedonistic, divine command, etc… One’s stated position was open to debate and analysis, and the ethical worldview was apparent. A proprietary, cloud-based, black-box, big data neural network system making a medical decision obscures, perhaps unintentionally, the ethics behind the decision. The “WHY” of a medical decision is as important as the “HOW”. What goes in to a medical decision often includes ethical weighting that ought to be as transparent as possible. These issues are presently not easily examined in AI decisions. The bioethics community therefore needs to be vigilant as more medical decisions begin to rely on AI. We should welcome AI as another tool in helping us provide good healthcare. Given the above concerns regarding AI bias and complexity, we should not however simply accept AI decisions as ethically neutral.

AI and the Trolley Car Dilemma

I have always hated the Trolley Car dilemma. The god of that dilemma universe has decreed that either one person or five people will die as a result of an energetic trolley car and a track switch position that only you control. Leave the switch in place and five people are run over by the trolley. Pulling the switch veers the trolley onto an alternate track, successfully saving the original five people but causing the death of a different lone person on the alternate track. Your action or inaction in this horrific Rube Goldberg contraption contributes to the death of either one or five people. Most people I know feel some sort of angst at making their decision. MIT has a website that allows you to pull the switch, so to speak, on several different variations of this dilemma and see how you compare with others who have played this game, if you enjoy that sort of thing.

Kris Hammond, Professor of Computer Science and Journalism at Northwestern, believes a robot would handle the trolley car problem far better than a human since they can just “run the numbers and do the right thing”. Moreover, says Professor Hammond, though we “will need them to be able to explain themselves in all aspects of their reasoning and action…[,his] guess is that they will be able to explain themselves better than we do.” Later in the article he claims that it is the very lack of angst regarding the decision-making process that makes the robot superior, not to mention the fact that the robot, as in the case of self-driving cars, would avoid placing us in the dilemma in the first place by collectively being better drivers.

For the sake of today’s blog, I am willing to grant that second claim to focus on the first: Is there really lack of angst and, if so, does that lack contribute to making the robot’s decision right and therefore superior?

Currently, no robot has sufficient artificial intelligence that might allow for self-awareness sufficient to create angst. Essentially, a robot lacks independent agency and as such cannot be held morally accountable for any actions resulting from its programming. The robot’s programmer certainly does and can. Presumably (hopefully) the programmer would feel some angst, at least eventually, when he or she reviews the results of the robot’s behavior that resulted directly from his or her program. Is the displaced decision-making really advantageous? Is the calculus inherent in the encoded binary utilitarian logic really that simple?

Watson, IBM’s artificial intelligence system, can finally best some human chess grand masters. Chess is a rule-based game with a large but not infinite set of possible moves. Could a robot really be programmed to handle every single variation of the trolley car dilemma? Are the five individuals on the first track or the single individual on the second track pastors, thieves, or some weird combination of both, one of whom recently saved your life? Should any of that matter? Who gets to decide?

Trolley car dilemmas seem to demand utilitarian reasoning. Robots are arguably great at making fast binary decisions so if the utilitarian reasoning can be broken down into binary logic, a robot can make utilitarian decisions faster than humans, and certainly without experiencing human angst. Prof Hammond claims the robots will simply “run the numbers and do the right thing”. But the decisions are only right or superior if we say they are.

Utilitarian decision-making is great if everyone agrees on the utility assigned to every decision.  But this is clearly not the case, as the summary results on the MIT website clearly show.  Further, I think that most normal people have angst over their own decisions in situations like these, even inconsequential decisions offered on MIT’s harmless website.  So in the case of the robot, the angst doesn’t occur when the robot is actualizing its program – it occurred months or years ago when the programmer assigned values to his or her utilitarian decision matrix.

Were those the right values? (hint: there is angst here)
Who gets to decide? (hint: even more angst here)

CGI Turing Test

[Star Wars fans spoiler alert: The following contains potential story information from “Rogue One: A Star Wars Story”, the Star Wars Episode IV prequel]

I confess that I am a Stars Wars geek in particular and a science fiction movie buff in general. Like many, I am old enough to have seen the first Star Wars movie at its 1977 release, before it was re-indexed as “Episode IV: A New Hope”. The computer generated imagery or CGI special effects in that movie revolutionized the science fiction genre. It is now commonplace to use CGI to accomplish all manner of special effects, transporting moviegoers into all sorts of fantastic virtual worlds and virtual characters that appear, frankly, real. Rogue One has taken CGI up to the next level with one particular character such that I would argue that Rogue One has passed what I am calling the CGI Turing Test.

The original Turing test was described by Alan Turing, a famous British mathematician who designed and built a mechanical computer in the 1940s that successfully decoded the Nazi Enigma machine, a previous unbreakable encoding device that had thwarted Allied efforts to eavesdrop on the Nazi military communications. The Turing test is commonly misconstrued as a test of a computer’s (artificial) intelligence, which it is not. It is actually a test to determine whether a computer can imitate a human well enough to convince an actual human that it (the computer) is human. This test was a variant of a party game known as the “Imitation Game” in which a man (person A) and a woman (person B) would try to convince a third party, called the interrogator (person C) who was in a separate room, that each was the other. The Turing test substitutes a computer for person A.

Rogue One plays a similar game. There is a character in the Star Wars films named Grand Moff Tarkin, a very evil general in the Empire played by British actor Peter Cushing. Cushing debuted his Grand Moff Tarkin character in the original 1977 Star Wars movie. He is again seen reprising this role in the new 2016 Rogue One installment. I thought he was as awesome as ever. Except that he wasn’t. Peter Cushing died 22 years ago in 1994. I promise if you watch Rogue One and put yourself in the role of person C, the interrogator, you will be convinced that the CGI Peter Cushing (person A) is the real Peter Cushing (person B). So, the Academy Award® for Best Actor in a supporting role goes to…a computer at Industrial Light & Magic?

What has this to do with bioethics in general or artificial intelligence in particular? Perhaps not much. The futurist Ray Kurzweil argued in his book “The Singularity is near” that a machine will pass the Turing test in 2029 and perhaps this will come true, though his previous predictions have been called into question. In keeping with this AI/Turing Test theme, I gave the gift of “Google Home” and “Alexa” to different family members this Christmas. I was pleasantly amazed by the speech recognition of both systems and fully expect the technology to rapidly improve. Despite this, the forgoing discussion, and the knowledge that Turing and Kurzweil both disagree with me, I remain convinced that our ability to create a computer to imitate a human, the Imago Hominis, so to speak, will always fall far short of His ability to create a human to reflect Himself, the Imago Dei.

As the interrogator, what do you think?

I am – is it?

This past summer, researchers at RPI’s Cognitive Science Department programmed three Nao robots to see if they could pass a test of self-awareness. Modeled after the classic “Wisemen Puzzle”, the robots were asked whether or not they had been given a “dumbing pill” (in this case, a tap on their head, which muted their verbal output) or a placebo. The test not only required the robots to respond to a verbal question (“Which pill did you receive?”) but also recognize its own voice as distinct from the others and correctly respond (“I was able to prove that I was not given the dumbing pill”). For a $9500 retail robot, this is an impressive artificial intelligence (AI) test and worth watching HERE.

Dr. Selmer Bringsjord, lead investigator and chair of the Cognitive Science Department at RPI is careful to point out that these robots have been programmed to be self-conscious in a specific situation and describes his work as making progress in logical and mathematical correlates to self-consciousness. His biography page on the RPI faculty website provides a rather tongue-in-cheek assessment of the results of his research: “I figure the ultimate growth industry will be building smarter and smarter such machines on the one hand, and philosophizing about whether they are truly conscious and free on the other. Nice job security.”

I believe philosophizing about whether the robots are truly self-conscious to be the more interesting topic. In their current form, while the robot appears to a human observer to be self-aware, it is really the algorithm or program that correctly indicates (realizes?) that the robot did not receive the dumbing pill. But the algorithm itself is not aware that it correctly determined which pill the robot received. One could make the algorithm more complex, such that the algorithm tests whether the algorithm correctly determined which pill the robot received. But would that algorithm really be aware that the algorithm was aware which pill the robot received? One can see the infinite regression building. (Google: “It’s turtles all the way down”)

Perhaps the more interesting question is how we humans will react as the robot AI algorithms appear more self-aware, whether or not they actually are. Taking Dr. Bringsjord’s lead, should I invest in the domain name “spcr.org”* now or give it some more time?

 

* Society for the Prevention of Cruelty to Robots