Several years ago I participated on a panel, hosted by the Kennedy Center for Human Development at Vanderbilt University, that discussed the possibility of conscious computers, their rights and roles. I didn't appreciate until midway through the panel that this bunch was probably not interested in conscious computers at all, but rather they were using the hypothetical, conscious computer as a surrogate for the mentally and physically handicapped, the terminally ill, perhaps even the fetus. The topic allowed a glancing discussion of issues such as personhood, health care and patient rights, abortion, without the emotional baggage that generally accompanies such topics.
In retrospect, the Kennedy-Center panel never addressed the ethical choices that a truly conscious, intelligent computer would force society to face. No one on the panel, which was composed of a philosopher, a psychologist, and a computer scientist, thought that human-level machine intelligence was a relevant concern in anything but the very long term. Rather, the perception of machine intelligence, whether faulty or accurate, was the relevant issue. The philosopher distinguished 'persons' from 'humans' -- there are certain characteristics of `personhood' that are not present in all humans, for example those with advanced Alzheimers he supposed, nor would these characteristics be present in machines. To paraphrase the philosopher's position:
"I and other responsible individuals have thought deeply about the distinctions between persons and non-persons. The distinguishing characteristics are such that they can and should inform treatment."If all people looked deeply, then I might have some sympathy for this view, but I subscribe to a view that
"The vast majority of people do not think deeply about person and non-person differences, but base judgements about personhood on surface features of being human. If society doesn't recognize differences between persons and non-persons because it doesn't look, then we need to be very careful about substantive differences in treatment."Today, I want to continue with my reasoning of several years ago, and explore the theological implications of machines and programs that much of humanity falsely perceives as intelligent.
Each of you probably received a flyer advertising the conference, with some questions in the margin that we might consider. I show these on the left, together with some corresponding questions stemming from a different slant:
Can actions taken be a machine, WILL actions taken be a machine, even if they are "automatic," even if they are "automatic," still be considered to have be considered BY MANY to have moral or religious worth? moral or religious worth? Do any "automatic" human Do any "automatic" human actions have moral worth? actions have moral worth TO SOME OF US? How close to human would a How close to human would a machine have to be before we machine have to be before CHILDREN felt the need to consider PARENTS its religious standing? PATIENTS DOCTORS PARISHIONERS CLERGY ...... consider its religious standing? What are the prospects of What human-like behaviors creating a machine that WILL BE taken as religiously exhibits religiously- relevant? BY WHOM? relevant human-like behaviors?As someone that is at the conference to represent Artificial Intelligence (AI), I have to tell you that I, and I suspect many AI researchers, have a public and a private face. When asked about what I do by family, old high school friends, students, and others not familiar with the field, I say something like:
I work in artificial intelligence, which is a field of computer science concerned with developing computers that exhibit intelligence; that can do the kind of mental tasks that people do, blah blah blah.This is a gross over-simplification that is designed to make a long conversation go away so that the questioner doesn't regret the question. Its an answer that everyone can think they understand, without really understanding, and it seems to work, because the reply is almost always something like "how interesting". The truth is that while interest in human-comparable intelligence motivated much of the very early work in artificial intelligence ( 1, 2) by young, very intelligent, very excited, and very naive individuals, and it motivated my own entry into the field after reading several Issac Asimov robot novels during one Summer break, but operationally AI researchers of my generation almost never seriously think of AI in these terms, though many of us probably entered the field because of Isaac Asimov.
AI has become largely an engineering discipline, but before talking about this engineering perspective of AI, I want to address the human-centered perspective of AI that probably dominates the views of those not in the field, particularly those not in academia. Those of us in AI let the cover story stand -- it is too easy for everyone concerned to rely on intuitions about human intelligence to explain artificial intelligence, leading to the anthropomorphisms that I suspect will contribute to theological implications of falsely-perceived artificial intelligences.
The Turing Test: A test of examiner, not examinee
An anthropocentric view of the world is so strong that the question of whether a machine is intelligent is often converted, with no justification necessary, to a question of whether a machine has human-comparable intelligence. In the the Turing Test, a judge interacts with a human or a machine. The computer is judged intelligent if the judge cannot determine whether the machine is human or machine after a protracted interaction, or if some number of judges incorrectly identify machine as human, or some variant of these criteria.
There are some problems with the Turing Test, as suggested by many (e.g., Hayes and Ford, 1996). Notably, the Turing Test is as much a test of the judge as it is the machine -- perhaps the judge simply didn't try hard enough or know enough to discriminate the human from machine.
Consider ELIZA a program developed by Joseph Weizenbaum in the late 60's and early 70's that takes the role of a psychoanalyst in a psychiatric interview. Here is a sample dialog, which I have simply made up -- but it captures the flavor of a real dialog I think:
ELIZA is a relatively simple pattern-matching program; it takes keywords from a subject's response and inserts these into one of a set of `skeletal' responses to the subject. It is an illusion that ELIZA understands a subject's statements. Weizenbaum found, however, that users were taking the dialog quite seriously. An amusing anecdote found Weizenbaum walking in on his secretary in the middle of a dialog with ELIZA, and the secretary asked him to leave so that the dialog could continue in privacy. A colleague seriously suggested that ELIZA-like systems could be used in place of real psychoanalysts. Needless to say, this scared Weizenbaum a bit. Apparently, users of ELIZA were easily fooled by the illusion that ELIZA comprehended the discussion. These users were judges that simply didn't work very hard.
Hayes and Ford point to the Loebner competition, a yearly event that pits machines against humans in Turing Tests. Eventually a $10,000 (?) prize may be claimed by a machine that passes a general Turing Test. Currently, contestants (machines) are allowed to compete in circumstances where the topic of discourse is very limited (e.g., the O. J. Simpson case). The winner of the Loebner prize might deliberately misspell words, back up over text to correct it, and type at `human' speed. Hayes and Ford argue that the development of such "artificial con-artists" is a result of using the Turing Test as a measure of success in AI. Ideally, researchers in the field of Artificial Intelligence should not pursue such demonstrations, but such tricks may be marketable.
Humanoid robots are the most likely candidates for fallacious humanization, but the challenges of perception and motor control have made it tough to create such machines convincingly in the first place (but see MIT's COG). Nonetheless, a question for psychologists is what are the features of robots that elicit the strongest reactions in us.
Successful intelligence/emotional mimicry may be more immediately possible by software robots, or softbots. These softbots might be regarded as expert systems that operate in a software environment such as a computer's operating system, or as part of a Web site. Consider a diet-planning softbot, which helps me plan my diet so that it adheres to certain nutritional and culinary constraints. Tools like this already exist. Of course, such specialized intelligences are far from general, but if we put the same talking head interface on many such "idiot savants" (Hayes and Ford), such as the diet advisor, a chess playing program, an artificial psychoanalyst, then how believable will be the illusion of general intelligence. Even to someone such as myself, the illusion may be quite believable, though I know that it is probably created by a relatively uninteresting conjoining of disparate systems, covered by the same high-tech, talking-head veneer.
Frankly, while I think that human-comparable machine intelligence is possible, it is not probable for a very long time (on the order of 100 years or more??). It is not too difficult to imagine, however, that machines and programs in the very near future will present powerful illusions of intelligence, and perhaps other aspects of personhood. The illusion will be different to different populations. The Turing Test that people like those at this conference will use to assess the technology, will not be the Turing Test that everyone uses. Some of us will be tenacious, knowledgeable judges. The majority of people will be neither tenacious or knowledgeable.
I don't want to underestimate machines that are powerful enough to sway our theology -- these machines of the future will undoubtedly be very impressive by today's standards, even if they fail a Turing Test by the most interested judges. However, we should not underestimate even simplistic, anthropomorphic influences either. Ive debated using several examples here to illustrate the power of cursory differences of appearance. It hasn't been easy to find non-trivial examples concerning humans and machines, without drawing from science fiction, but I will try the famous television footage of an encounter between a solitary man and a tank at Tienanmen Square. In all commentary that I have heard, the episode is a man/tank encounter, with the man illustrating courage and principle. But there are really two men here, one disguised as a tank. I don't want to take anything away from the man standing alone, but for once I would like it acknowledged that the person driving the tank probably or possibly demonstrated something worthy in stopping, in trying to detour around the man rather than crushing him, or even more compassionately in simply not forcing the man to move aside. I think that this example illustrates a twist on my point, but fundamentally the same point: many of us are persuaded by cursory appearances. In the near future, I suspect that the human appearance of a machine will persuade us to humanize it to some extent, but as the Tienanmen Square example illustrates, in the very long term (by my figuring), the machine appearance of a person (but not a human) will persuade us as well. I wonder if intelligent machines will be foils for human heros.
The Engineering Perspective of AI
While I think that many of the people can be fooled much of the time, I don't want to leave the impression that the illusion of machine intelligence or morality will stem simply from parlor tricks. Machines and programs can and will do impressive things, which most of us familiar with the technology would not call intelligent or moral, but which may nonetheless help cement the perception of both in the minds of many people.
The "intellectual" feats of current machines stem from their ability to explore vast numbers of possible actions and outcomes, choosing to take those actions that lead to the best outcomes. The most popular example is in the domain of chess, where computer programs such as Deep Blue may examine up to millions of possible board positions before choosing to make a move that appears to lead to the most favorable outcome for the machine.
Data mining applications also illustrate the computer's prodigious ability to explore many situations. An organization like WalMart may record millions of customer transactions a day. So-called market-basket analysis examines these transactions for items that tend to be bought together more than would be expected "by chance." Some co-occurrences are obvious to human analysts (e.g., baby food and diapers tend to be bought together), but some are less so (e.g., ice cream and Diet Coke). Having discovered such co-occurrences, WalMart can plan advertisting, sales, coupons, and store layout.
In general, AI is most appropriate in solving problems that are (1) hard to solve because so many possible steps may need to be considered, but (2) the solutions, once found, are "easy" to understand. A classic example of such a problem (labeled as NP by computer theoreticians for reasons that I won't get into), is the traveling salesman problem:
Is there a tour of client businesses in a strange city that is no more than N miles long?This is a hard problem because I can follow a certain route that appears promising, only to hit an unanticipated deadend or one-way street, thus requiring an annoyingly-roundabout route that makes the length of my complete tour greater than the allowable distance of N.
Once I find a tour that satisfies the stated constraints, however, I can repeat it as necessary, and easily communicate the route to another salesman. Roughly speaking, chess and data mining have this hard-to-solve, easy-to-comprehend property too -- easy-to-comprehend in the sense that I can verify that the solution is valid: the chess game was won via a sequence of legal moves, or a particular pattern of purchases is more common than I might expect if there were certain independence assumptions underlying the purchase of the individual items of the pattern.
Of course, there may be more to "validating" an answer in the chess and data mining domains than simply affirming the legality of the solution. I may want to know why certain moves were made when they were, and the criticality of each decision in leading to the final outcome, or why certain products are purchased in conjunction such as ice cream and Diet Coke.
While this "explanation/validation" step often requires a hefty bunch of exploration too, it often requires specialized knowledge that a human user possesses, and would be difficult to give a program. Interactive induction is an increasingly popular paradigm that pairs a computer program that explores a space of patterns using "weak" or relatively "generic" heuristics of interestingness or informativeness, and a human analyst who explains/validates in domain-specific terms the "nuggets" discovered by the program. For example, a data mining application in which I was involved, mitigated a long-standing process delay in the printing industry (Evans and Fisher, 1994; Handbook of Data Mining, 1999). At one plant of R R Donnelley and Sons, a process delay known as banding had reached a pace of 500-600 bands per year by 1989. A band is a line across a printed image like the magazine page pictured here, and slows the printing process considerably.
Data was collected along approximately 35 dimensions -- several hundred data points corresponding to cases when a band occurred, and several hundred data points corresponding to cases when printing jobs had run to completion without a band. A data mining program in an interactive induction context was used to discover and explain a number of patterns that distinguish cases in which banding is likely and in which banding is unlikely (e.g., when chrome solution ratio is high and ink temperature is low and ink viscosity is high, banding is highly unlikely). Bands have decreased dramatically over the years, and in 1997 there were fewer than 20 bands at the R R Donnelley plant in which this work began.
I do not think that AI programs per se are sophisticated enough at this point to tell us much about human nature (see the Conference pamphlet's questions), but I think that there is much to learn about human nature by watching humans interact with AI programs in interactive contexts. At least when users know something of the data being mined, human analysts have a healthy skepticism about the findings of the computer. In large part, interactive induction includes humans in the data analysis loop so that humans buy into the results of machine exploration.
An important result of interactive induction is that it demands discipline on the part of the analysts that use it. To explore data, a machine requires data. Banding had been a problem for decades in the printing industry, anecdotal solutions were tried and abandoned, but it was not until an interactive induction approach was taken that a systematic collection of data was performed.
In principle, I believe that the knowledge contributed by a human analyst in an interactive-induction setting could be encoded and provided to a machine. This is the dream of developers of systems such as Cyc, which has the goal of building programs that exploit encyclopedic knowledge. While this is probably a necessary step towards human-comparable intelligence, from an engineering standpoint this is not as cost effective as the interactive induction compromise, in which machine and human does what each does best: a machine explores and prunes a vast space of actions/conditions using weak heuristics, thus focusing the analyst's attention on a few conditions that need be explained in terms of specialized knowledge. The apparent general intelligence of conjoined, specialized intelligences, with like-appearing interfaces is what we will more probably see in the near term.
I am also ambivalent about Cyc-like approaches to building increasingly sophisticated, human-comparable machines for another reason -- they are more likely to forgive sloppy thinking. There are some fundamental lessons that are ideally learned, even in an introductory computer programming class: your thinking is almost always incomplete, and the world's stupidist entity capable of "speech" (i.e., a computer) is telling you so. The same lessons are highlighted in interactive induction settings, though the emphasis in this latter context is to push/encourage an analyst to hypothesize new knowledge. There are probably theological implications of stupid and intelligent machines alike, relating to the nature of faith and the tenacity with which we question results from what appears to be a black box.
I have spent some time discussing the area of AI in which my work focuses, notably machine learning and data mining. The possibility of intelligent machines learning about the spending habits of individuals, medical history, and the like is considerable. I also want to point out that humanoid robots will likely be performing tasks, including heart surgery, nuclear power cleanup, emergency medical response, that may promote anthropomorphism. (how much on this theme?)
I believe in theory that we can achieve human-comparable intelligence and morality in a machine. But practically speaking, I think that the theological implications of a comet hitting the Earth in the years before we acquire machine intelligence is a more relevant issue. Still, impressive computer power and substantial anthropomorphism will lead to machines that may influence our theology. Perhaps because I was a devotee of Asimov's early on, I think that the misconceptions and anthropomorphisms may have a positive impact on morality. I'll use some favorite sayings to introduce some misconceptions about machines that may nonetheless promote them as positive role models.
Jesus said, "Love your brother like your soul, guard him like the pupil of your eye." (Gospel of Thomas saying 25)I love this suggestion that love should be as automatic as the blink of your eye. Can (should) automatic actions be regarded as having moral worth? I think that they can if automaticity is the endpoint of an evolutionary struggle within an individual who seeks an ideal state. But I think that the more pressing question is, will automaticity of action (e.g., saving a life) by a humanoid machine be viewed by many as having moral worth? I think that the answer is "yes", even when you and I know that it will likely result from programming Asimov's laws of robotics in some direct fashion, and not as the result of some moral growth/struggle by the machine.
"The lion hath roared; who will not fear? The Lord Yahweh hath spoken; who will not prophesy?" (Amos, 3:7)Several years ago I discovered the Old Testament prophets, and read them feverishly for a time. My favorite has always been Amos, and the image of God in that book, standing by wall built with a plumb line, sweating, in khaki pants and a flannel shirt, and sleeves rolled to the elbows. But whether one is talking about Amos' call for justice, Isaiah walking naked for three years at God's command, or Micah's injunction to do justice, love kindness, and walk humbly with God, I find the purity of the prophets' obedience to God remarkable. I am not familiar with the Jewish tradition, but taking the prophets as standalone texts, their obedience seems unconditioned on any promise of personal reward -- this is not carrot theology. Again, machines may exhibit obedience, which is falsely perceived as stemming from caring selflessness, but they may be influential role models nonetheless.
Jesus said, ``If a blind man leads a blind man, they will both fall into a pit'' (Gospel of Thomas saying 34; also Matthew 15:14, Luke 6:30)An apparently intelligent and moral machine may serve as a positive role model, but the danger of course is that the deception will be discovered. What are the theological implications of that discovery?
Finally, I think that saying 29 from the Gospel of Thomas perfectly captures our fascination with AI and theology:
Jesus said, "If the flesh came into being because of spirit, then it is a wonder. But if the spirit came into being because of the body, it is a wonder of wonders. Indeed, I am amazed at how this great wealth has made its home in this poverty." (Gospel of Thomas saying 29)The theological implications of machines may stem from surprises over the years about the apparent wealth of spirit that a machine can attain, and disappointment, possibly dissolution, as we realize our self-deceptions.