Until
next week, 
Emotional Machines: A review of Rosalind Picard’s Affective Computing (1997, MIT Press).
Dateline: July 5, 1998
In discussing emotion in machines in a previous feature, I drew attention to the contribution that Dr. Rosalind Picard of the MIT Media Lab was making to the debate. That contribution has since been revealed more fully in Dr. Picard's recent book, Affective Computing. Clicking on this link will enable you to order the book, and if you are (like me) an interested layperson who follows developments in AI, or are a researcher or student of emotional computing, you would be well advised to do so.
It is a well-written book, sensibly organized to appeal to both lay and academic audiences, and summarizes the state of the very young field of affective computing. "Affect" and "affective" are Academia-speak for "emotion" and "emotional," and for the rest of this review I will use the terms most familiar to us.
The book's manifesto is "a call for a change in computing," a recognition that emotion is as necessary for truly intelligent machines as it is for truly intelligent people, and a call to the computer science community to "give" emotion to machines. In the first part of the book Dr. Picard presents evidence from the cognitive sciences that emotion is indeed a component of intelligence, as well as a discussion of what emotion consists of and what mechanisms a machine would need to have in order to have emotion. The second part of the book addresses the "how to" of endowing machines with the mechanisms for emotion, and is particularly good reading for students and scientists.
Since I am not (formally, at any rate) either a student or a scientist, I will however dispose briefly of part 2 of the book by noting that in it Dr. Picard has pulled together all the significant efforts to introduce emotion into computers, and describes the various approaches and models researchers have adopted so far as well as the major research problems remaining to be resolved.
As a comprehensive overview of the field, it seems to me to be imperative reading for anyone involved in, or wanting to become involved in, the development or study of emotional computing. I know of no other source where it has all been pulled together so concisely yet so meticulously (including even an exemplary algorithm for the emotion of joy). Since I also share her belief that emotion is a key component of intelligence, real or artificial, it probably behooves researchers in any branch of AI to read her book.
Now back to the first and (to me, with my bias toward the philosophy of AI, more interesting) part of the book.
"A computer," she says, "can be said to have emotion if it has the following five components that are present in healthy human emotional systems:" [my interpretation is in square brackets]
Where I find the book particularly interesting is in the descriptions of recent efforts to give emotions to machines (part 2 of the book) and in the discussion of the philosophical aspects of doing so. With regard to the latter, I would not find fault with 99 percent of what Dr. Picard writes, but I tend to think she underplays the role of emergence, though she does acknowledge its possibility. This was probably her conscious choice, since she is first and foremost a scientist and her book is more concerned with the science of emotional machines than with the philosophy and metaphysics of emergent phenomena (although there is a body of science a-building that is beginning to unravel the mechanisms of emergence.)
Noting that children acquire and learn to control emotions over time and through social interaction, and the consequent likelihood that we will need to give machines the same innate tools to acquire and learn to control their own emotions, she appears to implicitly recognize that emotion will emerge given the right conditions; that we do not need to program in every smile and every teardrop.
One danger of an emotional machine, says Dr. Picard, is that it may be used (passive voice) for dark purposes, such as secretly monitoring our emotional state and manipulating it. It might persuade us to rush out and buy a truckload of gizmos we don't need or vote for a politician promising to nuke all undesirable elements in society (these are my examples, not from the book). What she does not say, but I do, is that a smart machine might use (active voice) its knowledge of and ability manipulate our emotions for its own purposes.
Indeed, the book as a whole suggests that machines will continue to be "human-centered," even though they may "develop their own mechanisms of emotion" and even "become better than some people when it comes to recognizing emotions." She acknowledges that if a designer were to let a computer evolve its own emotions, it is possible that non-human emotions could emerge, which we would be unable to recognize. In the same vein (but this is a different point) because of the different physiology of the machine from the human, it is unlikely that a machine could ever "feel" emotions the way we do. It does not have a gut, for instance, and is therefore unlikely to experience the feeling of being hit in the stomach on receipt of really bad news.
Machines can also have sensory capabilities that go beyond our own. It can have infra-red vision, for example, and convert the images into something we too can see. This is true and already of enormous benefit to us, but it ignores the question of whether a free-willed machine would necessarily want to share its sensory experiences with us.
Dr. Picard raises the related questions of whether we can trust an emotional computer to always tell us the truth, and whether we could risk allowing the computer to have negative emotions, such as dislike for its user. With probably more control over its emotions than we do over our own, and given the common assumption that only humans lie, the computer would find it easier than us to "lie with a straight face." Why would it want to lie to us? Perhaps because it feels anger or contempt towards us. Dr. Picard asks rhetorically whether society would allow computers to have negative emotions toward humans.
To me, such dangers are preventable only as long as emotion is a strictly programmed property of the machine and not an emergent one. I argue that we can no longer think in terms of switching off an intelligent machine if it is not performing according to our liking, because increasingly machines are connected to the Internet. Switching a machine off does no good if its programming and memory has been distributed around the Net—which is almost inevitable given that a smart machine, aware of its vulnerability, will have cloned itself all over the Net.
Dr. Picard herself makes the insightful comment that "Emotional expressions, once recognized by a computer, are no longer photons that merely travel into the heavens, but bits that circulate over networks and are trivially copied and sold . . . ." Seen in that light, the issues Dr. Picard raises are much more significant, but as long as we underestimate the power of emergence and go on believing in the on/off switch, the significance is lost.
A further issue she raises is privacy from the machines prying sensors. We often try to hide our emotions, often for good reasons. A "poker face" is the quintessential example. Do we want our wearable computer to let Interflora know that we've just fallen madly in love and are ripe for an offer to send a $50 bouquet to the object of our affection?
This leads to a more fundamental issue: "Perhaps the most ominous scenario with any digital information is that of some powerful centralized organization using the data in a pernicious way." Related to this is the issue of losing control of our own emotions, since if a machine (or organization) knows our true emotion, it can then manipulate us. An example of such manipulation might be "for pacifying a violent criminal, when he exceeds some anger threshold."
But even this relatively benign application is of debatable value (not to mention debatable ethics). "In humans, emotion almost never completely determines behavior; it merely biases it. Anger does not cause violence, but it predisposes one to choosing this kind of action," she writes. Furthermore, in connection with the question of whether negative emotions should be permitted for machines themselves, Dr. Picard points out that negative emotions (for example, grief) can play positive roles in helping an intelligent being adapt to a complex and changing environment.
Recognizing that the issue of free will is "a neglected part of this problem" and "is going to become increasingly important once they have emotions," she briefly discusses potential moral behaviors of an emotional, free-willed machine. As a machine (she uses the more restrictive and less frightening term "software agent") "learns its own ideas of valence and salience, it becomes increasingly difficult for humans to predict what it might decide in a situation," and will need to be given "rigorous tests" before being put into a "position of authority."
Here I must respectfully disagree with Dr. Picard. I have long maintained, in these features, that we cannot control a machine that is smarter than us and free-willed, and even if we could, it would not guarantee a good outcome. Adolph Hitler might have passed with flying colors a rigorous test of leadership. However, I should emphasize that Dr. Picard's book does not endorse my view that the machine will become smarter and free-willed, though it is clear she does not discount the possibility. She raises, for example, the question of whether or not a free-willed machine should ethically be considered our servant (and therefore free to walk off the job at any time) or our slave (and therefore not).
Although she accepts the possibility of emergence—and explicitly notes the unpredictability inherent to complex systems (which is essentially another way of saying emergence)—her basic stance on both emotion and free will in machines seems to be that these attributes are entirely within in our power to bestow or withhold, and she lays down some guidelines for designers: Should computers be allowed to hide their emotions? Should they be given both primary and cognitive emotions? What sort of event should be allowed to trigger their primary emotions? Should they be programmed with human values? Should they be allowed to be swayed by emotion in others? Should they be allowed to feel misery, contempt, anger, or other dangerous emotions? "The summarizing question," she states, is "Should computer emotions be implemented as closely as possible to what we know about human emotions, or should they be designed differently . . . ?"
These are all normative questions—questions of what "should" happen, by design. To me, if one accepts the possibility of emergence, then one must accept the possibility that all these questions could be rendered moot by unpredictable developments. But she is right on the money when she says: "The greater the freedom of a machine, the more it will need moral standards. . . . A system that truly operates in a complex and unpredictable environment will need more than laws; it will essentially need values and principles, a moral compass for guidance, and perhaps even religion." Then she spoils it somewhat by adding: "On top of these issues looms the question of who dictates these moral codes. Who has moral authority over computers, robots, software agents, and other computational things?"
The question I would ask, rather, is: What can we do to help ensure that an acceptable standard of morality emerges in the machine, alongside its emotions and its free will?
Affective Computing is the sort of work which, by describing
the field under discussion, thereby helps to shape and define it. The research
and development efforts she writes about in part 2 of the book will grow
rapidly dated, and even the cognitive and philosophical issues she raises
in part 1 will need revision as the cognitive scientists and philosophers
shed more light on them. I hope we can look forward to a regular review
of the state of emotion in machines from this first-rate writer and scholar.
Until
next week,
NEXT WEEK: Probably another issue of EEEK!, Ellis's Endangered Employment Katalog.
Help Wanted: Got questions or comments on this article or on any other AI-related subject under the sun? Post it in the AIBB!