Imagine an artificial being with a humanoid body and humanlike senses. Imagine it on its way to carry out a task, perhaps to retrieve a certain book from a certain desk. Watch it walk through its environment, adjusting its path so that it doesn’t collide with people it encounters. Perhaps it recognizes someone, saying “Hello” and greeting the person by name; it might stop others and politely ask permission to scan their faces and ask their names, to hold in its memory. You might also see it giving reasonable answers to questions like “What’s your name?” “Where are you going?” or even—as it consults its built-in wireless connection to the Internet—“Is the stock market up or down, and by how much?” When it finally reaches the right desk, it identifies the book by its color or title, picks it up, and walks back to where it started.

Each action, from walking to seeing to talking, has been described in earlier chapters as a separate achievement, although not necessarily all in one being or as part of a robotic body. Some artificial capabilities have meaning and value even if implemented only on a computer. Artificial vision and speech, for instance, have been goals of AI research apart from their use in robots because they are significant parts of human intelligence and because they have useful applications such as the translation of language by a machine.

But the creature we are imagining uses all these abilities to function in the world. Such sophisticated functioning requires integrated guidance and control of disparate body parts and actions—that is, it requires a brain. In humans, the brain is an enormous collection of neurons that controls our behavior along with the sensory and motor functions of our bodies. In an artificial being, the equivalent is an enormous collection of electronic switches—transistors—etched into silicon to make digital microprocessors and memory chips. This type of brain might be all that is needed to steadily make artificial beings more capable because the speed and capacity of silicon chips continue to grow. But other approaches are possible as well: different forms of electronic brains, brains that combine organic and electronic elements, perhaps even actual living brains inserted into cyborgs.

What a brain does is think, and what thinking imparts to a being, natural or artificial, is intelligence. Whether a being with an artificial brain is actually thinking while electrons course through its circuits is still a matter for debate. It is easier to ask, “Is the being intelligent?” because that question can be answered by observing whether it exhibits intelligent behavior. Alfred Binet, the French psychologist who in 1905 laid the groundwork for the intelligence quotient (IQ) test, defined human intelligence as the sum of mental processes that come into play in adapting the individual to the environment. Modern definitions agree that intelligence is an adaptive property, meaning that it helps the organism survive and thrive by providing effective responses to changing situations.

This general definition, based on the idea of effective responses to the environment, can also be applied to artificial beings. However, in his book Behavior-Based Robotics, Ronald Arkin takes the definition a step further to make it more explicit for robots. Borrowing his version and extending it slightly gives this working definition:

Under this definition, a being that sees its surroundings and interprets them well enough to navigate, and recognizes people and their words

well enough to respond to them, is exhibiting some degree of intelligent behavior.

We can refine our judgment of the intelligence of an artificial being through better understanding of human intelligence and its measurement. Contemporary researchers and educators tend to reject the traditional form of IQ test. They believe that it defines intelligence too stringently by weighing only linguistic and logical-mathematical abilities. Intelligence is now seen as a complex phenomenon with different facets that cannot be summed up by a single IQ number.

One view of this multiplicity comes from the pioneering work of the psychologist Robert Sternberg of Yale University. Sternberg has developed a theory of threefold intelligence that gives weight to analytical, creative, and practical components. Another view comes from Howard Gardner, Hobbs Professor in Cognition and Education at Harvard, who has written extensively about seven diverse types or categories of intelligence. These make a useful grid against which to measure artificial mental functioning.

Gardner begins with two categories that are cornerstones of most definitions of intelligence and adds others that might be less familiar. Paraphrasing Gardner’s definitions, the seven are:

• Linguistic: Sensitivity to language, ability to use language to attain goals

• Logical-mathematical: Capacity for logical analysis and mathematical operations:

• Musical: Skill in recognizing and manipulating musical patterns

• Bodily-kinesthetic: Using one’s body or parts of it to solve problems

• Spatial: Recognizing and manipulating the patterns of space

• Interpersonal: Understanding the intentions and desires of other people and so working effectively with them

• Intrapersonal: Understanding oneself, creating an internal working model of oneself and using it to manage one’s own life

Where Gardner’s intelligences do not exactly match some types of robotic capability, they involve them. Spatial intelligence, for instance—which Gardner describes as the kind of ability used by navigators, sculptors, and surgeons—relies heavily on visual cognition. In the same way, consciousness of self plays a role in intrapersonal intelligence. Some writers have combined the interpersonal and intrapersonal categories into “emotional intelligence,” but it is useful to keep them separate when discussing whether artificial beings can be self-aware.

Gardner’s original definitions include volitional and creative components such as “fashioning products” under bodily-kinesthetic intelligence. My paraphrases omit these components because artificial beings are not yet creative; nevertheless, digital beings do show traces of Gardner’s intelligences. One of them, logical-mathematical ability, is essentially universal among artificial beings because each microprocessor chip has an arithmetic-logic unit (ALU) that accurately manipulates numbers at high speed. The ability to rapidly solve arithmetical problems and deal with areas, such as the stock market, that use arithmetic could be given to a digital being if it were a useful attribute. Artificial logical-mathematical ability could even extend to higher mathematics like algebra and calculus, using existing computer methods that manipulate mathematical ideas at an abstract symbolic level.

But some artificial creatures are more varied in their abilities. In fact, it seems that the seeds of each kind of intelligence can be found in existing beings, as we’ll see by examining three especially intelligent ones. Two we have described before; the third closely parallels the capabilities of our imaginary unit.

ASIMO (advanced step in innovative mobility)—the walking robot made by the Honda Corporation, described in Chapter 6—is the oldest of the three smart beings. It took more than two dozen engineers some 14 years to produce this accomplished walking robot,

which remains among the best. It displays a high level of kinesthetic intelligence. It can keep its balance as it walks on level ground or on a slope, and up and down steps. It can also balance on one leg, turn corners snappily, and walk backward.

Despite ASIMO’s long lineage, though, there is plenty of room for improvement. At ROBODEX 2003, an older model striding along confidently at 1 kilometer per hour (0.6 miles per hour) was out-paced by a newer model that walked faster though less smoothly. With its flexible physical platform that boasts many degrees of freedom, and its humanoid bodily configuration, ASIMO is capable of even more. Honda states that its long-term goal is to make robots that “can be helpful to humans as well as be of practical use in society.” At a height of 1.2 meters (48 inches), ASIMO can interact eye to eye with a person in a wheelchair or sitting up in bed, or with someone sitting at a desk. And with its childlike size, and its back-mounted computer that makes it resemble a student marching off to school with a backpack, ASIMO is quite the opposite of intimidating—it is endearing.

To extend its ability to deal with people, ASIMO’s kinesthetic intelligence has grown, and new communicative—if not linguistic—abilities have been added. Although ASIMO does not yet speak very well, it can listen, react with its body, and read human body language. Call it by name, and it turns its head toward you. It can wave hello, and safely carry out the social gesture of shaking hands with a person. It comes to a halt when it sees a hand held upright, traffic-cop style, understands that a pointing human hand is a directive to go to the indicated location and wait, and recognizes a goodbye wave. Capabilities to identify faces and plot walking routes—that is, elements of interpersonal and spatial intelligence—have been added, and ASIMO can also carry objects in its hands. These capacities, integrated in a single robot, permit ASIMO to act as a receptionist that can greet and recognize visitors and guide them to a specified location.

The initial goal in developing ASIMO was to produce a walking robot; the other abilities were added later. In contrast, the robot Kismet was built to display what its maker calls social intelligence, or in Gardner’s framework, interpersonal intelligence. Cynthia Breazeal,

who, as a graduate student in Rodney Brooks’s group at MIT, constructed the robot, is Kismet’s guiding spirit. When she was a little girl, Breazeal writes, she was “captivated and fascinated” by two “compelling characters” from the 1977 science fiction film Star Wars, the robots R2-D2 and C-3PO. Now she describes her “dream of a sociable robot:”

Breazel argues for the importance of socially intelligent artificial beings, as they enter the human world with potential uses from entertainment to healthcare to military applications. Noting that we humans have become competent at social interaction because dealing with each other has been essential to developing a sophisticated human culture, she suggests that people relate better to technology that displays “rich social behavior.” Socially intelligent beings could become colleagues with whom we communicate as easily as we do with people, and whom we even like having around. An important part of Breazeal’s thinking is the idea that the right kind of interaction will encourage people to teach a being just as they would teach a human child, with important consequences for the future of artificial beings.

Social capability has been built into Kismet, which has no body, hands, or legs, only a head and face meant to interact with people through expressive movements and speech, with appropriate use of vision and hearing. As described earlier, the robot has exaggerated, clownlike features—big blue eyes, prominent lips, and conspicuous animal-like ears—which it moves to convey emotional states that people immediately grasp. Kismet does not look human, however. Breazeal chose not to attempt that illusion because an imperfect simulation of humanity can be disturbing. Instead, the design suggests a fantasy creature that acts believably human, an approach that has been brought to perfection in Walt Disney’s animated offerings.

Further, Kismet’s looks and behavior were chosen to give the impression that it is young. It has big eyes, and its visual system seeks

primary colors and human skin tones and motion, so it reacts if a person shakes a brightly colored toy at it. Its auditory system listens both for words and, through prosodic analysis, the messages behind them, such as elements of praise, and it speaks in childlike tones. Judging by the responses of people who have chatted with the robot, Kismet’s behavior conveys a sense that it is a living being of a tender age. Women especially speak to Kismet in “motherese,” the exaggerated style they might use with a very young child. According to Breazeal, some even empathize with the creature, “often reporting feeling guilty or bad for scolding the robot and making it ‘sad.’”

Equipping Kismet with appropriate facial expressions might seem like a trivial engineering project compared to the difficult one of making ASIMO walk properly, but actually it requires substantial hardware. Moreover, Kismet’s actions must be integrated with the robot’s perceptions to give responses that make sense to people, which requires a large dose of social or interpersonal intelligence.

The electronic brain behind that interpersonal intelligence could not fit into Kismet’s head. It consists of 15 external networked computers that give Kismet “drives”—that is, built-in needs to be with people, and to be stimulated by them—and “emotions.” Each drive has a preferred operating point; for instance, to be neither over- nor understimulated. Depending on how well these ideal conditions are met as Kismet interacts with someone, the robot enters into one of three arousal states—bored, interested, or calm—and its face shows anger, disgust, fear, joy, sorrow, or surprise. Humans respond to these cues by modifying their conversation and actions; for instance, speaking soothingly if Kismet is angry or overstimulated, or waving a toy if it is bored. At their most empathetic, the human responses maintain Kismet’s drives at their ideal points, which puts the robot into a state of well-being; it is not overwhelmed, yet it is challenged to interact further—exactly what a good parent or teacher hopes to give a child. Breazeal has set up a complex, socially based feedback loop between robot and human that works to keep them mutually engaged.

ASIMO was built to emphasize kinesthetic intelligence, and Kismet the interpersonal type. A third smart robot that displays these two

categories of intelligence and more has been developed by the Sony Corporation, maker of the smart robotic dog AIBO. Sony began working on its humanoid Sony dream robot (SDR) in 1997. The latest version, significantly advanced, was originally called the SDR 4X II but now has the catchier name QRIO (Curio). It extends the kinesthetic intelligence of ASIMO and adds interpersonal, linguistic, and even musical intelligence in a unit only 58 centimeters (23 inches) tall. At the ROBODEX 2003 show, the broad abilities of this small unit made it the natural choice as master of ceremonies for the parade of varied robots, most of them much larger, that enlivened the event.

QRIO is literally 50 percent smarter than the preceding model because its brain consists of three microprocessor chips rather than two, all of which are specially optimized for high speed. The third chip provides speech synthesis and recognition, replacing an earlier external computer. The improvement in vocabulary is dramatic, from 20 words to at least 20,000, with which the robot is able to hold simple conversations. It also displays a degree of musical intelligence: It sings as well as talks, and can do so in harmony with other QRIO units. Its social interaction abilities, combining visual cognition and linguistic capability, have also come a long way: It greets people it knows by name. It learns new people by asking them to “Please hold still for a minute” then, using its dual camera stereoscopic vision and hearing, memorizes their faces and names. Once a face is learned, the unit can pick it out from a crowd.

The robot also has excellent kinesthetic intelligence along with a superior physical endowment. As I learned from Masahiro Fuita of Sony’s Digital Creatures Laboratory, one of the prime designers of the robot, specially designed actuators give it a particularly smooth gait and motion. In addition, sensors that feed it kinesthetic information keep it well balanced, even to the extent of adjusting its walking style depending on the surface under its feet. In a demonstration by Sony, one of the unit’s most lifelike acts was to squat down as it moved from one kind of flooring to another, examine the new surface, and then resume walking with a different gait. It is extraordinarily agile as well, moving with enough speed and coordination to dance—another facet

of musical intelligence. And it has sufficient spatial intelligence to plot a path and avoid obstacles as it walks.

QRIO and its predecessor, Sony’s AIBO dog, are carefully crafted to mesh with human expectations. Like Cynthia Breazeal, the design team at the Digital Creatures Laboratory and its consultant Ronald Arkin at Georgia Tech, draw on theories of emotion, and on ethology, the study of how living species behave in their natural settings. Insights from these areas are translated into programming architectures that create believable behavior patterns. (Interestingly, the designers note that ethological information for humans is more limited than for dogs, mostly because of privacy issues.)

These pioneering researchers have constructed a psychic space between human and robot where the person naturally tends to ascribe emotions and other characteristics of living beings to the creature. In addition to behaving in a variety of intelligent ways, the three smart beings, and others, seem to have emotional components too.

With their social capabilities, ASIMO, Kismet, and QRIO all show what look like emotions or frames of mind, although to vastly different extents. ASIMO’s aspect is similar to a spacesuit helmet visor rather than a face, but its confident walk and jaunty hand wave suggest a certain attitude. QRIO has a face, with appealingly big eyes, but the features are immobile; still, it can simulate emotional expressions by singing, dancing, and conversing appropriately.

Kismet shows emotions directly, however, in its face and voice, and as we have seen, these expressive features can make a bond between human and robot. Because Breazeal also uses the words “emotions” and “drives” to describe Kismet’s internal workings, it is natural to ask whether they have meaning to the robot itself—does it have innermost subjective feelings beyond what people ascribe to it? To put it another way, some people empathize with Kismet, feeling bad when they make it “sad.” But is anything like sadness going on inside Kismet? And might such feelings hint at intrapersonal intelligence, Gardner’s seventh category?

To Breazeal, the answer is unequivocally no. She writes, “Kismet is not conscious, so it does not have feelings…. That Kismet is not conscious (at least not yet) is [Breazeal’s] philosophical position.” Rodney Brooks agrees. “It is all very well,” he writes,

Brooks’s viewpoint is echoed by other researchers, such as Rosalind Picard of the MIT Media Lab. Picard is a pioneer in affective computing; as she defines it, “computing that relates to, arises from, or deliberately influences emotions.” She believes that “Computers do not have feelings in the way that people do … computers simply aren’t built the way we are to have that kind of an experience.”

Coming from robotics and computer experts, these comments about the lack of internal emotional states suggest that intrapersonal intelligence does not exist among artificial beings. But there are reasons to think that at least a low-level intrapersonal component is achievable, and that such an advance would represent an opening wedge for self-awareness, as we shall see later. Whether real or not, however, some familiarity with emotions can be more than a frill for artificial beings. If a being can smile at you, and recognize your own smile, your interaction is likely to go more smoothly than without these human attributes. This is also true for human–computer interactions: Picard notes that the appearance of emotions could, for instance, enhance computerized tutoring.

For artificial beings that move in the world, something akin to emotion can even be a survival factor. In his book Robot: Mere Machine to Transcendent Mind, Hans Moravec argues that advanced robots operating in the real world would need to deal with contingencies, and could do so through internal functions that parallel what real emotions do. These functions would take the form of “watchdog programs,” constantly operating within the robot to keep an eye out for trouble. If such a program senses danger ahead,

Similarly, Moravec postulates a doglike robotic love or loyalty, meaning that the unit could discern which of its activities especially please its human master, and would modify its behavior to keep its human happy. Anger, too, would play a role, for instance, in the behavior of a robot security guard. Upon detecting a human intruder, the being would move from requests for cooperation, to threats, to aggressive action. Carried to its logical though horrifying extreme, this scenario would resemble the scene in the film RoboCop where a police robot escalates its demands so aggressively that it shoots an innocent bystander before he can possibly comply. That response forces the corporate executives to decide they need a cyborg cop with human rather than robotic judgment.

There are obvious dangers in giving an artificial being emotions or their simulations, but surprisingly, it might also be that emotions are absolutely essential for a creature to think intelligently. As Rosalind Picard notes in her book Affective Computing:

Picard alludes to the fact that what seems to us an ingrained and strict distinction between rational and emotional thought—with the latter often dismissed as somehow less meaningful in human cognition—seems not to represent how the brain really works.

Varied evidence, from studies of people with damage to specific areas of the brain, to data from brain scans, shows intricate connections between the cortex, the part of the brain traditionally associated with rational thought, and the limbic system of the brain, parts of which are associated with the emotions. According to Antonio Damasio, whose researches and writings have been seminal in developing this view, “feelings are a powerful influence on reason … the brain systems required by the former are enmeshed in those needed by the latter.” High-order cognitive functions can be shaped and even

enhanced by emotions. Emotional states affect what the senses perceive; more surprisingly, too little emotion can prevent someone from narrowing down a menu of alternatives until a valid and rational choice emerges.

The links between reason and emotion might upset a treasured science-fiction notion, expressed in Isaac Asimov’s I, Robot and other stories, that an unemotional artificial brain can make better decisions for humanity than humanity, burdened by emotions, can make for itself. This perception might prove to be an illusion if machine emotions are necessary to produce fully capable machine intelligences. If that is the case, it adds a new dimension to the construction of an artificial brain: It might be necessary to ensure that the brain includes whatever combination of rational and emotional systems is necessary for full creative thought and consciousness.

Even without emotional components, however, it is a challenge to make digital microprocessors that process information as effectively as the human brain. New designs and modes of operation of computer chips might be required to improve sheer processing power, and eventually to incorporate new functions such as emotions. As Rosalind Picard has pointed out,

No one seems to be actually working on an “emotion chip” like the unit used by the android Commander Data in the Star Trek series, but researchers are working to improve artificial brains by increasing the speed and capacity of present-day digital chips, and by designing new types of chips. Further off on the horizon is the possibility of using living neurons or brains in cyborglike arrangements. To understand these possibilities, we need to examine both organic and artificial brains.

Look at the surface of a human brain and you see gray, distinct from the innermost parts that are white. The gray surface is the cortex, a

thin layer (6 millimeters or 0.25 inches thick) covering the brain (“cortex” comes from the Latin word for tree bark). It is the area where, according to traditional views of the brain, much of our thinking goes on; hence, the association of “gray matter” with intellectual activity. The cortex consists of layers of neurons, along with other cells that give them physical and physiological support. The neurons are arranged in orderly fashion, lying in vertical columns that are further grouped into functional clusters such as the visual and auditory cortices.

Much of the power of the brain comes from the sheer processing power of its 100 billion neurons. The cortex contains more of these than you might think because its wrinkles disguise its large surface area. Smoothed and spread out, the cortex would cover a square yard. Also significant is the number of interconnections among the neurons. Each neuron is connected to a thousand or more others, made possible partly by the fact that they link through three dimensions, vertically as well as horizontally.

Artificial brains—that is, microprocessor and memory chips that manipulate and store data—are also a form of gray matter, of a shinier sort, the color of pure silicon, polished to a reflective gunmetal sheen. Like neurons, the interconnected transistors etched into a chip transfer signals among themselves and come in huge numbers. Compared to a computer chip from the late 1970s, a modern Pentium processor has a thousand times more transistors and manipulates bits several thousand times faster, typically executing about 2,000 million instructions per second (MIPS). Today’s memory chips typically hold hundreds of megabytes of data.

But even the millions of transistors in a computer chip hardly compare to the billions of neurons in the brain. Further, the transistors are interconnected less abundantly than are neurons, partly because the chip is flat and not three-dimensional, and partly because it works serially, by executing one instruction after another. That scheme is less powerful than the massively parallel multiple processing carried out by an organic brain. In this respect, chips are inferior to natural brains, but they are also faster, transferring electronic signals in nanoseconds, a million times faster than organic systems. This enormous

edge in speed makes it possible to simulate in real time much of what the brain does, although the basic designs are different.

There is no question that artificial intelligence will grow as the processing power and storage capability of computer chips increase. But can chip-based intelligence reach the full power of the human brain? According to Hans Moravec, that goal would require microprocessors running at 100 million MIPS—50,000 times faster than a Pentium chip—supported by memory chips that store 100 million megabytes. Today’s most sophisticated computers, such as the unit that currently holds the world’s speed record—the Earth Simulator computer in Japan, which models our planet’s global behavior—are approaching this kind of speed. Even the Earth Simulator, however, is not remotely likely to form the brain of a mobile being: It cost about $500 million and occupies an entire building. The trick is to achieve the requisite speed in a tiny, low-power microprocessor chip. While that might seem a fantastic extension of what a Pentium chip can do, the pace of improvement has been staggering, with chip speeds roughly doubling every 18 months, so there is hope that the appropriate level can be achieved with our present silicon technology, perhaps in the 2020s as Moravec predicts.

But improvements in nanoscale silicon technology might run into roadblocks set by the behavior of matter at microscopic scales, which scientists do not fully understand, or by engineering problems, such as dealing with the heat thrown off by myriad transistors. If one of these roadblocks proves impassable, other approaches to nanoelectronics currently under study might still lead to more powerful artificial brains. A single molecule can now accomplish what a transistor does, suggesting the possibility of very compact processors and enormous memory banks. Scientists are also working on a new, exceedingly powerful type of computing based on quantum mechanics that might become a practical reality.

So far, these new approaches to computation are only in the experimental stage. Researchers are looking at other designs for building brains that bypass the limits of conventional chip technology, hoping to match the effectiveness of natural brain processes. One al-

ternate approach is to replace standard digital chips with neuromorphic chips—also made of silicon, but with a different mode of operation, one that mimics the workings of neurons and organic brains, which do not operate digitally. The transistors in a digital chip function as on-off switches that represent the binary digits 1 and 0, whereas the currents that flow among neurons do more than turn on and off. In neurons, the magnitude of the currents and how the currents change with time are essential information. Data transmitted in this way are said to be analog rather than digital in nature.

The differences between electronic digital processing and biological analog processing motivated two founders of the so-called neuromorphic approach: Carver Mead, Professor Emeritus at Caltech, who coined the term in the late 1980s, and Eric Vittoz of the Swiss Federal Institute of Technology. Compared to digital technology, writes Mead, “Biological information-processing systems operate on completely different principles,” and adds:

For Vittoz, an important advantage is that, “The collective computation carried out by the brain in its massively parallel architecture can be emulated on silicon.”

Kwabena Boahen, at the University of Pennsylvania, uses neuromorphic principles in the synthetic retina mentioned in the previous chapter, and emphasizes the efficiency of the method for the construction of an artificial brain. Transistors in a digital mode operate faster than neurons, but they also use more power. Boahen estimates that a digital chip as capable as the human brain would consume a billion watts of electrical power. If his estimate is anywhere near accurate, it is highly unlikely that we could ever build a humanlike digital brain in a form suitable for an artificial being. To explore the possibilities of using neuromorphic chips that copy how the brain operates, researchers in Boahen’s group have projects under way to develop a chip that follows the special pattern of connectivity among the six

layers of neurons in the visual cortex, and to implement the thousandfold connectivity among neurons that characterizes the brain.

Another new design with similarities to the brain is the adaptive or reconfigurable chip. In standard chip technology, once the transistors and other electronic elements are etched into the silicon, the circuitry they form is physically fixed, which also fixes what the chip can do. Normally the only way to change the circuit is to make a new chip. An adaptive chip, however, contains huge numbers of fundamental data-processing elements. Like myriad Lego blocks connected to form any one of an infinite number of structures, these elements can be connected in different configurations to perform different functions. Remarkably, this can be done on the fly with software commands rather than by the laborious process of changing hardware. Paul Master of QuickSilver Technology, a firm that along with Intel, IBM, and others is investigating adaptive chips, says: “Until now, the hardware had to match the problem. Now we can change that.”

According to proponents of the technology, adaptive chips offer speed and efficient use of power, in themselves important factors for artificial brains. But the truly novel aspect of the adaptive chip is its ability to change itself in real time, a crucial feature of the human brain not available in conventional chips: plasticity, the ability to change neural pathways according to new experience. Plasticity plays a large role in our development when we are young, and in learning and memory throughout our lives. It also gives the brain a way to compensate for functions that have been lost or reduced due to injury.

At this point in the development of adaptive chips, potential applications are limited to specialized devices such as cellphones, and there are questions about the long-term impact of the technology. But if it proves viable, then as is usually the case in nanoelectronics, improvement will come at a furious pace. We might expect to see second-generation silicon brains with the built-in ability to change and learn just as the human brain does—or perhaps solve for themselves the essential problems, such as finding the right mixture of rational and emotional thought.

There is, however, a different direction to consider in making

brains for artificial bodies, an ultimate possibility that avoids the difficult task of re-creating in silicon what nature has already worked out. This is the idea of merging living neurons or brains with artificial systems to create bionic or cyborglike arrangements. Whatever the advantages of neuronal processing, whatever the correct mix of reason and emotion, whatever the combination of brain structures and interactions that gives rise to consciousness, they already exist in living neural systems and brains.

The connection between organic brains and artificial constructions requires brain–machine interfaces (BMI), a technology still in its infancy. If true cyborgs, neural prosthesis, or the “jacked-in” world described by the science-fiction writer William Gibson ever come to pass, they will start with this research.

It is a serious step, scientifically and ethically, to go beyond emulating or copying the brain in silicon to merge living nervous systems with electronic networks, or to use living brains to operate nonliving devices. Some people feel a visceral shudder at the idea; it seems an unholy alliance, creating for them the same dread Victor Frankenstein felt as he animated a dead hulk. Grisly though the merging might seem, it is hard to argue with the main motivations for exploring this possibility; namely, the desire to aid paralyzed and handicapped people and also to develop new probes of the brain. Most of the scientists and physicians who work in this field (and the number has grown explosively in the last decade) seek to improve life for the ill and handicapped, but their research potentially also forms the basis for cyborg science.

One approach to making brain–machine interfaces focuses on the brain-wave method, where electrodes pressed against the scalp detect the brain’s neural activity, and the signals detected in this way are connected to an appropriate interface and computer. The information gained thereby can be used to train patients, who learn to tailor their brain activity to perform such useful functions as guiding a

computer cursor. The brain-wave method requires no brain surgery; however, it receives information from many neurons, a mixture that is difficult to interpret and put to optimum use. The pure signal from a single neuron is easier to analyze. For this reason, many BMI researchers want to implant electrodes in the brain to obtain well-defined signals from individual neurons. Several of these efforts are distinctly cyborglike, using a living brain to control an external device.

A case reported in 2000 might be the closest we have yet come to a natural brain devoted to an artificial body, like the dancing cyborg Deirdre in C.L. Moore’s story “No Woman Born”—except that this real-life cyborg did not use a human brain. Sandro Mussa-Ivaldi and his team at Northwestern University Medical School wanted to learn about connections between sensory information and motor responses in organic brains, so they created an animal cyborg. Its brain came from a sea lamprey, a predatory eel-like fish, chosen because its neurons are large and easy to manipulate. The portion used for the cyborg was the brainstem, which deals with vision and balance and ordinarily issues motor commands to control the animal’s swimming movements—but in the experiment, the fish’s body was replaced by a small two-wheeled mobile robot equipped with a set of light sensors. The brain, kept alive in a nutrient solution for experiments that ran as long as 8 hours, was connected to the light detectors and to the motors controlling the wheels by means of implanted electrodes.

To test the contrivance’s brain–body interaction, it was placed in a small arena ringed by light sources. As the lights were switched on, the brain received signals from its light sensors and responded by sending signals to the wheel motors, causing the robot to move. Not every installation of electrodes was successful, but when an installation worked well, the device moved consistently either toward or away from the lights, depending on how the electrodes were placed. In short, the artificial body was under the control of an organic brain that produced meaningful motor responses to sensory information, acting in a limited but truly cyborgian way.

Remarkably, cyborglike activity has also been achieved with the more powerful brains of monkeys and humans. In 2000, a group led

by neuroscientist Miguel Nicolelis at Duke University used the brain of a living monkey to control a robotic arm. The researchers implanted an array of fine wire electrodes in the brains of two owl monkeys. First they used the array to record neural patterns as they trained the monkeys to use their hands for specific tasks, such as reaching for food. These patterns gave enough information to predict where a monkey’s arm would be, milliseconds after the monkey’s brain produced the signal. The scientists used the data derived in this way to write a control program for robotic arms physically separate from the monkey. One arm, located in the same laboratory, was directly connected to the brain: a second one, distantly located in a laboratory at MIT, was connected via the Internet. For each arm, as the monkey moved its natural arm, the artificial limb faithfully followed its movements.

This astonishing result is not the final answer to the problem of giving a paralyzed person a useful BMI, which ideally should tap neural signals without any need for bodily motion. In 2003, Nicolelis and his group reported achieving that goal, although in monkeys, not humans. The scientists implanted electrodes in the brains of two macaque monkeys and recorded the neural patterns that arose as the animals learned to work a joystick that controlled a computer cursor; the monkeys were rewarded with a drink of juice when they used the cursor to reach and virtually grasp a target on the computer screen. Then the researchers disconnected the joystick but continued to tap the neural signals. Baffled at first, the monkeys soon realized that although manipulating the joystick no longer did anything, they could guide the cursor merely by thinking about it, with no muscular action. In a final step, the signals from the brain were rerouted to control a robot arm rather than the cursor. “By the end of training,” says Nicolelis, “I would say that these monkeys sensed they were reaching and grasping with their own arms instead of the robot arm.”

The ultimate goal is to develop similar brain–machine interfaces for humans. As an important step toward that goal, the Atlanta-based researcher and neurologist Philip Kennedy has, for the first time, enabled a human brain to control an external device with no associated bodily action. Kennedy implanted electrodes into the brain of a para-

lyzed man, who then learned to control a computer without moving a muscle.

Much of Kennedy’s work began at Emory University, and continues at his company Neural Signals. A visit to the company mostly reveals cluttered electronic workbenches where Kennedy and his colleagues work out how to process the signals elicited from implanted electrodes. But the heart of Kennedy’s achievement is the unique “neurotrophic” electrode that he developed, so small that it is not apparent on his workbench. Its design keeps it well anchored in the brain without undesirable motion or loss of electrical contact, problems that can plague conventional electrodes. The device consists of two fine gold wires attached to a tiny hollow glass cone the size of a pen tip. The inside of the cone is coated with nerve tissue, taken from another part of the patient’s body, which encourages the growth of nerve cells. When the cone is inserted into the brain, neurons grow through both its open ends to hold it firmly in place as they connect to the electrodes. The signals from the wires are sent first to an amplifier and then to a computer.

Kennedy’s great achievement has a second part, in clinical settings. He has implanted his electrode in patients who were severely paralyzed due to various causes including amyotrophic lateral sclerosis (ALS or Lou Gehrig’s disease), the condition that afflicts the British physicist Stephen Hawking. Among these patients was Johnny Ray, a Vietnam War veteran who was left “locked-in” at age 53 by a stroke; that is, he was left with full cognitive function but no bodily control except for limited face and neck motion. In 1998, Kennedy implanted two electrodes into Ray’s motor cortex (the part of the brain where movement commands originate as its neurons fire), specifically in the area devoted to moving the hands.

Ray sent signals through the electrodes to control a cursor he could watch on a computer screen, but achieving that control was not an overnight process. In initial training, Ray was asked to imagine he was moving his hands to activate the electrodes, but this strategy did not produce consistent results. After some months, however, he was able to move the cursor reliably, and these movements were correlated

with movements of his face and neck. Finally, Ray was able to accurately place the cursor wherever he wanted to with no bodily motion at all, which showed that he was controlling the cursor with his brain. When Ray reached this point, Kennedy asked him what he thought about as he moved the cursor. Ray’s reply (spelled out letter by letter with the cursor) was that he thought only about moving the cursor and not about any part of his body; so from Ray’s own viewpoint, he moved the cursor by mind alone.

One of the more interesting results of this feat relates to how the electrodes interact with the plasticity of the brain. Kennedy interprets his results as showing that as Ray was trained, the implanted “hand cortex” changed to “cursor cortex,” devoted only to cursor movement. Based on his work with monkeys, Nicolelis makes a similar point:

If human neurons actually adapt themselves in this way, that offers another potential benefit to patients; it gives great freedom as to where electrodes can be placed in the brain, regardless of where the brain is damaged. Changes in the brain due to neural interfacing might have other implications, too, that warrant careful consideration.

Although implanted electrodes enable a brain to communicate with the exterior world, efforts also continue to forge BMI connections that do not need surgery. One such BMI technology being pursued under DARPA auspices is based on magnetic effects. According to DARPA, the aim is to “communicate with the world directly through brain integration with peripheral devices and systems.” For example, pilots could control aircraft just by thinking about how they want them to move (as in the 1977 novel, Firefox, by Craig Thomas, and the 1982 film of the same name) or an infantryman could operate a powered exoskeleton. It has been known for some time that neural activity produces magnetic fields that extend through the skull and can be measured by sensitive detectors. Many questions remain,

though, such as whether the communication can work in both directions; that is, both from and to the brain, and whether the technique can be made fine-scaled enough to pick up signals from individual neurons. But the achievement of a noninvasive BMI would give a different meaning to bionic enhancements for civilians as well as soldiers; instead of submitting to a serious surgical procedure, a person could freely add or remove neural prosthetics, such as auxiliary memory or communication devices.

Other researchers are working on truly radical methods to interface neurons with electronic devices. One pioneer, Peter Fromherz of the Max Planck Institute for Biochemistry in Martinsried, Germany, aims to connect individual nerve cells with transistors by creating a “neuron on a chip.” He sees possibilities for neural prostheses and BMI units, and for new ways to study the brain, but there are difficulties. Although both chips and neural structures encode information electrically, the mechanisms differ. A chip carries electricity by means of a lightning-fast flow of electrons in a solid, whereas a neural system carries electricity by sluggish electrochemical means involving the movement of ions—charged atoms—in fluids. Hence the two systems operate at different time scales, nanoseconds for chips versus milliseconds for neurons. Finding ways to interface the systems is one challenge; ensuring that the silicon provides a proper base for nerve cells to grow and adhere is another.

Despite these challenges, Fromherz has created what might be called nanocyborgs, that is, prototypical neuron–chip hybrids, which consist of neurons from rat brains, leeches, or snails specially grown atop silicon chips so that the neurons overlay the transistors etched into the chips. Fromherz’s initial experiments show that the natural and the artificial systems indeed communicate directly; that is, electrical activity in the neuron causes activity in the chip, and vice versa. In a more complex demonstration, when two snail neurons were placed some distance apart on a silicon chip, the firing of one neuron induced the other to fire, although they were not linked by neuronal means but only through the silicon. Fromherz has also extended his technique to entire sections of a brain. In a structure consisting of a

thin slice of rat brain on a silicon chip containing a row of transistors, when the brain was electrically stimulated, so was the entire array of transistors.

These early, though encouraging, results are providing insight into the linkage between cell and silicon. Fromherz calls the two-neuron experiment “a silicon prosthesis,” suggesting that it might some day be possible to repair neuronal circuits electronically. Nevertheless, he urges caution, writing that “visionary dreams of bioelectronic neurocomputers and microelectronic neuroprostheses are unavoidable and exciting, but they should not obscure the numerous practical problems.”

A neuron on a chip is not conscious—although a network of them might be—but the other hybrid beings we have considered are. A locked-in paralyzed person like Johnny Ray brings a mind with subjective experience, a sense of self, and a personal history to whatever artificial extension is added through a BMI. It might never be desirable, ethical, or even possible to have a human brain operate a complete artificial body, rather than just a computer cursor or a mechanical arm. But if that were to happen, the result would be a being with the conscious personhood of the original mind (in Howard Gardner’s terms, intrapersonal intelligence), and the physical attributes of the new body—just as the cyborg dancer Deirdre kept her essential personality even as her brain functioned within a metal shell. (However, according to the findings that the brain changes as it operates new devices, that personality might change as well, as was feared would happen with Deirdre.)

Next, consider a hybrid being many steps below human mental capacity, a lamprey brain operating a robotic body. Does that cyborg possess consciousness? Scientists and philosophers would agree that even a living lamprey with brain and body intact lacks a sense of personal existence, what the consciousness theorist Gerald Edelman calls higher-order consciousness. But it does possess what Edelman

calls primary consciousness, the lower-level, nonlinguistic awareness of its immediate world through sensory information. The chunk of lamprey brain that operates the robotic body is even less of an individual being; nevertheless, in responding to sensory input and reacting by moving its artificial body, it shows primary awareness.

Now remove any living element and consider a completely artificial being. Does it possess consciousness? For the three smart robots we have examined, ASIMO, QRIO, and Kismet, there is no evidence for subjective experience, which is essential for higher consciousness and is implied in Gardner’s intrapersonal intelligence. But although the digital creatures are unaware of a self, they show low-level consciousness like that manifested by the lamprey cyborg. Through their senses, they know the world and respond to it.

They also have something else, internal body awareness through kinesthetic intelligence. As Hans Moravec points out, that kind of knowledge could easily be enhanced by designing a unit that registers, for instance, the power level of its own batteries and the operating temperature of its motors. Physical self-knowledge combined with knowledge of the world does not constitute consciousness of self, but if the unit uses information about its internal states to plan its actions, then it is displaying a salient characteristic of higher consciousness—namely, projection into the future—rather than responding only to the present moment. This state of consciousness is a more elevated one than that evinced by the lamprey cyborg.

There is another way that self-knowledge begins to touch on higher consciousness in artificial beings. According to theories of emotion, an event that elicits emotion triggers activity in the autonomic nervous system—the involuntary part of our nerve network that controls the glands, the heart, and more—to physiologically prepare the body to adapt to the stimulus. A person might experience nervousness by sensing the physiological reaction of “butterflies in the stomach,” along with a sense of fear. This might seem irrelevant to a robot like Kismet, which lacks any physiology. However, as Cynthia Breazeal points out, Kismet’s complex programming includes something roughly equivalent—a quantity that specifies its level of arousal,

depending on the stimulus it has been receiving. If Kismet itself reads this arousal tag, the robot not only is aroused, it knows it is aroused, and it can use this information to plan its future behavior. This, too, begins to resemble an important characteristic of higher consciousness; namely, the reflexive ability of a mind to examine itself over its own shoulder.

Unquestionably, these examples are a far cry from Gardner’s broad vision of intrapersonal intelligence, or that sense of selfhood projected forward and backward in time that defines the higher consciousness. Yet these tiny first steps toward self-knowledge might be the beginnings of an evolution toward full digital thought and consciousness—an evolution that has only begun, and whose prospects, along with other facets of the future and meaning of artificial beings, we examine in the last chapter.

This page intentionally left blank.