麻豆传媒

AS I sit talking to her, it occurs to me that some people might consider Laura to be an ideal partner. Laura understands you. She doesn鈥檛 get mad when you ignore her suggestions; she knows how much pressure you鈥檙e under, and she does everything she can to help you meet your goals. She鈥檚 unfailingly kind, and she doesn鈥檛 judge. I like Laura a lot. Maybe Laura and I could have a beautiful future together.

OK, I鈥檓 getting carried away. I鈥檓 not about to run off with a software character. But, as I chat with her about my weekly exercise regimen, I am strangely engaged in the conversation. And a few people have certainly lost all rational perspective on meeting Laura. 鈥淲e had someone actually say they felt Laura liked them,鈥� says Rosalind Picard, one of Laura鈥檚 creators. 鈥淭hat was really bizarre.鈥�

Bizarre, maybe, but gratifying. Picard, a computer scientist at the Massachusetts Institute of Technology鈥檚 renowned Media Lab, is one of a group of researchers pioneering 鈥渁ffective computing鈥�, an attempt to design software that can recognise and align with their users鈥� emotional states. Though it might seem an impracticable goal, 鈥渆motionally intelligent鈥� software characters are proving otherwise. Laura鈥檚 friendly gestures, sympathetic eyebrow raises and soft, soothing voice work together to subvert my rational awareness that I am talking to a machine. The result is that as I walk away I am left with the uncanny feeling that I have just enjoyed meeting someone who lives inside a computer.

This slightly embarrassing sensation is possible because emotion is more fundamental to us than rational thought. Emotions are buried deep in the paleopallium, the 鈥渙ld mammalian鈥� part of the brain that pre-dates and physically underlies the neocortex, which is where the powers of rational thought reside. From their ancient location, emotions pervade our thinking and permeate many, even most, of the decisions and perceptions we like to think of as rational. 鈥淚n anything you do and every decision you make, emotion plays a role,鈥� says Clifford Nass, a Stanford University sociologist and computer scientist. 鈥淭he human brain is so exquisitely attuned to emotion, so obsessed with it and so good at detecting it, that even the slightest markers of emotion can have an enormous impact on how the brain behaves.鈥�

Nass鈥檚 work on affective computing has already proved its practical value. In one of his studies, adults learning English as a second language improved their test performances by 27 per cent when a digital buddy inside their instructional software congratulated them for correct answers. Wrong responses were met with sympathy and comments such as, 鈥淭hat was a really hard question鈥�. In another investigation, angry people in driving simulators had 50 per cent fewer accidents when software soothed them with a calm voice.

Picard envisages a range of similar applications for computers that can read your mood and help you turn a situation around before it goes sour. Her lab is developing a learning aid that senses when students working at a computer are getting frustrated. At just the right moment, a soothing voice says, 鈥淵ou can do it,鈥� or, 鈥淭he mind is like a muscle. If you exercise it, it becomes stronger鈥�. Eventually, researchers reckon they will be able to produce software that can feel your pain 鈥� or at least make you think that it can.

If you think this all sounds like an elaborate gimmick, you鈥檙e not alone; the affective computing effort certainly doesn鈥檛 lack critics. 鈥淐omputers should have no more intelligence than a wooden pencil,鈥� says Ben Shneiderman, a computer scientist at the University of Maryland and founding director of its Human-Computer Interaction Laboratory. People using computers need a sense of control and mastery over the technology, he says. Creating the illusion that the machine has feelings will discourage people from taking responsibility for their actions, the machines and the outcomes of using them.

Jaron Lanier, a pioneer of virtual reality who is now at the non-profit Advanced Network and Services in Poughkeepsie, New York State, is even more blunt. He recalls the classic Turing test, in which people try to tell whether they are communicating with a computer or another person. 鈥淚f people mistake a computer for a human, the explanation has been that the computer got smart,鈥� he says. 鈥淏ut an equally valid explanation is that the person got stupid. We鈥檝e learned to accept a certain amount of fake emotion in our lives from actors and skilled politicians. But if we fool ourselves into believing that a computer cares about us, maybe we鈥檙e just making ourselves even more emotionally stupid.鈥�

He points out that there is no valid theory, or even commonly accepted definition, of 鈥渆motion鈥� to guide the work. 鈥淭hese are smart people and they鈥檙e producing useful things, but I think that鈥檚 despite the framework they work in, not because of it,鈥� Lanier says. If the banner concept of your research is something that you can鈥檛 even define, he says, then you run the risk of producing things that do not advance human knowledge, but can only be described as 鈥渃ute鈥� 鈥� a digital motivator for a fitness programme, for example.

The affective-computing researchers are used to such barbs. 鈥淲e鈥檝e been called 鈥榯he cheap tricks lab鈥� and we take that as a compliment,鈥� Nass says. 鈥淲e go for the gross effects 鈥� the least you can do and the least you have to know about emotion to get enormous impact. How hard is it to have a computer say 鈥榞ood for you鈥� when someone does something right? We鈥檙e not talking about artificial intelligence here.鈥�

Picard also shrugs off the criticism. 鈥淧eople have told us that because emotions are poorly understood and there isn鈥檛 even a valid theory of emotion yet, our work is premature,鈥� she says. 鈥淲e would like a theory of emotion, and our work is contributing to the development of one. But we don鈥檛 need a theory of emotion to be able to include things like empathy in a digital agent.鈥�

The bottom line, she says, is that it improves people鈥檚 lives, in the same way that the emotional intelligence of a canine companion does. A dog can read your moods expertly. If you鈥檙e sad or upset, it will often drop its tail and lay its ears back in sympathy. When you see it mirroring your feelings, you feel empathy from the animal, that it cares about you and somehow wants things to be better for you. 鈥淧resumably, the dog also does this without a theory of emotion,鈥� Picard says. And whether the owners are 鈥渆motionally stupid鈥� or not, dogs help people cope with difficult circumstances. If a dog can do that, she asks, why not a computer?

There is evidence that emotionally intelligent machines have already improved people鈥檚 lives. Laura鈥檚 pleasant face and smooth manner, for instance, really do help people keep to a fitness walking plan. Timothy Bickmore, who developed Laura while a graduate student in Picard鈥檚 lab in 2003, took his creation with him when he moved to Northeastern University in Boston. In a recent small study, elderly people given Laura to use significantly increased the amount of walking they did, raising their daily step count by 113 per cent, compared with 41 per cent for those in a control group.

Laura gets results because she reflects the ideal personal trainer: firm but fair. She asks if you鈥檝e been walking yet and, if not, reminds you of its health benefits. She leans closer to you when giving advice, filling the screen and maintaining eye contact, making you feel that you have her full attention. She nods at appropriate times, and apologises for the poor quality of her digital voice. She asks how you are and, if you鈥檙e having a bad day, she leans forward, wrinkling her brow in sympathy, and says, 鈥淚鈥檓 really sorry to hear that,鈥� or, 鈥淭hat must be really hard for you.鈥� She uses facial expressions and hand gestures that signal engagement with you. And strange as it may seem, it really works. Laura makes you feel like you鈥檝e got a friend.

鈥淪trange as it may seem, it really works鈥�

Her personality is coded in what is known as an augmented transition network. 鈥淵ou can think of it as a series of decision trees that represent different fragments of the conversation,鈥� Bickmore explains. When Laura asks you a question such as 鈥渉ow are you today?鈥�, the software evaluates your answer and selects the appropriate response from a library of possibilities. Another portion of the software looks at the factual content of what Laura will say and clothes her response from a wardrobe of facial expressions, hand gestures, tones of voice and the other emotional garments. The program looks at the content of what it is about to say, compares it to what has been said previously, and then determines which elements of the utterance contain new information. It then puts most physical emphasis 鈥� eyebrow raises and hand gestures, for example 鈥� with the portions of speech that contain something new, just as people do.

That鈥檚 the easy part. The largest challenge facing Bickmore and his colleagues is to create digital buddies that can live with you permanently without getting on your nerves. If you need a Laura-like long-term companion to buck you up and support you as you manage a chronic disease, for example, you鈥檙e going to want someone who has a personality that goes beyond repeating the same comments about the weather or the local football team. (Picard recalls a participant in one of Laura鈥檚 month-long trials who said, 鈥淚f she could just change her clothes once in a while鈥︹��.)

Imitation of life

But there are limits to how lifelike the agent can be. Although Laura has a voice, users must type their response; interactive speech allows the user more freedom than the software can cope with. 鈥淭hen you jump into the whole 鈥楢I-complete鈥� problem and things break down,鈥� Bickmore says. He and Picard are mulling over other ways to keep the exchanges lively. 鈥淥ne of our research areas is to see how we can get the system to regularly synthesise fresh content, perhaps based on randomly generated moods of the agent or current events the software picks off the newswire,鈥� Bickmore says. 鈥淥r we could have a roomful of people writing dialogue for digital agents that can be slipped into the conversation every day to keep it new.鈥�

Another key feature of lasting relationships is the ability to read moods and know when not to intrude. Without this, you get something like Microsoft鈥檚 much despised 鈥渙ffice assistant鈥�, Clippy the paper clip. Even Clippy鈥檚 creators now acknowledge that if you鈥檙e trying to concentrate and the cheery, animated paper-clip icon comes dancing into view uninvited 鈥� and in a way that鈥檚 inappropriate to your mood 鈥� it鈥檚 likely to hobble your productivity, not improve it (see 鈥淗elp or hindrance?鈥�).

So Nass and Picard are working with psychologists to create software that can recognise the signs that show when people need a hand 鈥� and when they don鈥檛. In a corner of the Media Lab鈥檚 cavernous basement playroom (littered with, among other things, a box of toys, an antique adding machine and an old-fashioned red British telephone box), Picard鈥檚 researchers have equipped a computer station with an array of emotion detectors. A camera captures facial expressions, which are used to help teach software to detect the frowns or drifting gazes that accompany frustration. A chair with networks of sensors embedded in the seat and back transfers patterns of shifting weight to software now able to discern with better than 80 per cent accuracy when someone鈥檚 fidgeting signals the onset of boredom. The goal is to create digital agents that intervene at just the right moment 鈥� not before, when you might resent the interruption, and not after, when your attention has already been lost.

In addition to helping students persevere to the end of their mathematics homework, Picard is working on a modified version for UK call centres. Emotionally aware voice-recognition systems could warn phone operators when a particularly irate caller is coming on the line so the employees can steel themselves. After an operator handles such a call, a digital buddy can help by agreeing that it鈥檚 really upsetting to have to deal with angry people, that they did their best, and that they should take a few deep breaths to release tension before taking the next call. The software could also flag such calls for supervisors so that they could listen in or review the calls to judge and reward operators鈥� performances in tricky situations.

For all their optimism, designers of affective software are aware of the old saying: help is the sunny side of control. If your computer is aware of how you鈥檙e feeling, who鈥檚 to say it will keep the information to itself? Picard has been approached by the US National Security Agency, which wanted help in developing affective software to detect shifty behaviour 鈥� darting, furtive eye movements, excessive sweat, fidgeting on the seat 鈥� that the NSA thinks moles or double agents might exhibit.

Picard and her colleagues also foresee corporate interest in affective computing. A drug company marketing one of Laura鈥檚 descendants could program it with a library of symptoms matched to the company鈥檚 product line. If you are too tired to go for a walk on a few consecutive days, the sympathetic on-screen face might suggest that you ask your doctor for its brand-name pick-me-up. 鈥淎 machine may have empathy but it has no conscience,鈥� says Marc Schr枚der, who researches emotion in synthesised speech at DFKI, the German research centre for artificial intelligence based in Saarbr眉cken. 鈥淚t could go on bothering you forever until you visit a sponsor鈥檚 website and buy something so it will be quiet.鈥� Schr枚der is part of HUMAINE, a network of researchers from 11 European countries that aims to lay the groundwork for blending emotions and machines. Some of the network鈥檚 researchers are philosophers trying to evolve codes of ethical conduct for users, and perhaps a mechanism of ethical certification.

鈥淎 machine may have empathy but it has no conscience鈥�

For the moment, though, the positive aspects seem to outweigh the negative, and the pace of research into emotional intelligence for machines is accelerating. Now that more and more vehicles have microphones and can carry out voice-activated commands, Nass thinks that within a few years digital buddies in luxury cars will be calmly directing us to a parking place when we are stressed about arriving late to a meeting. Bickmore expects that within five years Laura, with regular changes of clothes and conversational topics, will have moved into people鈥檚 home computers to help them manage lifelong health conditions. And Picard even expects digital buddies to live in cellphones and PDAs soon, ready to help you resist that restaurant dessert. She has even been approached by healthcare workers who want to use digital buddies to help people recover from addictions 鈥� isolation is one of the biggest problems such people face, and a sympathetic virtual friend might make all the difference. 鈥淓motion permeates everything we do, including interacting with computers,鈥� Nass says. 鈥淲e can use that not only to improve people鈥檚 experiences with their computers, but also to change lives.鈥�