Artficial Intelligence

36 C O M M U N I C AT I O N S O F T H E A C M | J A N U A R Y 2 0 1 7 | V O L . 6 0 | N O . 1

V viewpoints

I M

A G

E B

Y O

L E

G D

O R

O S

H I

N /S

H U

T T

E R

S T

O C

K .C

O M

as natural-sounding voices, facial ex- pressions, and simulated displays of human emotions. Each of these tech- niques has valuable application to human-computer interfaces, but not when their primary effect is to fool or mislead. Attempts to dress up signifi- cant AI accomplishments with human- oid flourishes does the field a disser- vice by raising inappropriate questions and implying there is more there than meets the eye. Was IBM’s Watson pleased with its “Jeopardy!” win? It sure looked like it. This made for great

T H E D O M I N A N T P U B L I C narra- tive about artificial intelli- gence is that we are building increasingly intelligent ma- chines that will ultimately

surpass human capabilities, steal our jobs, possibly even escape human control and kill us all. This misguided perception, not widely shared by AI researchers, runs a significant risk of delaying or derailing practical applica- tions and influencing public policy in counterproductive ways. A more appro- priate framing—better supported by historical progress and current devel- opments—is that AI is simply a natural continuation of longstanding efforts to automate tasks, dating back at least to the start of the industrial revolution. Stripping the field of its gee-whiz apoc- alyptic gloss makes it easier to evaluate the likely benefits and pitfalls of this important technology, not to mention dampen the self-destructive cycles of hype and disappointment that have plagued the field since its inception.

At the core of this problem is the tendency for respected public figures outside the field, and even a few within the field, to tolerate or sanction over- blown press reports that herald each advance as startling and unexpected leaps toward general human-level in- telligence (or beyond), fanning fears that “the robots” are coming to take over the world. Headlines often tout noteworthy engineering accomplish- ments in a context suggesting they con- stitute unwelcome assaults on human

uniqueness and supremacy. If comput- ers can trade stocks and drive cars, will they soon outperform our best sales people, replace court judges, win Os- cars and Grammys, buy up and develop prime parcels of real estate for their own purposes? And what will “they” think of “us”?

The plain fact is there is no “they.” This is an anthropomorphic conceit borne of endless Hollywood blockbust- ers, reinforced by the gratuitous inclu- sion of human-like features in public AI technology demonstrations, such

Viewpoint Artificial Intelligence: Think Again Social and cultural conventions are an often-neglected aspect of intelligent-machine development.

DOI:10.1145/2950039 Jerry Kaplan

V viewpoints

J A N U A R Y 2 0 1 7 | V O L . 6 0 | N O . 1 | C O M M U N I C AT I O N S O F T H E A C M 37

viewpoints

V viewpoints

ket transitions can take considerable time, causing serious hardships for displaced workers. And if AI acceler- ates the pace of automation, as many predict, this rapid transition may cause significant social disruption.

But which jobs are most at risk? To answer this question, its useful to ob- serve that we don’t actually automate jobs, we automate tasks. So whether a worker will be replaced or made more productive depends on the nature of the tasks they perform. If their job in- volves repetitive or well-defined proce- dures and a clear-cut goal, then indeed their continued employment is at risk. But if it involves a variety of activities, solving novel challenges in chaotic or changing environments, or the authen- tic expression of human emotions, they are at far lower risk.

So what are the jobs of the future? While many people tend to think of jobs as transactional, there are plenty of professions that rely instead on building trust or rapport with other people. If your goal is to withdraw some spare cash for the weekend, an ATM is as effective as a teller. But if you want to secure an investor to help you build your new business, you won’t be pitching a machine anytime soon.

This is not to say that machines will never sense or express emotions; indeed, work on affective computing is proceeding rapidly. The question is how these capabilities will be per- ceived by users. If they are understood simply as aids to communication, they are likely to be broadly accepted. But if they are seen as attempts to fake sym- pathy or allay legitimate concerns, they are likely to foster mistrust and rejec- tion—as anyone can attest who has waited on hold listening to a recorded loop proclaim how important their call is. No one wants a robotic priest to take their confession, or a mechanical un- dertaker to console them on the loss of a loved one.

Then there are the jobs that involve demonstrations of skill or convey the comforting feeling that someone is paying attention to your needs. Except as a novelty, who wants to watch a self- driving racecar, or have a mechanical bartender ask about your day while it tops up your drink? Lots of profes- sions require these more social skills, and the demand for them is only going

television, but it also encouraged the audience to overinterpret the actual significance of this important achieve- ment. Machines don’t have minds, and there is precious little evidence to sug- gest they ever will.

The recent wave of public suc- cesses, remarkable as they are, arise from the application of a growing col- lection of tools and techniques that allow us to take better advantage of advances in computing power, stor- age, and the wide availability of large datasets. This is certainly great com- puter science, but it is not evidence of progress toward a superintelli- gence that can outperform humans at any task it may choose to undertake. While some of the new tools—most notably in the field of machine learn- ing—can be broadly applied to classes of tasks that may appear unrelated to the non-technical eye, in practice they often rely upon certain common attributes of the problem domains, such as enormous collections of ex- amples in digital form. High-speed trading algorithms, tracking objects in videos, and predicting the spread of infectious diseases all rely on tech- niques for finding subtle patterns in noisy streams of real-time data, and so many of the tools applied to these apparently diverse tasks are similar.

We are certainly using machines to perform all sorts of real-world tasks that people perform using their native intelligence, but this does not mean the computers are intelligent. It mere- ly means there are other ways to solve these problems. People and computers can play chess, but it is far from clear that they do it the same way. Recent advances in machine translation are re- markably successful, but they rely more on statistical correlations gleaned from large bodies of concorded texts than on fundamental advances in the under- standing of natural language.

Machines have always automated tasks that previously required human ef- fort and attention—both physical and mental—usually by employing very dif- ferent techniques. And they often do these tasks better than people can, at lower cost, or both—otherwise they would not be useful. Factory automation has replaced myriad highly skilled and highly trained workers, from sheet met- al workers to coffee tasters. Arithmetic

problems that used to be the exclusive domain of human “calculators” are now performed by tools so inexpensive they are given away as promotional trinkets at trade shows. It used to take an army of artists to animate Cinderella’s hair, but now CGI techniques render Rapunzel’s flowing locks. These advances do not de- mean or challenge human capabilities; instead they liberate us to perform ever more ambitious tasks.

Some pundits warn that computers in general, and AI in particular, will lead to widespread unemployment. What will we do for a living when ma- chines can perform nearly all of to- day’s jobs? A historical perspective re- veals a potential flaw in this concern. The labor market constantly evolves in response to automation. Two hun- dred years ago, more than 90% of the U.S. labor force worked on farms. Now, barely 2% produce far more food at a fraction of the cost. Yet, everyone isn’t out of work. In fact, more people are employed today than ever before, and most would agree their jobs are far less taxing and more rewarding than the backbreaking toil of their ancestors. This is because the benefits of automa- tion make society wealthier, which in turn generates demand for all sorts of new products and services, ultimately expanding the need for workers. Our technology continually obsoletes pro- fessions, but our economy eventually replaces them with new and different ones. It is certainly true that recent advances in AI are likely to enable the automation of many or most of today’s jobs, but there is no reason to believe the historical pattern of job creation will cease.

That’s the good news. The bad news is that technology-driven labor mar-

Machines don’t have minds, and there is precious little evidence to suggest they ever will.

38 C O M M U N I C AT I O N S O F T H E A C M | J A N U A R Y 2 0 1 7 | V O L . 6 0 | N O . 1

viewpoints

sign problems, not the unpredictable consequences of tinkering with some presumed natural universal order. Good products, including increasingly autonomous machines and applica- tions, don’t go haywire unless we de- sign them poorly. If the HAL 9000 kills its crewmates to avoid being deactivat- ed, it is because its designers failed to prioritize its goals properly.

To address these challenges, we need to develop engineering stan- dards for increasingly autonomous systems, perhaps by borrowing con- cepts from other potentially hazard- ous fields such as civil engineering. For instance, such systems could in- corporate a model of their intended theater of operation, (known as a Standard Operating Environment, or SOE), and enter a well-defined “safe mode” when they drift out of bounds. We need to study how people naturally moderate their own goal-seeking be- havior to accommodate the interests and rights of others. Systems should pass certification exams before de- ployment, the behavioral equivalent of automotive crash tests. Finally, we need a programmatic notion of basic ethics to guide actions in unantici- pated circumstances. This is not to say machines have to be moral, simply that they have to behave morally in rel- evant situations. How do we prioritize human life, animal life, private prop- erty, self-preservation? When is it ac- ceptable to break the law?

None of this matters when comput- ers operate in limited, well-defined domains, but if we want AI systems to be broadly trusted and utilized, we should undertake a careful reassess- ment of the purpose, goals, and po- tential of the field, as least as it is per- ceived by the general public. The plain fact is that AI has a public relations problem that may work against its own interests. We need to tamp down the hyperbolic rhetoric favored by the pop- ular press, avoid fanning the flames of public hysteria, and focus on the chal- lenge of building civilized machines for a human world.

Jerry Kaplan ([email protected]) is a visiting lecturer in computer science at Stanford University and a fellow at the Stanford Center for Legal Informatics. His latest book is Artificial Intelligence: What Everyone Needs to Know (Oxford University Press, 2016).

Copyright held by author.

to grow as our disposable income in- creases. There’s no reason in principle we can’t become a society of well-paid professional artisans, designers, per- sonal shoppers, performers, caregiv- ers, online gamers, concierges, cura- tors, and advisors of every sort. And just as many of today’s jobs did not ex- ist even a few decades ago, it is likely a new crop of professions will arise that we can’t quite envision today.

So the robots are certainly com- ing, but not quite in the way most people think. Concerns that they are going to obsolete us, rise up, and take over, are misguided at best. Worrying about superintelligent machines dis- tracts us from the very real obstacles we will face as increasingly capable machines become more intricately intertwined with our lives and be- gin to share our physical and public spaces. The difficult challenge is to ensure these machines respect our often-unstated social conventions. Should a robot be permitted to stand in line for you, put money in your parking meter to extend your time, use a crowded sidewalk to make de- liveries, commit you to a purchase, enter into a contract, vote on your behalf, or take up a seat on a bus? Philosophers focus on the more ob- vious and serious ethical concerns— such as whether your autonomous vehicle should risk your life to save two pedestrians—but the practical questions are much broader. Most AI researchers naturally focus on solv- ing some immediate problem, but in the coming decades a significant im- pediment to widespread acceptance of their work will likely be how well their systems abide by our social and cultural customs.

Science fiction is rife with stories of robots run amok, but seen from an engineering perspective, these are de-

Model Learning

Smart Machines Are Not A Threat to Humanity

AI Dangers: Imagined and Real

Computing History Outside the U.K. And U.S.: Some Selected Landmarks from Continental Europe

Copyright Enforcement in the Digital Age

A Messy State of the Union: Taming the Composite State of Machines of TLS

Life Beyond Distributed Transactions

BBR: Congestion-Based Congestion Control

Are You Load Balancing Wrong?

Plus the latest news about closing the back door, predictive policing, and secure quantum communications.

C om

in g N

ex t

M on

th in

C O

M M

U N

IC A

T IO

N S So the robots

are certainly coming, but not in the way most people think.