artificial intelligence: Past and future - Computer Science

newslion, and Rutter, whose total winningsamounted to $3.25 million, the mostmoney ever won by a single “Jeopardy!”player. At the end of the three-dayevent, Watson finished with $77,147,beating Jennings, who had $24,000,and Rutter, who had $21,600. Themillion-dollar prize money awarded toWatson went to charity.Named after IBM founder ThomasJ. Watson, the Watson system was builtby a team of IBM scientists whose goalwas to create a standalone platformthat could rival a human’s ability toanswer questions posed in naturallanguage. During the “Jeopardy!” challenge,Watson was not connected to theInternet or any external data sources.Instead, Watson operated as an independentsystem contained in severallarge floor units housing 90 IBM Power750 servers with a total of 2,880 processingcores and 15 terabytes of memory.Watson’s technology, developed byIBM and several contributing universities,was guided by principles describedin the Open Advancement of Question-Answering (OAQA) framework, which isstill operating today and facilitating ongoinginput from outside institutions.Judging by the sizeable coverage ofthe event, Watson piqued the interestof technology enthusiasts and the generalpublic alike, earning “Jeopardy!”the highest viewer numbers it hadachieved in several years and leadingto analysts and other industry observersspeculating about whether Watsonrepresents a fundamental new ideain computer science or merely a solidfeat of engineering. Richard Doherty,the research director at EnvisioneeringGroup, a technology consulting firmbased in Seaford, NY, was quoted in anAssociated Press story as saying thatWatson is “the most significant breakthroughof this century.”Doherty was not alone in makingsuch claims, although the researcherson the IBM team responsible fordesigning Watson have been far moremodest in their assessment of thetechnology they created. “Watson is anovel approach and a powerful architecture,”says David Ferrucci, directorof the IBM DeepQA research team thatcreated Watson. Ferrucci does characterizeWatson as a breakthrough in artificialintelligence, but he is careful toqualify this assertion by saying that thebreakthrough is in the development ofartificial-intelligence systems.“The breakthrough is how we pulledeverything together, how we integratednatural language processing, informationretrieval, knowledge representation,machine learning, and a generalreasoning paradigm,” says Ferrucci. “Ithink this represents a breakthrough.We would have failed had we not investedin a rigorous scientific methodand systems engineering. Both wereneeded to succeed.”Contextual EvidenceThe DeepQA team was inspired byseveral overarching design principles,with the core idea being that no singlealgorithm or formula would accuratelyunderstand or answer all questions,Watson’s on-stage persona simulates the system’s processing activity and relative answerconfidence through moving lines and colors. Watson is shown here in a practice match withKen Jennings, left, and Brad Rutter at IBM’s Watson Research Center in January.says Ferrucci. Rather, the idea was tobuild Watson’s intelligence from abroad collection of algorithms thatwould probabilistically and imperfectlyinterpret language and scoreevidence from different perspectives.Watson’s candidate answers, those answersin which Watson has the mostconfidence, are produced from hundredsof parallel hypotheses collectedand scored from contextual evidence.Ferrucci says this approach requiredinnovation at the systemslevel so individual algorithms couldbe developed independently, thenevaluated for their contribution to thesystem’s overall performance. The approachallowed for loosely coupled interactionbetween algorithm components,which Ferrucci says ultimatelyreduced the need for team-wide agreement.“If every algorithm developerhad to agree with every other or reachsome sort of consensus, progresswould have been slowed,” he says.“The key was to let different membersof the team develop diverse algorithmsindependently, but regularlyperform rigorous integration testingto evaluate relative impact in the contextof the whole system.”Ferrucci and the DeepQA team areexpected to release more details laterthis year in a series of papers that willoutline how they dealt with specific aspectsof the Watson design. For now,only bits and pieces of the completepicture are being disclosed. Ferruccisays that, looking ahead, his team’s researchagenda is to focus on how Watsoncan understand, learn, and interactmore effectively. “Natural language understandingremains a tremendouslydifficult challenge, and while Watsondemonstrated a powerful approach,we have only scratched the surface,” hesays. “The challenge continues to beabout how you build systems to accuratelyconnect language to some representation,so the system can automaticallylearn from text and then reason todiscover evidence and answers.”Lillian Lee, a professor in the computerscience department at CornellUniversity, says the reactions aboutWatson’s victory echo the reactions followingDeep Blue’s 1997 victory overchess champion Garry Kasparov, butwith several important differences.Lee, whose research focus is naturalphoto courtesy ibm14 communications of the acm | july 2011 | vol. 54 | no. 7

newslanguage processing, points out thatsome observers were dismissive aboutDeep Blue’s victory, suggesting thatthe system’s capability was due largelyto brute-force reasoning rather thanmachine learning. The same criticism,she says, cannot be leveled at Watsonbecause the overall system needed todetermine how to assess and integratediverse responses.“Watson incorporates machinelearning in several crucial stages of itsprocessing pipeline,” Lee says. “Forexample, reinforcement learning wasused to enable Watson to engage instrategic game play, and the key problemof determining how confident tobe in an answer was approached usingmachine-learning techniques, too.”Lee says that while there has beensubstantial research on the particularproblems the “Jeopardy!” challengeinvolved for Watson, that prior workshould not diminish the team’s accomplishmentin advancing the stateof the art to Watson’s championshipperformance. “The contest reallyshowcased real-time, broad-domainquestion-answering, and provided ascomparison points two extremely formidablecontestants,” she says. “Watsonrepresents an absolutely extraordinaryachievement.”Lee suggests that with languageprocessingtechnologies now maturing,with the most recent example ofsuch maturation being Watson, thefield appears to have passed throughan important early stage. It now facesan unprecedented opportunity in helpingsift through the massive amountsof user-generated content online, suchas opinion-oriented information inproduct reviews or political analysis,according to Lee.While natural-language processingis already used, with varying degreesof success, in search engines andother applications, it might be sometime before Watson’s unique question-answeringcapabilities will helpsift through online reviews and otheruser-generated content. Even so, thatday might not be too far off, as IBMhas already begun work with NuanceCommunications to commercializethe technology for medical applications.The idea is for Watson to assistphysicians and nurses in finding informationburied in medical tomes, prior“Natural languageunderstandingremains atremendouslydifficult challenge,and while Watsondemonstrateda powerful approach,we have onlyscratchedthe surface,”says David Ferrucci.cases, and the latest science journals.The first commercial offerings fromthe collaboration are expected to beavailable within two years.Beyond medicine, likely applicationareas for Watson’s technology wouldbe in law, education, or the financialindustry. Of course, as with any technology,glitches and inconsistencieswill have to be worked out for each newdomain. Glitches notwithstanding,technology analysts say that Watsonliketechnologies will have a significantimpact on computing in particular andhuman life in general. Ferrucci, for hispart, says these new technologies likelywill mean a demand for higher-densityhardware and for tools to help developersunderstand and debug machinelearningsystems more effectively.Ferrucci also says it’s likely that userexpectations will be raised, leading tosystems that do a better job at interactingin natural language and siftingthrough unstructured content.To this end, explains Ferrucci, theDeepQA team is moving away from attemptingto squeeze ever-diminishingperformance improvements out ofWatson in terms of parsers and localcomponents. Instead, they are focusingon how to use context and informationto evaluate competing interpretationsmore effectively. “What we learned isthat, for this approach to extend beyondone domain, you need to implement apositive feedback loop of extracting basicsyntax and local semantics from language,learning from context, and theninteracting with users and a broadercommunity to acquire knowledge thatis otherwise difficult to extract,” hesays. “The system must be able to bootstrapand learn from its own failingwith the help of this loop.”In an ideal future, says Ferrucci, Watsonwill operate much like the ship computeron “Star Trek,” where the inputcan be expressed in human terms andthe output is accurate and understandable.Of course, the “Star Trek” ship computerwas largely humorless and devoidof personality, responding to queriesand commands with a consistently eventone. If the “Jeopardy!” challenge servesas a small glimpse of things to come forWatson—in particular, Watson’s precisewagers, which produced laughterin the audience, and Watson’s visualizationcomponent, which appeared to expressthe state of a contemplative mindthrough moving lines and colors—theDeepQA team’s focus on active learningmight also include a personality loop soWatson can accommodate subtle emotionalcues and engage in dialogue withthe kind of good humor reminiscent ofthe most personable artificial intelligencesin fiction.Further ReadingBaker, S.Final Jeopardy: Man vs. Machine and theQuest to Know Everything. Houghton MifflinHarcourt, New York, NY, 2011.Ferrucci, D., Brown, E., Chu-Carroll, J., Fan, J.,Gondek, D., Kalyanpur, A.A., Lally, A., Murdock,J.W., Nyberg, E., Prager, J., Schlaefer, N.,and Welty, C.Building Watson: An overview of theDeepQA project, AI Magazine 59, Fall 2010.Ferrucci, D., et al.Towards the Open Advancement of QuestionAnswering Systems. IBM Research ReportRC24789 (W0904-093), April 2009.Simmons, R.F.Natural language question-answeringsystems, Communications of the ACM 13, 1,Jan. 1970.Strzalkowski, T., and Harabagiu, S. (Eds.)Advances in Open Domain QuestionAnswering. Springer-Verlag, Secaucus, NJ,2006.Based in Los Angeles, Kirk L. Kroeker is a freelanceeditor and writer specializing in science and technology.© 2011 ACM 0001-0782/11/07 $10.00july 2011 | vol. 54 | no. 7 | communications of the acm 15

NnewsScience | doi:10.1145/2063176.2063182Better Medicine ThroughMachine LearningComputers that tease out patterns from clinical datacould improve patient diagnosis and care.Neil SavageImage courtesy of Kenji SuzukiMedicine can be as muchart as science, a detectivestory in which doctors relynot only on lab tests andx-rays, but on their ownexperience and clues from a patient’shistory to develop diagnoses or predictfuture health problems. But all of thoselab tests, blood pressure readings, magneticresonance imaging (MRI) scans,electrocardiograms, and billing codesadd up to reams of data, which beforetoo long will be joined by individualgene sequencing. Computer scientistsare increasingly applying machinelearning techniques to all that data,searching for patterns that can aid diagnosisand improve clinical care.“Machine learning plays, I think, anessential role in medical image analysisnowadays,” says Kenji Suzuki, assistantprofessor of radiology and medicalphysics at the University of Chicago’sComprehensive Cancer Center. Suzukihas been working on automating thedetection of cancerous lesions in imagesfrom x-rays or computed tomographyscans. Considering that radiologistsmay miss 12%–30% of lung cancers insuch scans, a machine learning tool offersgreat potential.Since the mid-1980s, computer sci-(a) (b) (c) (d)Kenji Suzuki and colleagues’ comparison of their rib-suppressed temporal-subtraction(TS) images with conventional TS images: (a) previous chest radiographs, (b) current chestradiographs of the same patient, (c) rib-suppressed TS images with fewer rib artifacts, and(d) conventional TS images.january 2012 | vol. 55 | no. 1 | communications of the acm 17

newsentists have tried to improve on thatperformance using feature-based machinelearning in which the computerwould pick out morphological features,texture differences, and more to identifyabnormal tissue. But such catalogingof features still misses some cancersthat doctors are able to spot with theirown eyes. Sometimes the features thecomputer is seeking can be subtle oroverlap with normal anatomical structuressuch as bone, making them moredifficult to spot. So Suzuki asks the machineto instead focus on the values,such as intensity, of individual pixels.“Because the computing power has increaseddramatically in recent years, wecan process the pixel values directly,” hesays. The resulting system is highly sensitive,achieving up to 97% accuracy.But one concern is making the programso sensitive that it starts findingnonexistent lesions. In Suzuki’s lungcancer tests, the feature-based algorithmfalsely identified five lesions perpatient, while the pixel-based methodproduced less than one false positiveper patient and no false negatives.Suzuki says each method produces adifferent type of false positive, so combiningthe two approaches leads to themost accurate outcome. Suzuki is nowworking on expanding the technique toother types of cancer and other imagingmethods, such as MRI and PET. “Youjust need to train the machine learningAs the useof electronicmedical recordsgains acceptance,machine learningis likely to playan even larger rolein clinical medicine.technique with new images,” he says.A Guide for DiagnosisAt IBM Almaden Research Center, theAdvanced Analytics for InformationManagement (AALIM) project appliesmachine learning to a wide variety ofdata—readings of vital signs, tests suchas echocardiograms, and demographicinformation—to chart the medical historiesof patients over several years. Bycomparing the history of hundreds orthousands of people, the system canidentify previous patients who are similarto a current patient, then apply collaborativefiltering to suggest the bestdiagnosis and treatment options for thenew patient.“With a large number of pre-diagnosedpatient datasets available inelectronic health records, physicianscan now benefit from the opinion oftheir peers on cases similar to their patients,”says Tanveer Syeda-Mahmood,head of IBM’s Multimodal Mining forHealthcare project. The hope is that byhelping doctors base their decisions onquantitative information, the numberof diagnostic errors can be reduced. AA-LIM, Syeda-Mahmood says, provides “aholistic view of the patient’s condition,”producing one-page summaries, longtermprofiles of various measurementsof health, and detailed comparisonsshowing diagnosis, treatment, and outcomesfor similar patients.The computer might, for instance,help a relatively new doctor decide thatshe needs to consult a specialist. Approximately5% of cases, says Syeda-Mahmood, are ambiguous enough thateven senior clinicians ask other doctorsfor their opinions, which AALIM easilyprovides. In emergency rooms, it can cutthe time doctors spend flipping throughcharts by half. Although researchershave tested the system on patient data,it would likely require approval from theFood and Drug Administration before itcould be used in a hospital.Finding ways to use all this medicaldata often requires new developmentsin machine learning. For instance,Syeda-Mahmood wanted to give theEducationTech-Rich Learning EnvironmentsWhat’s a solution to classroomsplagued by students’ lowretention rates, high withdrawals,and failing grades? Toss outtraditional lectures and create atechnology-rich environment, sayresearchers at Rochester Instituteof Technology (RIT).Data from six years ofresearch involving 500-plusundergraduates taking threedifferent engineering coursesrevealed that 90% of the studentssaid they learned and retainedinformation better when acombination of tablet PCs,collaborative software, andmultiple projection screens wereused as teaching tools.The study targeted threefoundation courses forengineering-technology degreeprograms—pneumatics andhydraulics, applied dynamics,and applied fluid mechanics—all of which suffered fromhigher-than-average studentwithdrawals and low retentionrates. In one class, for example,27% of the students hadpreviously received low or failinggrades and had to repeat thecourse or withdraw from it.“The students were just notgetting the material,” says RobertGarrick, associate professor atRIT’s College of Applied Scienceand Technology. “They weren’tunderstanding the intendedlearning objectives.”In a traditional laboratorysetting, students work atdifferent workstations while theinstructor walks around, answersquestions, and reviews circuitoperations.For the study, RIT researchersredesigned the courses toinclude a tablet PC and DyKnowcollaborative software for eachstudent. Classes took place in aninteractive technology classroomthat featured a multiscreendisplay connected to facultyand student tablet PCs. Classinformation and any notationswere captured, recorded, andarchived.The students told theresearchers they preferred thetablets for note taking, in-classwork, test preparation, andclassroom layout.The combination of thetechnology resources improvedthe visual connection to thematerial, student-facultyinteraction, and it enhancedthe modeling of engineeringproblems, three areas seenas critical to retaining thetechnical information, explainsRIT assistant professor LarryVillasmil.The next step for theresearchers is to study howtech-rich learning environmentscan work for underrepresentedgroups, such as deaf and hardof-hearingstudents, who makeup about 10% of the studentsin engineering and technologycourses.—Paul Hyman18 communications of the acm | january 2012 | vol. 55 | no. 1

newscomputer a doctor’s ability to recognizesome types of heart disease by the characteristicshape of waves produced byan electrocardiogram. She developeda new function, called a constrainednon-rigid translation transform, whichcould identify the similarity betweenshapes in different ECG readouts.The Power of PredictionAlthough diagnosis and treatment arekey aspects of medical care, predictionis also important, especially when it canlead to early interventions. We all know,for instance, that factors such as weightand blood pressure can give an idea ofa person’s risk of heart disease. Butthose sorts of risk scores are based onpopulation-wide models, says ShyamVisweswaran, assistant professor of biomedicalinformatics at the University ofPittsburgh. “If you build a model froma group of people who are kind of similarto the current patient, you might dobetter,” he says. Visweswaran has developedan algorithm that lets a computeruse clinical data to learn a model tailoredto one specific patient and predictoutcomes for that person.The computer takes all the data it hason the patient, such as age, blood pressure,and lab results, and then picks onevariable and builds a model of all thepatients in its database who share thatvariable. It could, for instance, compareeveryone in the 50–55 age group.“With a large numberof pre-diagnosedpatient datasetsavailable in electronichealth records,physicians can nowbenefit from theopinion of their peerson cases similar totheir patients,”says TanveerSyeda-Mahmood.It builds a model for each variable itcan find, looks at which ones best fitthe patient at hand, and then averagesthe best models to make a personalizedprediction of that patient’s outcome.Whereas a population model only usesa handful of variables considered to bethe best—it could be a simple checklistof several risk factors, for instance—this approach can potentially use any ofhundreds of variables. One additionaladvantage is the machine might identifysome factor that is predictive, butthat medical science was not previouslyaware of, opening up new areas for research,Visweswaran says.As with Suzuki’s pixel-based processing,this is another machine learningmethod that has benefited from thegrowth of processing power. As recentlyas five years ago it might have taken ahalf-hour to build all these models. Todayit takes less than a minute, so themachine can guide diagnoses in realtime during patient visits.This approach could help predict anintensive-care patient’s risk of an infectionspreading to other organs, whichis a notoriously difficult task, and leadto earlier or more aggressive treatment.It might help doctors decide, forinstance, which pneumonia patientsneed to be admitted to the hospitaland which patients could be sent homewith antibiotics.In the informatics program at Children’sHospital Boston, assistant professorBen Reis and his colleagues areworking on predicting a patient’s futurediagnoses years in advance. They havedeveloped Bayesian models that theycall Intelligent Histories, which combthrough the standard diagnostic codesused for billing, to find patterns in a patient’shistory that predict risk. In theirfirst application of the work, they discoveredthey could identify patients atrisk of domestic abuse as much as twoyears before the doctors seeing thosepatients first discovered the problem.Doctors are supposed to screen patientsfor domestic abuse, but oftenmiss it until the problem becomesacute, says Reis. Not only can the computeraid in screening for known signsof abuse, it also picked up other diagnosticcodes in the test that had notbeen thought of as predictive, such asinfections, which might teach doctorssomething about domestic abuse.The Children’s Hospital Boston teamis working to expand their modeling toother types of diagnoses. At the sametime, they are refining the machinelearning itself. For instance, “we’re tryingto quantify how the quality of thedata that goes into the model affects theresults that come out,” says Reis.As the world moves to a greater useof electronic medical records, machinelearning is likely to play an even largerrole in clinical medicine, researcherspredict. Visweswaran says genetic data,in particular, is going to require complicatedcomputational models if it is goingto be of value. Soon, experts expect,the cost of gene sequencing will dropto the point that individual genomeswill become part of people’s medical records,and will be available to the samedata mining and pattern recognitionapproaches being applied to other data.Genetic data will be too complicatedand too voluminous to be handled withold-fashioned charting systems. “Youhave to have computational tools to querythis data,” Visweswaran says. “There’sno way it can be done on paper.”Further ReadingGruhl, D., et al.Aalim: A cardiac clinical decision supportsystem powered by advanced multi-modalanalytics, International Medical InformaticsConference, Cape Town, South Africa, Sept.12–15, 2010.Reis, B.Y., Kohane, I.S., and Mandl, K.D.Longitudinal histories as predictors offuture diagnoses of domestic abuse:modeling study, British Medical Journal,339, Sept. 2009.Syeda-Mahood, T., Beymer, D., and Wang, F.Shape-based matching of ECG recordings,29 th Annual International Conference of theIEEE Engineering in Medicine and BiologySociety, August 23–27, 2007, Lyon, France.Suzuki K., Zhang J., and Xu, J.Massive-training artificial neural networkcoupled with Laplacian-eigen functionbaseddimensionality reduction forcomputer-aided detection of polyps inCT colonography, IEEE Transactions onMedical Imaging 29, 11, Nov. 2010.Visweswaran, S., Angus, D.C., Hsieh, M.,Weissfeld, L., Yealy, D., and Cooper, G.F.Learning patient-specific predictive modelsfrom clinical data, Journal of BiomedicalInformatics 43, 5, Oct. 2010.Neil Savage is a science and technology writer based inLowell, MA.© 2012 ACM 0001-0782/12/01 $10.00january 2012 | vol. 55 | no. 1 | communications of the acm 19

newsTechnology | doi:10.1145/1721654.1721662Robots Gear Up forDisaster ResponseAfter 15 years of research, robots for searchand rescue may be nearing prime time.Gary AnthesPhotograph courtesy of the Center for Robot-Assisted Search and RescueOn January 17, 1995, an earthquakenear Kobe, Japankilled approximately6,434 people, caused thecollapse of 200,000 buildings,and resulted in $102.5 billion indamages. Three months later a truckbomb exploded outside a federal governmentbuilding in Oklahoma City,OK, and claimed 168 lives and damagedor destroyed 324 buildings withina 16-block radius. While it was anawfully memorable year for disasters,both of these tragedies set in motiona flurry of research in robotics that observerssay could save countless livesin future disasters.Indeed, developers of search andrescue robots say the technology—which spans such diverse disciplinesas artificial intelligence, sensing, communications,materials, and mechanicalengineering—is nearly ready for deployment.Applications could includesearch, reconnaissance and mapping,removing or shoring up rubble, deliveryof supplies, medical treatment, andevacuation of casualties.However, a host of technical challengesremain. Also, researchers areconcerned that a lack of standards,scarce federal funding, and tepid interestfrom companies that don’t yet seea big market for robotic rescuers standin the way of the miniaturization, devicehardening, and systems integrationthat are needed to make the technologymature.Only one emergency response teamin the U.S.—New Jersey Task ForceOne—so far owns a robot. And the robotstested by researchers in a handfulof disasters in recent years have produceddecidedly mixed performances.“We still don’t know how to use thesethings,” says Robin Murphy, a professorof computer science and engineeringand director of the Center for Robot-A camera robot being inserted in a bore hole at Crandall Canyon Mine in Utah.Assisted Search and Rescue at TexasA&M University. “Real disasters are infrequent,and every one is different. Therobots never get used exactly the way youthink they will, and they keep uncoveringnew bottlenecks and problems. Soit’s an emerging technology.”Murphy says the devices are oftentested in unrealistically robot-friendlylabs or via simulations that don’t quiteduplicate the realities of real-life situationsthat involve dirt and sand, steepchanges in elevation, or radio-blockingmetal structures. Some of the most vexingproblems seem simple yet remainfrustratingly intractable. For example,at the Crandall Canyon Mine in Utah,where six miners and three rescue workerswere killed in 2007, mud greatly hinderedthe effectiveness of the workers’camera robot. “We steered the robot toplaces where water was dripping andturned it face-up to rinse off some of themud,” Murphy says. However, the camerarobot was eventually trapped by arock slide, causing the robot’s tether tosnap and for it to be lost.Howie Choset, associate professorof robotics at Carnegie Mellon University(CMU), specializes in snake robots,which are thin, legless devices with multiplejoints. They can go places moretraditional, track- or wheel-propelledrobots can’t, but the technology stillneeds work. “My last trial at a rubblepile in Texas didn’t go so well,” Chosetnotes. “They didn’t get over little obstaclesI thought they should have. Ourcontrol laws are still not well defined;we don’t have good feedback; we don’thave enough sensing in the robots; andtheir skins have to be better designed.”So while his mechanical snakes canperform remarkable feats such as crawlingup the interior of a vertical pipe orswimming across a pool, Choset saysa lack of funding stands in the way ofmaking the snakes truly versatile andapril 2010 | vol. 53 | no. 4 | communications of the acm 15

newsrobust. Choset needs, for instance, todevelop more mechanical snake gaits.He’d like his snakes to be able to changefrom a vertical undulating gait to a sidewindergait on command or, better yet,to autonomously switch gaits to suitnew conditions. And he’d like the mechanicalsnake to know how to executeone gait in its front segments and a differentgait at the rear segments. “Wehave developed the greatest variety ofsnake gaits in the world,” says Choset,“but a rubble pile has that many moresituations than we can anticipate.”Asked if further animal study wouldhelp, Choset replies, “It’s true you areinspired by biology, but snakes have200 bones and the snake robot has just15 links. Snakes have a material calledmuscles, but we are not going to bemaking muscles any time soon.” And,he adds, snakes have marvelous sensorsfor heat and pressure in their skins,something else technology has yet toeasily match.Three Levels of ChallengesUsers of search and rescue robots facechallenges at three levels, says SanjivSingh, a research professor at CMU’sRobotics Institute. At the lowest levellies information processing—gettingand managing information about theenvironment. At the next level comesmobility—getting the robot to whereit is needed. And at the highest levelcomes manipulation—enabling therobot to perform the appropriate physicaltask once it is in place. Singh’sEmber project, partially funded bythe U.S. National Science Foundation,seeks to aid first responders at the firsttwo levels, and in situations that aredynamic, chaotic, and often providingpoor visibility.Inside a burning building, for example,it is unlikely that the structure’scommunication systems will remainworking, and first responders won’thave a map or plan for the building.Singh’s group has developed technologywhereby a firefighter or a robot can scattersmart radio beacons inside the building.Some beacons are stationary andsome are attached to a human or robot.These nodes begin talking to each otherand autonomously organize themselvesinto an ad hoc sensor network. The radiosmeasure distances to each other and,using algorithms developed by Singh’sMiniaturization,device hardening,and systemsintegration areall needed for thematuring of searchand rescue robots.group, construct a map of their physicallayout—or a map of conditions such astemperature—and track the movementof robots or people.“Imagine there is a commanderstanding outside the building, and helooks at a screen and he can see whereall his people are inside the building,”Singh says. “And they can do this withoutany prior survey of the building,and without any power or prior communicationsinfrastructure inside thebuilding.”Constructing spatial maps fromdistance-only data has been feasible forsome time. It’s possible to draw a mapshowing the positions of cities in theU.S. solely from the intercity mileagetable at the back of an atlas, Singh says.But his innovation was the developmentof algorithms—based on Kalman filtering,Markhov methods, and Monte Carlolocalization—that can do the job witha sparsely populated distance table.Singh has also made progress at thesecond level of the hierarchy, the onedealing with robot mobility. He has developeda suite of search algorithms forteams of robots to use in spaces humanscan’t or don’t want to go. Some are suitedto looking for an immobile person,while others are geared to looking formoving people, such as an intruder. Inthe latter case, the robots might post“guards” at various locations in a buildingto spot the intruder’s movement.In addition, Singh’s algorithms canbe classified as “efficient” (find a targetin the lowest expected time); “guaranteed”(clear the environment so captureis assured); or “constrained” (maintainrobot positions that ensure networkconnectivity or line-of-sight communication).“You could combine these algorithmsif you have a team of robots ora team of robots and humans,” he says.Combining an efficient search with aguaranteed search would tend to minimizesearch time while still making surethe search ultimately succeeds.Murphy, who has become a kind ofevangelist for the search and rescuerobotics community in the U.S., saysthe technical problems associatedwith the devices will be solved in duecourse. But she says strong governmentfunding and support is neededif search and rescue robots are to seewidespread use in fewer than 10 years.The standards being developed nowat the National Institute of Standardsand Technology will also be a big help,Murphy predicts.Brilliant robotic technology exists,says Murphy, but it needs to be integratedinto complete, robust systems, andsensors and other components must bemade smaller, stronger, and cheaper.All of this requires corporate effort, shenotes. “We are just inches away,” Murphysays. “A lot of the software is justwaiting for the hardware to catch up.”Further ReadingMurphy, R., Tadokoro, S., Nardi, D., Jacoff, A.,Fiorini, P., Choset, H., Erkmen, A.Search and rescue robotics. SpringerHandbook of Robotics, Siciliano, B. andKhatib, O. (eds.). Springer Science andBusiness Media, Secaucus, NJ, 2008.Kumar, V., Rus, D., Singh, S.Robot and sensor networks for firstresponders. IEEE Pervasive Computing 3, 4,October-December, 2004.Shammas, E., Choset, H., Rizzi, A.Geometric motion planning analysis fortwo classes of underactuated mechanicalsystems. International Journal of RoboticsResearch 26, 10, October 2007.International Rescue System Institutehttp://www.rescuesystem.org/IRSweb/en/IRSU.htmlBiorobotics Laboratory, Carnegie MellonUniversityhttp://www.cs.cmu.edu/~biorobotics/Dekker, S.The Field Guide to Understanding HumanError. Ashgate Publishing, Farnham, Surrey,U.K., 2006.Gary Anthes is a technology writer and editor based inArlington, VA.© 2010 ACM 0001-0782/10/0400 $10.0016 communications of the acm | april 2010 | vol. 53 | no. 4

newsTechnology | doi:10.1145/1941487.1941494I, Domestic RobotWith recent advances in laser rangefinders, faster algorithms,and open source robotic operating systems, researchers are increasingdomestic robots’ semantic and situational awareness.Gregory GothIn d u s t rial robots, fixed-loplesof advanced manufactur-cat i o n and single-functionmachines, have long been sta-ing settings. Medical robots,which can help surgeons operate withsmaller incisions and cause less bloodloss than traditional surgical methods,are making fast inroads in metropolitanand suburban hospitals. Rescue robots,included wheeled and snake-likerobots, are increasingly common, andwere deployed in the search for survivorsin the aftermath of the earthquakeand tsunami that recently struck Japan.On the other hand, the promise ofmultipurpose domestic assistance robots,capable of a wide range of tasks,has been a distant goal.However, recent advances in hardwaresuch as laser rangefinders, opensource robotic operating systems, andfaster algorithms have emboldened researchers.Robots are now capable offolding laundry, discerning where toplace an object on cluttered surfaces,and detecting the presence of peoplein a typical room setting.“It’s easy for me to be optimistic,but if robots aren’t actually being usefuland fairly widespread in 10 years,then I will be fairly disappointed,” saysCharles Kemp, assistant professor ofbiomedical engineering at GeorgiaTech University.Sensors Enable AwarenessIn recent months, numerous researchteams have published papers detailingadvances in robots’ perceptual capabilities.These perceptual advancesenable the robots’ mechanical componentsto complete domestic tasks hithertoimpossible.Kemp and his research team havepioneered semantic and situationalawareness in robots through severalmethods, including the creation ofradio frequency identification (RFID)Willow Garage’s PR2, an open sourcerobotics research and development platform.semantic tags on common objectssuch as light switches, and by combiningsensor data taken from both twodimensionalcamera data and threedimensionalpoint clouds gathered bylaser rangefinders.University of Bonn researchers JörgStückler and Sven Behnke also demonstratedsuccess, using a combinationof 2D laser and camera sensors.They programmed a mobile servicerobot to combine laser rangefinderdata that hypothesizes the presence ofa person’s legs and torso with 2D frontaland profile images of the detectedface.Stückler and Behnke also modeledthe semantic probability of detectinga person’s presence in differentlocations of a room—highprobability in a chair and low probabilityon a bookshelf, for instance—and supplied the robot with thatknowledge. The prior knowledge ofthe room semantics and precalculatedrange of likely valid facial heighthelps the Bonn researchers discernfalse positive returns.Steve Cousins, CEO of Willow Garage,which manufactures the openplatform general-purpose PR2 robot,says further advances in perceptual capabilitiesmay be even more likely withthe recent debut of sensing technologythat enables a computer to analyze anarea in three dimensions and then tocreate what the technology’s manufacturer,PrimeSense, calls a synchronizeddepth image. The technology sells forless than 1/20th of the de facto standardresearch rangefinder, which costsabout $5,000. Both Cousins and Kempbelieve the low cost of the PrimeSensesensor (it is a key component of Microsoft’sKinect gaming system) may leadto a surge in situational and semanticrobotic research. Kemp says his teamrecently installed one of the new sensorsto its PR2.In essence, Kemp says its real-timetechnology greatly simplifies a robot’sdata-gathering process.Prior to installing the new sensor,on projects such as the work on makingthe robot discern clutter, he says“we had to tilt the laser rangefinderup and down, then snap a picture andrelate those two things. That’s a prettyslow process and really expensive.”A Semantic DatabaseKemp says there are two distinct researchareas for similar problem setsin domestic robotics: those related toperceptual problem sets, and thoserelated to mechanical awareness. Forexample, a roving robot meant to helpa person with basic housekeepingchores must not only know how to differentiatea refrigerator door handlefrom a light switch, but it must also beable to calculate which approach itsarms must take, and how firmly it mustgrip the respective levers.PhotoGraph courtesy of Willow Garage16 communications of the acm | may 2011 | vol. 54 | no. 5

newsIn the experiment using RFID tags,Kemp created a semantic databasethe robot could refer to after identifyingan object. The database containsinstructions on how the robot shouldact upon an object. For example, under“actions,” after a robot identifies andcontacts a light switch, the commandsare “off: push bottom” and “on: pushtop.” Each of these actions is furthersub-programmed with a force thresholdthe robot should not exceed.Kemp is also investigating anotherapproach to providing robots withsuch situational awareness that entailsequipping human subjects with touchsensors. The sensors are held duringthe completion of common tasks suchas opening refrigerators and cabinetdoors in multiple settings. The informationon the kinematics and forces ofsuch actions is then entered into a databasea service robot can access whenit approaches one of these objects enroute to performing a task.“If the robot knows it is a refrigerator,it doesn’t have to have worked withthat specific refrigerator before,” hesays. “If the semantic class is ‘refrigerator’it can know what to expect and bemore intelligent about its manipulation.This can make it more robust andintroduces this notion of physicallygrounded common sense about thingslike how hard you should pull whenopening a door.”Offboard computation akin to thekinematic database is also being doneto improve already successful robotictasks. A team of researchers led by PieterAbbeel, an assistant professor ofcomputer science at the University ofCalifornia, Berkeley, programmed ageneral-purpose Willow Garage PR2 robotto fold towels randomly laid downon a tabletop by using a dense opticalflow algorithm and high-resolutionstereo perception of the towels’ edgesand likely corners. Abbeel’s experimentyielded a perfect 50-out-of-50-attemptsuccess rate; the robot was able to recalculatefailures in the 22 instances thatwere not initially successful by droppingthe towel, regrasping a corner, and carryingon until the task was completed.Abbeel says his team has been ableto greatly reduce the amount of timenecessary to fold each towel in subsequentexperiments, from 25 minutesto approximately four minutes, by utilizinga new approach: rather than relyheavily upon onboard perceptual data,Abbeel has performed parallel computationson the Amazon cloud on meshmodels. Those models, he says, are“triangles essentially put together likepeople using computer graphics orphysics-based simulations. Once youhave that mesh model, you can do asimulation of how this article of clothingwould behave depending on whereyou pick it up.”The new approach, he says, relieson observations that the bottommostpoint of any hanging article is usuallya corner. Two consecutive grasps of atowel, he says, will be highly likely toyield two diagonally opposed corners.For t-shirts, he says, likely consecutivegrasps will be at the end of two sleevesfor a long-sleeved shirt or the end ofone sleeve and diagonally across at thehip for a short-sleeved shirt.“There are a few of these configurationsyou are very likely to end up in,then all you need to do perception-wiseis to differentiate between these veryfew possibilities,” Abbeel says.ROS is BossAnother hallmark advance of the domesticrobot community is the growthof an open-source ecosystem, builtaround the BSD-licensed Robot OperatingSystem (ROS), largely maintained byWillow Garage and Stanford University.“Our goal has basically been to setthe foundation for a new industry tostart,” Cousins says. “We want twopeople to be able to get together in agarage and get a robotics business offthe ground really quickly. If you haveto build software as well as hardwarefrom scratch, it’s nearly impossible todo that.”Abbeel says the ROS ecosystem maygo a long way to taking the robots outof the lab and into real-world locations.“In order for these robots to maketheir way into houses and becomecommercially viable, there will needto be some sort of bootstrapping,” Abbeelsays. “It will be very importantfor people to do some applications extremelywell, and there has to be morethan one. So I hope what may be happening,with robots in different places,is that different schools will develop atrue sensibility for the robot, and thesethings could potentially bootstrap theprocess and bring the price down. Asingle app won’t be enough.”Cousins says the combination offalling hardware prices for devicessuch as the PrimeSense sensor, andthe blooming ROS ecosystem might beanalogous to the personal computerresearch of the early 1970s, specificallycomparing the PR2 to the iconic XeroxAlto desktop computer. List price onthe PR2 is $400,000.“Right now the PR2 is the platformto work on if you want to do mobilemanipulation research,” Cousins says.“It’s a little expensive, but in today’sdollars it’s about the same as the Alto.It’s not going to be the robot you putinto your grandmother’s home, but thesoftware we develop on the PR2 willlikely be a key component of the market.I think ROS is going to be drivingthose future personal robots.”Further ReadingStückler, J. and Behnke, S.Improving people awareness of servicerobots by semantic scene knowledge,Proceedings of RoboCup InternationalSymposium, Singapore, June 25, 2010.Maitin-Shepard, J., Cusumano-Towner, M.,Lei, J., and Abbeel, P.Cloth grasp point detection based onmultiple-view geometric cues withapplication to robot towel folding, 2010IEEE International Conference on Roboticsand Automation, Anchorage, AK, May 3–8,2010.Schuster, M.J., Okerman, J., Nguyen, H.,Rehg, J.M., and Kemp, C.C.Perceiving clutter and surfaces for objectplacement in indoor environments, 2010IEEE-RAS International Conference onHumanoid Robots, Nashville, TN, Dec. 6–8,2010.Yamazaki, A., Yamazaki, K., Burdelski, M.,Kuno, Y., and Fukushima, M.Coordination of verbal and non-verbalactions in human–robot interaction atmuseums and exhibitions, Journal ofPragmatics 42, 9, Sept. 2010.Attamimi, M., Mizutani, A., Nakamura, T.,Sugiura, K., Nagai, T., Iwahashi, N.,Okada, H., and Omori, T.Learning novel objects using out-ofvocabularyword segmentation and objectextraction for home assistant robots, 2010IEEE International Conference on Roboticsand Automation, Anchorage, AK, May 3–8,2010.Gregory Goth is an Oakville, CT-based writer whospecializes in science and technology.© 2011 ACM 0001-0782/11/05 $10.00may 2011 | vol. 54 | no. 5 | communications of the acm 17

Vviewpointsdoi:10.1145/1785414.1785428Computing EthicsWork Life inthe Robotic AgeTechnological change results in changes inexpectations, in this case affecting the workplace.Jason BorensteinRobots are beiNG designedto perform a broader arrayof work-related tasks.Global economic hardshipsmay be (temporarily)causing the demand for industrial robotsto decline, 4 but improvements inartificial intelligence and the drive forefficiency will likely encourage companiesto develop and use increasingamounts of robotic workers. Thoughthe justification for automation isoften couched in the language of liberation,this oversimplifies the complexitiesassociated with technologicalchange. Merely because technologyis well designed from an engineeringperspective, it does not follow thatsociety’s problems are solved. This isnot to say that efforts to create roboticworkers must stop, but the roboticscommunity must be diligent in dealingwith emerging ethical issues. Designpathways must be selected thateither mitigate or prevent the negativeconsequences of using robots inthe workplace. Otherwise, troublinghistorical occurrences, such as thedecimation of certain segments of thework force, might be repeated.A view of a robot arm used in world’s first remote heart operation performed at GlenfieldHospital in Leichester, U.K., on April 28, 2010.With each significant technologicalchange, visions of how improvedand efficient our lives will become aretypically offered. To some degree, thepromise that we will be “liberated”from performing repetitive and mundanetasks has held true. Most of usdo not mourn the passing of havingto wash clothes or dishes by hand.Yet expectations in both our personaland professional lives tend to shiftcorrespondingly, which in many wayscounterbalances the “liberating” featuresthat technology offers. RuthSchwartz Cowan recognized years agothat the introduction of electronic devicesinto the home did not free womenfrom the burden of doing householdchores. As Cowan states, “Whata strange paradox that in the face ofso many labor-saving devices, littlelabor appears to have been saved!” 1In short, increasing expectations absorbedall of the extra time that wassupposed to be freed up.Similarly, we need to seriously considerhow the increased use of robotswill alter workplace expectations.For instance, if robots can help surgicalprocedures to be completed morerapidly, will demands on surgeonsincrease so they will have to performmore procedures per day? Expectationsin terms of what it means to be a “good”professional are also likely to change,especially if a robot’s error rate is lowerthan a human’s. Briefly put, we shouldbe wary of predictions that robots willbe our liberators considering how thePhotoGRAPh by Rui VieIRA/AP photo30 communications of the acm | july 2010 | vol. 53 | no. 7

viewpointstypical workweek does not seem to begetting shorter or less demanding inthe digital age.The U.S. military is enjoying thebenefits of robots since they can complete“dull, dirty, or dangerous” tasks,and their labor is very useful in the civilianrealm as well. Yet automationcan eliminate job opportunities andusually causes the demographics of thework force to be significantly altered ina relatively short amount of time. Employersfind robots to be rather enticingsince they do not receive benefits orrequest vacation time. Through the designchoices they make, scientists andengineers play a key role in determiningthe kinds of employment practicesthat can and will transpire.Employment Impactsand ImplicationsIf categories of jobs do indeed vanish asa result of robots, will the relevant skillsof displaced workers be transferred toanother application or will those skillsbe rendered obsolete? This concernis not unique to robots. But what maybe a new variation now is that the jobsavailable to humans may be drasticallyreduced as computers, the Internet,and robots replace humans in employmentsectors that used to be thought ofas immune to automation. At present, itis fairly difficult for people to find workthat is not connected in some way tothese technologies. This developmentmight not be conducive to the flourishingof each person’s respective talents,and robots are likely to exacerbate thissituation. Also, the type of skills that willbe in demand if and when the roboticage takes hold might be obvious in someways but not so apparent in others. aThe impact of robotic workers canand will extend beyond the eliminationof labor-intensive jobs, whichcaptures a key reason why the ethicaldimensions of robots seem to be drawingincreased attention. It is not onlypossible to eliminate “dangerous” and“boring” work but at least some jobsrequiring specialized expertise, suchaFor example, in Wired for War, P.W. Singerdiscusses how cooks might have more job securitythan military pilots because they canprepare food in creative ways. In the civilianrealm, he reassures hairstylists by suggestingtheir specific abilities may keep them employed;Penguin Press, NY, 2009, 130–132.Scientists andengineers shouldreflect on their ethicalresponsibilities tocommunicate withthe public abouta robot’s capabilitiesand limitations.as being a surgeon, may start to disappear.A decade ago, Bill Joy, the cofounderof Sun Microsystems, famouslywarned against this. 2 Even if we don’tshare Joy’s apprehension about the futureof robotics, we can still appreciatethe perils of trying to replace “uniquelyhuman” abilities such as critical thinkingand intuition.To illustrate this point, we can lookat the robots being created to assist withthe health care needs of elderly populations.An outgrowth of this effort is thatit could subtly or perhaps dramaticallychange how nursing homes function. Inprinciple, robots could free up the timeof nursing home staff; for example, a roboticassistant can provide medicationreminders or warnings if a resident isin danger. Such a robotic counterpartmight enable human workers to bemore caring and productive. However,nursing homes and other care facilitieswill be tempted to downsize theirhuman staff when a robot is “hired”instead of freeing up human staff togive more time to residents. 3 Sincemany nursing home residents in theU.S. and elsewhere already do not getenough care and individualized attention,this is a very troubling possibility.Theoretically, an increased emphasison in-home care could for example leadto the creation of other types of jobsbut we should be skeptical about this.Financial considerations, the drive forefficiency, and overconfidence in technologyare strong driving forces that canpush humans “out of the loop.”On a related note, reliance on automationmay exacerbate a common humantendency to shift our attention to adifferent task when we believe (perhapsfalsely) that we can trust someone orsomething else to deal with the task athand. b Returning to the issue of healthcare, will nursing home staff be less attentiveif a robotic assistant is placed ina resident’s room? The more reliable wethink automated systems are, the morelikely it is our attention will stray. Whatcomplicates matters is that this typeof behavioral shift might not be consciouslydetected. Hence, it would bewise to temper the confidence that usersplace in robots and other automatedsystems, especially when people couldbe significantly harmed. This could beaccomplished in part by ensuring thatrisks are transparently presented tousers. To that end, scientists and engineersshould reflect on their ethical responsibilitiesto communicate with thepublic about a robot’s capabilities andlimitations, and not merely leave it tomarketers, sales departments, and othersto fill this role.ConclusionEthical concerns about integrating robotsinto the workplace are becomingincreasingly pronounced. Again, the intentionhere is not to stop innovation.Rather, the hope is to inform the designprocess. Ideally, the robotics communitywill select design pathways thatmitigate the associated concerns andthereby enhance the public’s lives.b Placing too much confidence in technology, oftenat the expense of other sources of information,seems to be a growing problem with GPSin automobiles; see for example, Is your GPSnavigator a friend or foe?” The Sydney MorningHerald, (Jan. 12, 2010); http://www.smh.com.au/executive-style/gadgets/is-your-gps-navigator-afriend-or-foe-20100112-m4ei.htmlReferences1. Cowan, R.S. More Work For Mother: The Ironies OfHousehold Technology From The Open Hearth To TheMicrowave. Basic Books, 1983, 44.2. Joy, B. Why the future doesn’t need us. Wired 8, 4 (Apr.2000).3. Sparrow, R. and Sparrow, L. In the hands ofmachines? The future of aged care. Minds andMachines 16, 2 (May 2006), 141–161.4. Tabuchi, H. In Japan, machines for work and play areidle. The New York Times (July 12, 2009); http://www.nytimes.com/2009/07/13/technology/13robot.htmlJason Borenstein (borenstein@gatech.edu) is the directorof Graduate Research Ethics Programs in the School ofPublic Policy at Georgia Tech in Atlanta, GA.The author would like to thank Rachelle Hollander, KeithW. Miller, and the anonymous reviewers for their helpfulinsights and guidance.Copyright held by author.july 2010 | vol. 53 | no. 7 | communications of the acm 31

Vviewpointsdoi:10.1145/1787234.1787248ViewpointRights for AutonomousArtificial Agents?The growing role of artificial agents necessitates modifyinglegal frameworks to better address human interests.Samir ChopraIt is a commonplace occurrencetoday that computerprograms, which arise fromthe area of research in artificialintelligence knownas intelligent agents, function autonomouslyand competently; 1 theywork without human supervision,learn, and, while remaining ‘justprogrammed entities’, are capable ofdoing things that might not be anticipatedby their creators or users.In short, leaving philosophicaldebates about the true meaning of‘autonomy’ aside, they are worthyof being termed ‘autonomous artificialagents’. a And on present trends,we, along with our current social andeconomic institutions, will increasinglyinteract with them. They will buygoods for us, possibly after carryingout negotiations with other artificialagents, process our applications forcredit cards or visas, and even makedecisions on our behalf (in smarterversions of governmental systemssuch as TIERS 2 and in the ever-increasingarray of systems supporting legaldecision-making 3 ). As we interact withthese artificial agents in unsupervisedsettings with no human mediators,their increasingly sophisticated functionalityand behavior create awkwardaJim Cunningham has pointed out that a certaindegree of autonomy is present in allprograms; consider Web servers or emaildaemons for instance. One might think of intelligentagents as a move toward one end ofthe spectrum of autonomy.The artificial agentis better understoodas the means bywhich the contractoffer is constituted.questions. If it is a reasonable assumptionthat the degree of their autonomywill increase, how should wecome to treat these entities?Societal norms and the legal systemconstrain our interactions with otherhuman beings (our fellow citizens orpeople of other nations), other legalpersons (corporations and public bodies),or animal entities. There are, inparallel, rich philosophical discussionsof the normative aspects of these interactionsin social, political, and moralphilosophy, and in epistemology andmetaphysics. The law, taking its cuesfrom these traditions, strives to providestructure to these interactions. Itanswers questions such as: What rightsdo our fellow citizens have? How dowe judge them liable for their actions?When do we attribute knowledge tothem? What sorts of responsibilitiescan (or should) be assigned to them?It is becoming increasingly clear thesequestions must be addressed withrespect to artificial agents. 4 So, whatplace within our legal system shouldthese entities occupy so that we may dojustice to the present system of socioeconomic-legalarrangements, whilecontinuing to safeguard our interests?The Contracting ProblemDiscussing rights and responsibilitiesfor programs tends to trigger thoughtsof civil rights for robots, or taking themto trial for having committed a crime orsomething else similarly fanciful. Thisis the stuff of good, bad, and simplisticscience fiction. But the legal problemscreated by the increasing use of artificialagents today are many and varied.Consider one problem, present in e-commerce: If two programs negotiatea deal (that is, my shopping bot makesa purchase for me at a Web site), doesthat mean a legally binding contract isformed between their legal principals(the company and me)?A traditional statement of the requirementsof a legally valid contract isthat “there must be two or more separateand definite parties to the contract;those parties must be in agreementi.e., there must be a consensusad idem; those parties must intend tocreate legal relations in the sense thepromises of each side are to be enforceablesimply because they are contractualpromises; the promises of each partymust be supported by considerationi.e., something valuable given in returnfor the promise.” 538 communications of the acm | august 2010 | vol. 53 | no. 8

viewpointsIllustration by john herseyThese requirements give rise to difficultiesin accounting for contractsreached through artificial agents andhave sparked a lively debate as to howthe law should account for contractsthat are concluded in this way. Mostfundamentally, doctrinal difficultiesstem from the requirement there betwo parties involved in contracting: asartificial agents are not considered legalpersons, they are not parties to thecontract. Therefore, in a sale broughtabout by means of an artificial agent,only the buyer and seller can be therelevant parties to the contract. Thisentails difficulties in satisfying therequirement the two parties shouldbe in agreement, since in many casesone party will be unaware of the termsof the particular contract entered intoby its artificial agent. Furthermore,in relation to the requirement thereshould be an intention to form legalrelations between the parties, if theagent’s principal is not aware of theparticular contract being concluded,how can the required intention be attributed?Legal scholarship has suggesteda variety of solutions, 6 ranging fromthe idea programs should be treatedas “mere tools” of their principals tothose suggesting programs be grantedfull legal personhood in order togrant legal efficacy to the deals enteredinto by them. Some of the suggestedsolutions struggle to solve this problemwhen: protocols between buyersand sellers (and their agents) are notspecified in advance; the terms of usegoverning individual transactions arenot specified; the terms of a contractare not finalized via human review; orwhen agents capable of determiningthe terms of contracts are employed.In these settings, agents might arriveat negotiated or reasoned decisions,which their principals might not haveagreed to had they been given the opportunityto review the decision. Giventhis fact the agent cannot just be understoodas a ‘mere tool’ or ‘meansof communication’ of the principal;rather, the artificial agent is better understoodas the means by which thecontract offer is constituted. bb This discussion is considerably oversimplifiedbut I hope the outlines of the legal problemare clear.Artificial Agents as Legal AgentsOne possible solution, which would requireus to grant some legal standingto the programs themselves, 7 wouldbe to treat programs as legal agents oftheir principals, empowered by law toengage in all those transactions coveredby the scope of their authority.We would understand the program ashaving the authority to enter into contractswith customers, much as humanagents do for a corporate principal.Some of its actions will be attributedto its corporate principal (for instance,the contracts it enters into), whilethose outside the scope of its authoritywill not. The ‘knowledge’ it acquiresduring transactions, such as customerinformation, can be attributed to thecorporate principal, in the way thatArt in Developmentknowledge of human agents is. Lastly,the established theory of liability forprincipal-agent relationships can beapplied to this situation. The details ofthis solution aside, the most importantaspect here is that, unlike a car, a programis neither a thing nor a tool; rather,it is an entity with legal standing inour system.In granting the status of a legalagent to a computer program, weare not so much granting rights toprograms as protecting those thatemploy and interact with them. Understandingappropriately sophisticatedprograms as legal agents of theirprincipals could be a crucial step toregulating their presence in our lives.It will enable us to draw upon a vastbody of well-developed law that dealswith the agent-principal relationship,and in a way that safeguards the rightsof the principal user and all concernedthird parties. Without this framework,neither third parties nor principals areadequately protected. Instead, we findourselves in a situation where increasinglysophisticated entities determinethe terms of transactions that affectothers and place constraints on theiractions, though with no well-definedlegal standing of their own. Viewing aprogram as a legal agent of the employercould represent an economically efficient,doctrinally satisfying, and fairresolution that protects our interests,without in any way diminishing oursense of ourselves.Rights and Legal Personhoodfor Artificial AgentsThere are two ways to understand thegranting of rights, such as legal agency,to artificial agents. Rights might begranted to artificial agents as a way ofprotecting the interests of others; andartificial agents might interact with,and impinge on, social, political, andlegal institutions in such a way thatthe only coherent understanding oftheir social role emerges by modifyingtheir status in our legal system—perhapstreating them as legal agents oftheir principals, or perhaps treatingthem as legal persons like we do corporationsor other human beings. Andwhen they enjoy such elevation, theyaugust 2010 | vol. 53 | no. 8 | communications of the acm 39

viewpointsArtificial agentshave a long way togo before we cancountenance themas philosophicalpersons.must conform to the standards expectedof the other entities that enjoystanding in our legal system.The question of legal personalitysuggests the candidate entity’s presencein our networks of legal and socialmeanings has attained a level ofsignificance that demands reclassification.An entity is a viable candidate forlegal personality in this sense providedit fits within our networks of social, political,and economic relations in sucha way that it can coherently be a subjectof legal rulings. Thus, the real questionis whether the scope and extent of artificialagent interactions have reachedsuch a stage. Answers to this questionwill reveal what we take to be valuableand useful in our future society as well,for we will be engaged in determiningwhat sorts of interactions artificialagents should be engaged in for us tobe convinced that the question of legalpersonality has become a live issue.While the idea of computer programsbeing legal persons might sound fanciful,it is worth noting the law has neverconsidered humanity a necessary orsufficient condition for being a person.For example, in 19th century England,women were not full persons; and, in themodern era, the corporation has beengranted legal personhood. c The decisionto grant personhood to corporations isinstructive because it shows that grantingpersonhood is a pragmatic decisiontaken in order to best facilitate humancommerce and interests. In so doing, wedid not promote or elevate corporations;cIn his Max Weber Lecture, “Rights of Non-humans?Electronic Agents and Animals as NewActors in Politics and Law,” Gunther Teubnernotes that animals were often treated as legalactors including being brought to trial.we attended to the interests of humans.Artificial agents have a long way togo before we can countenance themas philosophical persons. But theirroles in our society might grow to apoint where the optimal strategy is togrant them some form of limited legalpersonhood. Until then, we should acknowledgetheir growing roles in ourlives and make appropriate adjustmentsto our legal frameworks so thatour interests are best addressed. Indeed,this area requires an internationallegal framework to address the ubiquityof artificial agents on the Internet,and their deployment across nationalborders. d I have merely scratched thesurface of a huge, complex, multidisciplinarydebate; in the years to come, wecan only expect that more complexitiesand subtleties will arise.d The work being done on the “Alfabiite” project(at Imperial College London, NRCCL Oslo,CIRFID Bologna) may be of interest in providingguidance in this regard.References and Further Reading1. The fast-growing literature on agent technologies istruly gigantic; for introductions, see M.J. Wooldridge,Reasoning about Rational Agents, MIT Press,Cambridge, MA, 2000, and N.R. Jennings and M.J.Wooldridge, Eds., Agent Technology: Foundations,Applications and Markets, Springer Verlag, 1998.2. See http://www.hhs.state.tx.us/consolidation/IE/TIERS.shtml3. The Proceedings of the International Conferences onArtificial Intelligence and Law (http://www.iaail.org/past-icail-conferences/index.html), and the journalArtificial Intelligence and Law are rich sources ofinformation on these systems.4. A very good source of material on the legal issuesgenerated by the increasing use of artificial agentsmay be found at the Law and Electronic AgentsWorkshops site: http://www.lea-online.net/pubs5. Halsbury’s Laws of England (4th edition) Vol.9 paragraph 203; c.f. Restatement (Second) ofContracts, § 36. There is a large amount of literature in this area; somevery good treatments of the contracting problem maybe found in: T. Allan and R. Widdison, “Can computersmake contracts?”, Harvard Journal of Law andTechnology 9 (1996), 25–52,; A. Bellia Jr., “Contractingwith electronic agents,” Emory Law Journal 50,4 (2001), 1063; I. Kerr, “Ensuring the success ofcontract formation in agent-mediated electroniccommerce,” Electronic Commerce Research 1, (2001),183–202; E. Weitzenboeck, “Electronic agents and theformation of contracts”, International Journal of Lawand Information Technology 9, 3 (2001), 204–234.Various international trade agreements such as thoseformulated by the UNCITRAL or national legislationssuch as the UCITA have not as yet resulted in clarityin these areas.7. S. Chopra and L. White, “Artificial agents—Personhoodin law and philosophy,” in Proceedings of the EuropeanConference on Artificial Intelligence, 2004 and S.Chopra and L. White, A Legal Theory for AutonomousArtificial Agents, University of Michigan Press, to bepublished.Samir Chopra (schopra@sci.brooklyn.cuny.edu) is anassociate professor in the Department of Philosophy atBrooklyn College of the City University of New York.Copyright held by author.40 communications of the acm | august 2010 | vol. 53 | no. 8

last byteFuture Tense, one of the revolving features on this page, presents stories andessays from the intersection of computational science and technological speculation,their boundaries limited only by our ability to imagine what will and could be.DOI:10.1145/1562164.1562192Jaron LanierFuture TenseConfusions of the Hive MindCherish the individual.Be cautious about the artificial intelligenceapproach to computer science.It is impossible to differentiatethe actual achievement of AI from thedegree to which people change whenconfronted with what is purportedto be intelligent technology. We humansare vulnerable to bending overbackward, sometimes making ourselvessignificantly stupider, in orderto make an algorithm seem smart. Agreat many people in the U.S., as wellas elsewhere, demonstrated this dangerwhen they interacted foolishlywith deeply flawed algorithms relatedto the credit and mortgage industries.There is an even greater economicdanger ahead as it relates to the ideaof AI. If we are gullible enough to expectemergent large-scale intelligenceto arise from the vast connections ofthe worldwide Internet, as has beenproposed with increasing frequencyin Communications and elsewhere,then we risk undermining the valuewe place on human labor and creativity.We might thus ruin the most successfuldesign yet invented for thepurpose of generating and preservingindividual human dignity and liberty—capitalism.Those who believe in the imminentarrival of global AI (possibly emergingfrom the computing clouds) pretendthat all the information we humansupload actually comes from somemysterious supernatural dimension.There’s an economic component tothe way we lie to ourselves to supportthis confusion. Millions of us anonymouslyupload our online offerings—thoughts, pictures, videos, links, votes,and more. Or, if not anonymously,we often express ourselves in such afragmentary way, as with tweets, thatthere is no room left for personality.Under these circumstances we acceptthat we will not be paid for our acts ofexpression, as if we are engaged in amassive economic ritual to reify thefalsehood that a global supernaturalbrain is speaking, instead of us.The idea of creativity emergingautonomously from the computingclouds has the potential to ruin whatmight be called the endgame of basictechnological development. Willtechnology good enough to providecomfort and security usher in a goldenage for all? Or will we diverge intotwo species, one relatively lucky, theother relatively left out, as predictedby H.G. Wells in his novel The TimeMachine in 1898?The rarified beneficiaries mightturn out to be the owners of the computingclouds, while the rest might beinundated with [continued on p. 111]PHOTOGRAPH BY Rami K112 communications of the acm | september 2009 | vol. 52 | no. 9

last byteLAST BYTE[continued from p. 112] advertising.The bifurcation of humanity could besustained only so long as those on thereceiving end have money to spend.But as more things become free inorder to support advertising, fewer ofus will be making money. The dénouementwould probably be some sort ofviolent swing toward socialism.This might sound like an extremescenario, but consider how muchmore difficult it is for certain creativepeople to earn a living today than theydid before the public Internet becamea global social phenomenon. Themost tormented examples are probablyrecording musicians and investigativejournalists.Alas, it is now common to hearsuggestions that people in this predicamentshould revert to retro (inevitablymore physical) strategies ofsustenance, like selling branded T-shirts and other merchandise. Thisis a sad reversal of what had been oneof the brightest aspects of technologicalprogress. Prior to the centrality of“open culture” and the rise of onlinecollectivization, technological progressgenerally supported ever morecerebral, creative, and comfortablemeans of making a living.Now extrapolate: How long will itbe before cheap fabricating robots areable to download T-shirt designs fromthe cloud and automatically manufacturecustomized clothing as easily asone downloads music today? And howlong after that will it be before personalrobots are able to build copiesof the latest medical implant or othergadgets from an online design? Theanswers are likely to be measured indecades, not centuries. If robotics iseventually good enough to harvest thegarbage dumps of the world for materialsand transform them into manufacturedproducts, then a plateau willhave been reached. At that time, allconsumer technology will becomemedia technology. Even those whohoped to make a living from T-shirtswill join the investigative journalistand recording musician in poverty.How far back in history toward thestone age will people have to devolvein order to find a way to make a livingHow long willit be before cheapfabricating robotsare able to downloadT-shirt designsfrom the cloudand automaticallymanufacturecustomized clothingas easily as onedownloads musictoday?when fabricating robots are that good?Will people be forced by the marketplaceto work the fields, as academicsdid under various Maoist-type regimes?Not with good robots around.Surely, robots will eventually also do abetter job tending the crops.If you go back to some of the earliestthinking about how informationtechnology might interact withthe patterns of human life, you’llfind examples of people who thoughtahead to this potential dilemma. Forinstance, Ted Nelson, probably thefirst person to really think throughhow something like the Web might bebuilt and how it would influence humansociety, proposed in the 1960sa design in which each copy of a fileexisted, from a logical point of view,in only one instance. Any user couldmake micropayments to gain access.The conflict between file sharing andDRM would be defused because therewould be little motivation to makecopies. Accessing files would be enticinglycheap, but everyone would makesome incremental amount of moneyfrom sharing files with everyone else.A new social contract would emergebased on self-interest. This was notjust a proposal to extend capitalism,but to broaden its benefits to a greatervariety of people, since all wouldbe able to upload interesting bits asneeded.A popular objection when Nelsonproposed this design was that fewpeople had anything of interest orvalue to say, and if they tried to saywhat they could, no one else would beinterested. Fortunately, the rise of socialnetworking has proved these objectionsunfounded.I directly experienced a later period,in the 1970s and 1980s, when Nelsonwas no longer a solitary pioneer.Much of the underlying architectureand ideology that guides the publicInternet today appeared in rough cutduring those years. The ideas hadshifted. Nelson was attacked by thecampus left of the time over his willingnessto imagine a future in whichmoney continued to be important.Meanwhile, the culture of AI fascinatedengineers, drawing their attentionaway from the problem of how toreward human creativity that had sofascinated Nelson.We ended up with an Internet andWeb that is, for the moment, a sort ofcross between mass collective implementationof a Turing Test, throughdesigns like Twitter, and the clumsyfantasy of armchair pseudo-Maoists. Irealize these words could strike manyas alarmist. If this is the case for you,please look into the history of collectivistdesign in human affairs. Suchdesigns often appear enlightened atfirst, with a special way of enchantingidealistic young people. But they havealso engendered the worst social disastersof the past century.That’s why I reject the idea that acollective or emergent intelligenceis appearing through the computingclouds. We’ll never know if it’s reallythere, or if we have collectively becomeidiots.Jaron Lanier is a computer scientist interested ininterpersonal perception, biomimetic computing, and newdisplays and sensors. He received a Career Award fromthe IEEE in 2009 for his lifetime contributions to virtualreality research and is presently working at Microsoft onintriguing unannounced projects.© 2009 ACM 0001-0782/09/0900 $10.00september 2009 | vol. 52 | no. 9 | communications of the acm 111