BEN SHNEIDERMAN Manipulation - ACS Lab

20
42 interactions...november + december 1997 v s BEN SHNEIDERMAN B en Shneiderman is a long-time proponent of direct manipulation for user interfaces. Direct manipulation affords the user control and predictability in their interfaces. Pattie Maes believes direct manipulation will have to give way to some form of delegation— namely software agents. Should users give up complete control of their interaction with interfaces? Will users want to risk depending on “agents” that learn their likes and dislikes and act on a user’s behalf? Ben and Pattie debated these issues and more at both IUI 97 (Intelligent User Interfaces conference - January 6–9, 1997) and again at CHI 97 in Atlanta (March 22–27, 1997). Read on and decide for yourself where the future of interfaces should be headed—and why.

Transcript of BEN SHNEIDERMAN Manipulation - ACS Lab

Page 1: BEN SHNEIDERMAN Manipulation - ACS Lab

42 i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

vsDirect

Manipulation

B E N S H N E I D E R M A N

Ben Shneiderman is a long-time proponent of direct manipulation

for user interfaces. Direct manipulation affords the user control

and predictability in their interfaces. Pattie Maes believes direct

manipulation will have to give way to some form of delegation—

namely software agents. Should users give up complete control of

their interaction with interfaces? Will users want to risk depending on

“agents” that learn their likes and dislikes and act on a user’s behalf?

Ben and Pattie debated these issues and more at both IUI 97

(Intelligent User Interfaces conference - January 6–9, 1997) and again

at CHI 97 in Atlanta (March 22–27, 1997). Read on and decide for

yourself where the future of interfaces should be headed—and why.

Page 2: BEN SHNEIDERMAN Manipulation - ACS Lab

43i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

d e b a t e

P A T T I E M A E S

vsvs Interface

Agents

on

Excerpts from debatesat IUI 97 and CHI 97

Page 3: BEN SHNEIDERMAN Manipulation - ACS Lab

44 i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

Jim: Okay, welcome all to the afternoon ses-sion. My name is Jim Alty. I think I am sup-posed to be the moderator, whatever thatmeans, for this session. I hope things don’t gettoo rough. The debate topic this afternoon, asyou all know, is direct manipulation versusintelligent agents. On my right I’ve got BenShneiderman—fresh from his triumphs thismorning when he’s been gnawing away at theagent people—from the University of Mary-land. On my left, of course, fresh for a fight isPattie Maes from the MIT Media Laboratory.

Let me just explain. It’s 15 minutes fromeach speaker to place their position, and then5 minutes allowed each for rebuttal. Then weopen it up to the floor for about 30 minutesor so. Could you please use the microphones?If you want to line up behind the micro-phones, I will select you to make your com-ments. Then at the end, there will be two5-minute summing ups. So, let us commencethe debate.

Ben: First, my thanks to the organizers of thisIntelligent User Interfaces workshop for dar-ing to arrange this debate. It was Angel Puer-ta’s careful attention to my concerns thatmade me feel comfortable in coming to speakhere, and so I want to offer him a souvenirfrom our lab—the usual T-shirt. And to JimAlty, a cup of tea for when he needs to relax,from our group at the lab.

I am pleased to represent the University ofMaryland’s Human-Computer InteractionLab, which is now 14 years young. Over theyears we’ve explored a host of user-interfacedesign issues in an interdisciplinary wayinvolving computer science, psychology, andthe library school. Our goal is to create envi-ronments where users comprehend the dis-play, where they feel in control, where thesystem is predictable, and where they are will-ing to take responsibility for their actions. Tome, responsibility will be the central issues inthis debate.

My pitch over 20 years has been to make ascience out of user interface research. I want toget past the notion of “user-friendly,” a vague,and misleading term, and to be really clearabout specifying who the users are and whattheir tasks are. Then we can make a theory

that allows us to predict the time it takes for aspecific user to learn a specific task, the speedof performance on bench-mark tasks, the rateand distribution of errors, and the retention ofthose tasks over time. We look at subjectivesatisfaction as a secondary measure and havedeveloped a standardized Questionnaire ofUser Interaction Satisfaction (QUIS) that theuniversity has licensed to more than 100 orga-nizations around the world. QUIS consists of71 items, from low-level details to higher levelconcepts in the interface (Office of Technolo-gy Liaison, +1-301-405-4210).

We think accommodating individual dif-ferences is important—not just mentioningexperts and novices, but understanding quan-titatively what performance differences weanticipate. Do we expect experts to performtwice as fast or twenty times as fast as novices?Would men perform differently from womenin the use of interfaces, or prefer differentinterfaces? And then we try to deal withbroader cultural issues that are even more dif-ficult to measure.

For me, the future is most clearly movingin the direction of “information visualiza-tion.” I think we would do best to focus onthe remarkable human capabilities in the visu-al domain, which I think are largely underuti-lized by the current designs with 40 icons in2-3 windows. I think we should have two orthree orders of magnitude more: 4,000 ormore items on the screen in an orderly waythat enables people to see all of the possibili-ties and navigate among them.

I will show you three brief videotapedexamples. They all show applications andextensions of the strategy of “direct manipula-tion,” a term I coined in 1982 to describe theexisting successful systems. These systems allhad rapid, incremental, and reversible actions,selection by pointing, and immediate feedback(100-millisecond updates for all actions). Ibelieve that this strategy reduces errors andencourages exploration. The current manifes-tations of direct manipulation are the visualways of searching in databases and on the Webaccompanied by visual presentation of results.

Let’s take a look at an example that goesback 4 years in the first videotape called theFilmFinder.

vsdirect

manipulation

interfaceagents

I think wewould do bestto focus on theremarkablehuman capabilities inthe visualdomain, which Ithink are largelyunderutilized by the currentdesigns with 40 icons in 2-3windows. Ithink we shouldhave two orthree orders ofmagnitudemore: 4,000 ormore items onthe screen in anorderly way thatenables peopleto see all of thepossibilities andnavigate amongthem.

Page 4: BEN SHNEIDERMAN Manipulation - ACS Lab

“Users havegreat controlover the displayand as theyselect items, thedetails appearin windows onthe sides.

45i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

FilmFinder video (see Figures 1a-c, CHI94videotape or HCIL 1995 Video tape reports):This display shows thousands of filmsarranged on the x-axis from 1920 to 1993and on the y-axis from low to high populari-ty. We can use the length slider to trim the setof films by the number of minutes in the filmso we do not have to see films that are too

long, and then we can use the category buttonto show only drama or only action films. Wecan zoom in on more recent pictures and takeonly the more popular ones. And when thereare fewer than 25, the titles will appear auto-matically. When we select one of them, we geta description of the film and informationabout the actress and actor. We can also hunt

d e b a t e

Figure 1(a): FilmFinder

shows 1500 films in a

starfield display where

the location of each

point is determined by

the year of the film (x-

axis) and its popularity

in video store rentals

(y-axis). The color

encodes the film type

(Ahlberg & Shneider-

man, 1994).

ftp://www.cs.umd.edu/pro

jects/hcil/Screen-

dumps/Film/film-

alldots.gif

Figure 1(b): FilmFinder

after zooming in on

recent popular films.

When fewer than 25

films remain, the titles

appear automatically.

ftp://www.cs.umd.edu/pro

jects/hcil/Screen-

dumps/Film/titles.gif

Page 5: BEN SHNEIDERMAN Manipulation - ACS Lab

for films organized by actors. In this case, youmight be interested in Sean Connery, and hisfilms will appear on the screen.

Ben: Okay. I think you get an idea that thecontrols are visually apparent as you dragthem. The updates occur in 100 millisecondsand users get a clear understanding of whatthe situation is. This work goes back to 1993,and the 1994 CHI conference has a couple ofpapers describing it [1, 2]. A general purposeversion of that software was used for theDepartment of Juvenile Justice project, whichyou will hear about shortly.

Here is a FilmFinder done in the UNIXversion of the product called SpotFire (Figure2). Chris Ahlberg has made a commercialproduct out of this and you can downloadthe demo version off of the Web (http://www.ivee.com).

It would be hard to see how to programany kind of agent tool to anticipate all of thepossibilities that your eye would pick up. Weshow the age of the youthful offenders on thex-axis. There are 4,700 of them, from 10 to19 years old. The number of days until a deci-sion was made on their treatment plan isshown on the y-axis. The rules of this organi-zation say that decisions must be made with-in 25 days, but you can see a lot more thanthey anticipated are well above the 25-daylimit.

Interesting things pop up whenever you trya visualization. I hope you will spot these yel-low lines—those were a surprise. We thoughtthere was a bug in the program, but it turnsout that if you start clicking on them to getthe details-on-demand, you’ll find out they alloccur in a Hartford County. They were allbrought in on a certain day. They all have the

46 i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

Figure 1(c): FilmFinder after selecting a single film. The info card pops up with details-on-demand.

ftp://www.cs.umd.edu/projects/hcil/Screen-dumps/Film/film-sean.gif

vsdirect

manipulation

interfaceagents

Page 6: BEN SHNEIDERMAN Manipulation - ACS Lab

47i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

same charge, which is narcotics possession.These were drug busts and they were all puton treatment plans at the same time. Thosekind of patterns happen to anyone who triesvisualization programs, and it would be hardto see how you could program an agent toanticipate all of the possibilities that your eyecan pick up in 1/10th of a second.

Another problem we have dealt with in thesecond videotape has to do with visualizing per-sonal histories. It is called LifeLines (Figure 3).

The placement line shows this youth iscurrently placed at Waxter, a detention center.We also see a placement in a drug-abuse pro-gram and a placement at Cheltenham. Thisone is thicker because the court found himguilty of auto theft. When I click on the lineof the drug-abuse program, all of the relatedinformation is highlighted. I can tell that theplacement was for a breaking and entering

d e b a t e

Figure 2: Spotfire version of FilmFinder provides increased user controls.

Users can set axes (set to length in minues and year) and glyph attributes

(color is set to subject and larger size indicates award winning film). Spot-

fire is available in Unix, Windows and Java versions (http://www.ivee.com)

Figure 3: Youth Record prototype using the Lifelines display to show a case history for the Maryland Department of Juvenile

Justice. (CHI96 videotape or HCIL 1996 Video Tape Reports; 4) ftp://www.cs.umd.edu/projects/hcil/Research/1997/patientrecord.html

Page 7: BEN SHNEIDERMAN Manipulation - ACS Lab

48 i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

case and was requested by Brown. A click onBrown’s name gives the contact information.Here are the reviews which were written. Aclick on the code brings the text of the review.In the same way, I can get more details abouteach case and placement by clicking on thelabels. Those screens are all for the old system,showing that LifeLines can be used as a frontend and acts as a large menu for all of thescreens. The top buttons can access generalinformation, but any critical information willappear on the overview screen. For example,here’s a mention of suicide risk. Seeing theoverview gives the user a better sense of howmuch information is available and what typeof information it is. Of course, this impliesthat all of the information can be presented inone screen.

Ben: We think that LifeLines can also beapplied for medical records, and we are nowapplying it in a project with IBM. We give anoverview of patient histories that contains theconsultations, conditions, documents, hospi-talizations, and medications. Users have greatcontrol over the display and as they selectitems, the details appear in windows on thesides. We think this strategy has great powerin providing convenient access to large andcomplex databases in a way that gives the usersan overview of what is happening and anappreciation of where the details fit into theoverall context. The visual presentation givesusers enormous bandwidth and there arepotentially thousands of selectable items onthe screen at once offering rapid access to theitem that you are seeking.

The third and final example, is a visualdatabase of the human body called the VisibleHuman. Our role was to develop a browsinguser interface to the 15 gigabytes of images atthe National Library of Medicine (3; CHI96videotape and HCIL 1996 Video Reports;free software available for Sun workstationsfrom http://www.nlm.nih.gov and select theVisible Human links till you find our VisibleHuman Explorer).

This direct manipulation interface presentsa thumbnail image of a coronal cross sectionof the body reconstructed directly from theaxial cryosections (Figure 4). This coronal

image acts as an overview, giving a visual rep-resentation of the entire body. The axialcryosections are a local view showing athumbnail corresponding to the slider posi-tion on the overview. We can explore the bodyby simply dragging the slider. It updates inreal time, giving an experience of flyingthrough the human body. Here we see thebrain, the shoulders, the torso, the abdomen,the thighs, the knees, and all of the way downto the toes. We press the retrieve button andthe corresponding full-size image is retrievedover the network from the NLM archive in acouple of minutes. We provide several usefulalternative overviews and also the ability togenerate any coronal section overview, forexample, near the back of the body or near thefront.

Ben: Other labs are working on related ideasof information visualization. From the PacificNorthwest National Labs, this textual data-base that has been presented in a two-dimen-sional mountains-and-clusters visualization togive users an idea of the volume and interrela-tionship of items. Steve Eick at AT&T Labshas these wonderful visual overviews of largetextual documents.

Here, the characters in a children’s book arecolor coded so that you can see the progress ofthe story as it moves on to different characters.Departmental e-mail networks and richerinformation, such as 3D network representa-tions, are part of the things he likes to showwith a variety of user controls to filter the traf-fic and reveal patterns of usage that might bedifficult to see with other data presentationstrategies.

The closing slide says that the overview isthe most important. It gives users a sense ofcontext, of what to look at—the big picture.Then they zoom in on what they want, filterout what they don’t want, and finally go fordetails-on-demand. My claim is that this givesusers the feeling of being in control and there-fore they can be responsible for the decisionsthey make. Thank you.

Jim: Okay, thanks very much, Ben. That wasperfect timing. We now hand it over to Pattie.

Ben:

During the CHI97 debate,

Pattie made a point about

my use of autofocus cam-

eras, suggesting that they

were some sort of agent.

As we were speaking Alan

Wexelblat was taking pic-

tures of the events using

my camera. When I went

through the color slides I

found that most of the

pictures he took were out

of focus! As with most

autofocus cameras, there

is a narrow area in the

center that is used for

focusing and this must be

placed on the intended

subjects of the photo.

Unfortunately Alan didn’t

know this and simply

pointed at the stage area

but the focus point was

on the background poster.

Almost all the photos

were unusable.

We can all find this amus-

ing and leave it at that,

but I think there is a seri-

ous point which is that

agents don’t always do

what we expect them to

do, and it takes some

knowledge to make

effective use of agents (or

auto-focus cameras). If we

were to assess responsi-

bility — I would take part

of it in failing to give ade-

quate instruction to Alan,

he might take part

because he was the direct

user, and maybe the

manufacturer has another

part for a poor design

which fails to provide

appropriate feedback.

Page 8: BEN SHNEIDERMAN Manipulation - ACS Lab

49i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

Pattie: I’m not going to bribe the moderatorwith tea or T-shirts or anything. I hope thatthe work will speak for itself. The word “agent”is used in a lot of different ways. I want to startthis presentation by explaining what I mean bythe word “agent,” and in a particular, “softwareagent.” Basically, software agents are a newapproach to user software, a new way of think-ing about software that end-users have to use.In particular, the way in which agents differfrom the software that we use today is that asoftware agent is personalized.

A software agent knows the individual user’shabits, preferences, and interests. Second, asoftware agent is proactive. It can take initia-

tive because it knows what your interests are. Itcan, for example, tell you about somethingthat you may want to know about based on thefact that you have particular interests. Currentsoftware, again, is not at all proactive. It does-n’t take any initiative. All of the initiative has tocome from the user. A third difference withcurrent software is that software agents aremore long-lived. They keep running, and theycan run autonomously while the user goesabout and does other things. Finally, softwareagents are adaptive in that they track the user’sinterests as they change over time.

So, you can ask, well, why do you call it anagent? Why call it an agent given that the term

d e b a t e

Figure 4: Visible Human Explorer user interface, showing a reconstructed coronal section overview

(on the left) and an axial preview image of the upper abdominal region (on the upper right). Drag-

ging the sliders animates the cross-sections through the body (North et al., 1996).

ftp://www.cs.umd.edu/projects/hcil/Research/1995/visible-human.html

vsdirect

manipulation

interfaceagents

Page 9: BEN SHNEIDERMAN Manipulation - ACS Lab

50 i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

is already so overloaded, and given the factthat it’s really software that is slightly differentfrom existing software? Well, we call it anagent to emphasize the fact that agent soft-ware can act on your behalf while you aredoing other things. We also want to empha-size that it does this based on its knowledge ofyour preferences, just like a travel agent willact on your behalf by buying you a travel tick-et based on the information that the travelagent has about your preferences. Note that Iprefer not to use the term “intelligent agents”nor “autonomous agents” because those termshave even more problems associated withthem.

Now, why do we need software agents?Why does our software need to become morepersonalized? Why does our software need totake initiative to help us as a user? This needsto happen because our current computer envi-ronment is getting more and more complex,and the users are becoming more and morenaive. Finally, the number of tasks to take careof, and the number of issues to keep track of,are continuously increasing. Let me tell youmore about this. First of all, the nature of ourcomputer environment is radically differenttoday from 20 years ago, back when the cur-rent style of computer interaction was invent-ed. Twenty years ago, one typically had oneuser using one computer, and everything inthat computer, every file, every object, was ina particular place because the user put it there.There was a limit to the amount of informa-tion on that computer. It was completely stat-ic. Nothing changed unless the user made itchange. It was completely structured and wellorganized. Today, our computer environmentsare completely different.

More and more the World Wide Web andour browser is becoming the one and onlyinterface. It’s not quite the case yet today, butit will be a year from now. In that situation,our computer is no longer this closed environ-ment that we have complete control over.Instead, our computer or the screen is a win-dow onto this vast network, this vast networkof information and other people. That net-work is continuously changing. It is dynamic.Something may be in one place today, and thenext day it may be in another place or may be

gone. Continuously, new information is beingcreated, new versions of software are beingadded. Also, that environment is completelyunstructured. Take a look at the World WideWeb, for example. You couldn’t possibly try tovisualize the World Wide Web in any waybecause it is completely unstructured andbecause it has been built by so many differentpeople and is continuously changing. I believethat the dominant metaphor that we havetoday is a mismatch for the computer envi-ronment we are dealing with tomorrow.

Second, the user 20 years ago was differentfrom the typical user today. Twenty years agowe mostly dealt with professional users ofcomputers. Today and tomorrow the con-sumer electronics market is going to be theone that dominates, and those users do noteven know how to program their VCRs. Howare they going to deal with user interfaces?

Third, the number of things that peoplehave to keep track of and the number of tasksthat they use their computers for is huge andis increasing all of the time. As we know fromother domains, whenever workload or infor-mation load gets too high, there is a pointwhere a person has to delegate. There is noother solution than to delegate. For example,many of you may have students that you del-egate certain tasks to, or you may have per-sonal assistants that you delegate certain tasksto, not because you can’t deal with those tasksyourself, but because you are overloaded withwork and information. I think the same willhappen with our computer environments:that they become just so complex and we usethem for so many different things that weneed to be able to delegate.

We need to be able to delegate to whatcould be thought of as “extra eyes or extraears” that are on the lookout for things thatyou may be interested in. We also need “extrahands or extra brains,” so to speak, becausethere will be tasks that we just cannot dealwith because of our limited attention span orlimited time, and we need other entities to beable to represent us and act on our behalf.Some examples to make this more concrete (Ididn’t bring any videos because of the limitedamount of time), but most of you have seen atleast one of these agents. These are some of

“Why do we needsoftware agents?… Take a look atthe World WideWeb, for exam-ple. You couldn’tpossibly try tovisualize theWorld Wide Webin any waybecause it iscompletelyunstructuredand because ithas been builtby so many different peopleand is continu-ously changing.I believe thatthe dominantmetaphor thatwe have today isa mismatch forthe computerenvironment weare dealing withtomorrow.

vsdirect

manipulation

interfaceagents

Page 10: BEN SHNEIDERMAN Manipulation - ACS Lab

“We need to beable to delegateto what couldbe thought ofas “extra eyesor extra ears”that are on thelookout forthings that you may beinterested in.

51i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

the ones we built in our lab. Letizia, built byHenry Lieberman, who is here at the confer-ence, is an agent which continuously watchesyou as you browse the Web, analyzing andmemorizing all of the pages that you visit. Itextracts from those pages the common key-words.

Whenever you are using your Web brows-er, Letizia always looks ahead and checkswhether within a certain depth of the currentpage, there happen to be any pages that youmay be interested in. So, for example, if I aminterested in scuba diving, my agent may havepicked it up because I look at a lot of pagesabout scuba diving. If I go to a particularentertainment site, it may look ahead and say,hey did you realize that if you follow that linkthat there are some pages about scuba divingin the Florida area? The Remembrance Agent isanother agent that continuously tracks thebehavior of the user. It helps you remembercertain things. It helps you remember who

sent e-mail or whether you already replied toa certain e-mail message. It may proactivelyremind you of information related to theinformation you are currently looking at. Itworks in EMACS. When I am, for example,looking at an e-mail message from a particularperson, it proactively reminds me of the previ-ous e-mail messages from that same person,which is very useful because I may have for-gotten to reply to one of them.

Firefly, some of you may have tried thatagent, is basically a personal filterer for enter-tainment, not unlike the movie applicationthat Ben talked about, except that this agentwill again keep track of your interests, yourpreferences, and proactively tell you aboutnew movies that you may be interested in whichyou even forgot to ask about in the first place.Yenta is another agent that we built whichtracks what the user’s interests are by lookingat your e-mail and files and extracting key-words. It talks to other Yenta agents belonging

d e b a t e

Page 11: BEN SHNEIDERMAN Manipulation - ACS Lab

52 i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

to other users, and if it notices that anotheruser shares some of your interests, especially ifthose interests are very rare, then it introducesyou to that other user. It may say “hey did yourealize that at this IUI conference there isanother person who is interested in goingscuba diving in Florida” so that maybe thenwe can decide to go scuba diving together.Again, it’s suggesting something that you

wouldn’t have thought of yourself. Kasbah isanother set of agents that buy and sell onbehalf of users. We are currently setting upthis experiment MIT-wide, meaning for15,000 people. We have already done testswith 200 people. It’s basically a marketplacewhere you can create an agent who will buy orsell a second hand book or a second handmusic CD for you. You just tell the agent, “Iwant to sell The Joshua Tree by U2. I want toask $12 for it at first. You are allowed to go aslow as $9. You have two months to sell thisCD. You should be really tough and onlychange the price all the way at the end, nearwhen the 2 months are over.” That agent willrepresent you in that marketplace, negotiatingon your behalf with other people or otheragents who may be interested in buying thatCD from you. Again, it is sort of acting onyour behalf. You don’t have to waste any timetrying to make 10 bucks, but the agent will dothis for you.

I think it’s important to address some com-mon misconceptions about agents: First of all,sorry to say so, but agents are not an alterna-

vsdirect

manipulation

interfaceagents

Page 12: BEN SHNEIDERMAN Manipulation - ACS Lab

53i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

tive for direct manipulation. A lot of confer-ences and magazines pitch agents againstdirect manipulation. They are actually com-plementary metaphors. Whenever you havean agent interface, typically you also need avery good application interface because anagent is not a substitute for an interface. Anagent basically interacts with the applicationjust like you interact with the application. It’sas if you had someone else looking over yourshoulder as you are using the application,noticing some of your preferences and habitsand then offering to automate some of thetasks for you. So you still need a very gooddirect manipulation interface—visualiza-tion—all of these wonderful tools so that theuser can personally interact with the applica-tion. An agent can never predict all of themovies that I may possibly be interested in. Itmay be able to make some interesting sugges-tions to me, but I will still need to look upparticular movies myself.

A second misconception is that some peo-ple think that agents are necessarily personi-fied or anthropomorphized. In fact, mostagents are not. Most of them don’t even dealwith a natural language interaction interface.

A third misconception is that agents necessar-ily rely on traditional AI (artificial intelligence)techniques, like knowledge representation andinferencing. In fact, most of the agents that arecommercially available and have proven success-ful with large numbers of users rely on eitheruser programming or on machine learningrather than traditional AI techniques.

I want to conclude by addressing some crit-icisms of software agents that Ben has comeup with—as well as people like Jaron Lanier(see “A Conversation with Jaron Lanier,”interactions ii.3 (July 1995), pp. 46-65), whois also very vocal about all of this. Opponentsof agents typically argue that well-designedvisualization interfaces are better. Like I saidbefore, you still need a well-designed interfacewhen incorporating agents in an application.However, some tasks I may just not want todo myself even if the interface was perfect. Ifmy car had a perfect interface for fixing theengine, I still wouldn’t fix it. I just don’t wantto bother with fixing cars. I want someone elseto do it.

d e b a t e

Pattie Maes is an Asso-

ciate Professor at MIT’s

Media Laboratory,

where she founded and

directs the Software

Agents Group. She cur-

rently holds the Sony Corporation Career Development

Chair. Previously, she was a visiting Professor and a

Research Scientist at the MIT Artificial Intelligence Labora-

tory. She holds a Bachelor’s degree and Ph.D. degree in

Computer Science from the Vrije Universiteit Brussel in Bel-

gium. Her areas of expertise are Human Computer Interac-

tion, Electronic Publishing and Electronic Commerce,

Artificial Intelligence, and Artificial Life.

Pattie is one of the pioneers of a new research area

called Software Agents, that is, semi-intelligent computer

programs which assist a user with the overload of infor-

mation and the complexity of the online world. She is one

of the organizers for the leading conferences in this area

such as the annual Autonomous Agents conference and

the annual Practical Applications of Agents and Multi-

Agent Systems conference. She is a founder and board

member of the Agent Society, an international industry

and professional organization established to assist in the

widespread development and emergence of intelligent

agent technologies and markets.

Pattie is a consultant in the area of Software Agents for

several major companies such as Apple Computer, Hughes

Research, etc. She is the editor of three books and is an

editorial board member and reviewer for numerous profes-

sional journals and conferences, such as the User Modeling

journal, the Personal Computing journal and the Artificial

Life journal. Her work has achieved several prizes, includ-

ing the IBM best Bachelor’s thesis award (1984), the OOP-

SLA-1987 best paper award, the Ars Electronica 1995

award, the Interactive Media Festival Award 1995, the

ArcTec 1995 award and so on.

Finally, she is a founder of Firefly Network, Inc. in

Boston, Massachusetts—one of the first companies to com-

mercialize software agent technology.

Page 13: BEN SHNEIDERMAN Manipulation - ACS Lab

“Opponents ofagents typicallyargue that well-designed visual-izationinterfaces arebetter. Like Isaid before, youstill need a well-designed inter-face whenincorporatingagents in anapplication.However, sometasks I may justnot want to domyself even ifthe interfacewas perfect. Ifmy car had aperfect interfacefor fixing theengine, I stillwouldn’t fix it. I just don’t wantto bother withfixing cars. Iwant someoneelse to do it.

54 i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

A second criticism is that agents make theuser dumb. That’s actually more one of Jaron’sobjections rather than Ben’s, I think. To someextent it’s true. If I don’t fix my car then I’m notgoing to learn about fixing cars. However, thisdoes not constitute a problem. As long as there’salways an agent available or I can call one by amotor association like AAA, then that’s fine. It’stoo bad that I will never learn about cars, but Iwant to learn about other things instead.

A third criticism expressed is that usingagents implies giving up all control. That’s incor-rect. I think you do give up some control whenyou deal with an agent. I tell the car mechanic tofix my car or to fix this or that part of the car. Idon’t know how exactly he or she is going to dothat. I don’t mind giving up some control, actu-ally, and giving up control over the details as longas the job is done in a more-or-less satisfactoryway, and it saves me a lot of time.

Okay, just very briefly, I want to say that Ithink where the true challenge lies is indesigning the right user-agent interface. Inparticular, we need to take care of these twoissues: understanding and control. Under-standing means that the agent-user collabora-tion can only be successful if the user canunderstand and trust the agent, and controlmeans that users must be able to turn overcontrol of tasks to agents but users must neverfeel out of control. I believe that this is a won-derful interface-design challenge, and we havecome up with a lot of solutions to actuallymake sure that the agent’s user interface hasthese two properties that the user feels in con-trol or has control when he or she wants it, aswell as that the user understands what theagent does and what its limitations are. Let mesave that for later, maybe. Thanks.

Jim: Thanks very much. Okay, Ben’s going togo up to 5 minutes to say whatever he likes.

Ben: How interesting. We are debating, butpart of me is drawn to the idea of celebratingPattie Maes and encouraging you to followher example. I want to draw the audience’sattention to her transformation during themonths we’ve had these discussions. As I goback to Pattie Maes’s work and I read her ear-lier papers and her Web sites, she promotes

autonomous agents and presents an anthropo-morphic vision. Even in the current proceed-ings her article is titled “Intelligent Software,”so I was delighted with her opening remarksthat rejected intelligent and anthropomorphicdesigns. The old Pattie Maes wrote “agentswill appear as living entities on the screen,conveying their current state of behavior withanimated facial expression or body languagerather than windows text, graphics, and fig-ures.” So we’ve got two Pattie Maes. I willchoose the newer one that demonstratesmovement in my direction including her lastslide which might have been written by me:“User understanding is central, and user con-trol is vital for people to be successful.”

In fact, I have other ways of celebratingPattie Maes. I encourage you to look at herFirefly Web site, which is an interesting appli-cation. Collaborative filtering, I think, willbecome an important approach for manydomains. But as a user, I can’t find the agentson the Firefly Web site. In fact, as I searchedto find the agents, all I came up with was thatthe company had previously been calledAgents, Inc. and is now called Firefly. If youread the Firefly Web site, you will not find theword “agents” in the description of this sys-tem. In fact, the interface is a quite lovely,direct manipulation environment, allowingusers to make choices by clicking in a verydirect manipulation way.

So, I think we’ve made progress in clarify-ing the issues in the past year of our ongoingdiscussions. For example I think we can sepa-rate out the issue of natural language interac-tion, which as far as I can see, has not been asuccess. The systems that were offered com-mercially even a few years ago, like Q&A fromSymantec or Intellect from AI Corporation todo database query, and Lotus HAL for spread-sheets, are gone, and direct manipulation isthe surviving technology.

A second issue is anthropomorphic inter-faces such as chatty bank tellers and the PostalBuddy or the proposed Knowledge Navigatorof Apple’s 1988 video. Microsoft’s playfulattempt at a social interface in BOB is also afailed product. As far as I can see the anthro-pomorphic or social interface is not to be thefuture of computing.

vsdirect

manipulation

interfaceagents

Page 14: BEN SHNEIDERMAN Manipulation - ACS Lab

55i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

A third issue of adaptive interfaces is quiteinteresting. I would concede half a point andsay that we now see two levels: the user inter-face level, which users want to be predictableand controllable, as Pattie has stated, and thelevel below the table, where there may besome interesting algorithms such as collabora-tive filtering. If those can be adaptive theremay be benefits in the way that Pattiedescribes. This is related to other adaptationssuch as when I save a file to disk, I see it saved,and it is retrievable by me. Under the table,there’s a great deal of adaptation dealing withspace allocation, disk fragmentation, andcompression strategies, but from the user’spoint of view, there’s no adaptation. It’s quitepredictable. The same goes for engines in

d e b a t e

Page 15: BEN SHNEIDERMAN Manipulation - ACS Lab

56 i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

automobiles: I turn the key, I step on the gas,it goes. Underneath the hood, below the user’sconcern, there are adaptive algorithms thatwill set the engine speed based on many fac-tors, such as the temperature, gas mixture, etc.That level of adaptivity is important as long asit does not interfere with the user’s predictionof behavior.

I am concerned about the anthropomor-phic representation: it misleads the designers,it deceives the users; it increases anxiety aboutcomputer usage, interferes with predictability,reduces user control, and undermines users’responsibility—which I think is central. Ithink anthropomorphic representationsdestroy the users’ sense of accomplishment; Ithink users want to have the feeling they didthe job—not some magical agent.

Finally I am concerned about the confu-sion of human and machine capabilities. Imake the basic assertion that people are notmachines and machines are not people. I donot think that human-to-human interaction isa good model for the design of user interfaces.

Jim: Okay, Pattie, would you like to respondto that?

Pattie: First of all, I should clarify thatautonomous agents or the word “agents” has amuch broader meaning than the words “soft-ware agent,” and my group at the Media Labdoes research on autonomous agents moregenerally, as well as software agents. So whenBen was quoting from our Web site, he’s actu-ally quoting other work that we do, for exam-ple, work on synthetic characters that caninteract with people in a virtual environment,which doesn’t have anything to do with thesoftware agent’s work. In fact, it has less andless to do with it than it may have at onepoint. So, if you go to autonomous agent con-ferences, for example, the Agents Conferencein Marina del Rey in February (First Interna-tional Conference on Autonomous Agents,see interactions iii.6, Conference Preview),you’ll see work being discussed that relates torobots, autonomous robots. You’ll see workabout synthetic believable characters, andyou’ll see work about software agents. So that’sone thing I wanted to respond to. It’s impor-

tant to distinguish these different types ofagents and not lump them all together.

Second, I absolutely agree with Ben that sofar the most successful interfaces are the oneswhere the agents are pretty much invisible.They are not visualized as a little anthropo-morphic character. For example, in Firefly,that is the case. That doesn’t mean that thereisn’t an agent operating. For example, in Fire-fly, the Firefly agent will proactively tell youabout other users that you may want to talkto. It will warn you when there is somethingthat has changed somewhere that you may beinterested in. There is still an agent theremonitoring all of your preferences and proac-tively making recommendations to you, butthat doesn’t mean that there has to be this lit-tle cute character on the screen.

Now, I think one of the reasons that Benand I disagree is actually that we are focusingon completely different problem domains. Inpretty much all of the problem domains thatBen looks at we are dealing with a user who isa professional user, and we are dealing with atask domain that is very well structured and aninformation domain that is very well orga-nized, so that it lends itself to visualizing all ofthe different dimensions. The kind of prob-lems that we have typically been dealing withare very different because we are dealing withend users who are not necessarily very trained.They may use the Web for a couple of hoursper week, but that is about it. We are dealingwith a very different information domain, aninformation domain that may be very ill struc-tured and very dynamic. For example, theWorld Wide Web is actually sort of one of thekey domains that we do all of our research on.

Finally, to illustrate that these approachesaren’t necessarily incompatible, I could envi-sion a version of Ben’s movie finder which usesBen’s nice visualization interface, where anagent is continuously monitoring what moviesyou seem to be interested in. That agent may,for example, highlight particular areas in theinterface which it thinks you will be specifi-cally interested in. That kind of interfacewould actually combine an agent that learnsabout your preferences and proactively makessuggestions to you, with a nice visualizationinterface. The reason why you want that kind

“I am concernedabout the confusion ofhuman andmachine capabilities. Imake the basicassertion thatpeople are notmachines andmachines arenot people. I donot think thathuman-to-humaninteraction is agood model forthe design ofuser interfaces.

vsdirect

manipulation

interfaceagents

Page 16: BEN SHNEIDERMAN Manipulation - ACS Lab

57i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

of proactive software is that the user does notnecessarily always want to have all of that con-trol when searching for a movie. I believe thatusers sometimes want to be couch-potatoesand wait for an agent to suggest a movie tothem to look at, rather than using 4,000 slid-ers, or however many it is, to come up with amovie that they may want to see.

Question: Okay, I have a remark to Ben, whichthen transitions into a question for both speakers.Ben, I was a little bit irritated, I think Pattie wastoo, with your lovely presentation on informationvisualization, which seems to be entirely besidethe point. As Pattie said, we will take the bestdirect manipulation and visualization as we canpossibly get. It seems to me the contribution to thisdiscussion would be some negative examples ofwhere agentlike things are bad. At some level, Itake the thrust of your position, Ben, as beingreactionary, to put it in simpler words, sort offear-driven. I would like to test my theory by ask-ing the following question, which is, we have nowa new medium on the interface of playing whichis speech. Speech is now practical. IBM makesspeech systems that are being used and so on.

My question actually to both of you is to seewhat your reactions are to this new technology. Ipredict that if you are going to have speech witha computer, first of all, research at Stanfordrecently has shown that once you have a comput-er talking you cannot prevent people fromanthropomorphizing the computer. I do not seehow you are going to have a coherent speechinterface without using human communicationprinciples. So I predict that you will say just don’tdo it. I also want to hear Pattie‘s commentsabout speech technology.

Ben: Do we have another half an hour here? Ithought I said very positive things about Fire-fly and its agent and adaptation, and I cer-tainly like to see automaticity built intointerfaces that amplify the user’s capabilities. Ihave trouble with the words like “agents,” and“expert,” and “smart,” and “intelligent”because, they mislead the designer, anddesigners wind-up leaving out importantthings. In fact, I love Pattie’s slide up here. Ifthe agent-oriented community would adopt

d e b a t e

Ben Shneiderman is

a Professor in the

Department of

Computer Science,

Head of the Human-

Computer Interac-

tion Laboratory,

and Member of the Institute for Systems Research, all

at the University of Maryland at College Park. He has

taught previously at the State University of New York

and at Indiana University. He regularly teaches popu-

lar short courses and organizes an annual satellite

television presentation on User Interface Strategies

seen by thousands of professionals since 1987. The

third edition of his book Designing the User Inter-

face:tion (1987) has recently been published by Addi-

son-Wesley Longman.

Ben is the author of Software Psychology: Human

Factors in Computer and Information Systems (1980)

Addison-Wesley Longman, Reading, MA and his 1989

book, co-authored with Greg Kearsley, Hypertext

Hands-On!, contains a hypertext version on two disks.

He is the originator of the Hyperties hypermedia sys-

tem, now produced by Cognetics Corp., Princeton

Junction, NJ. In addition he has co-authored two text-

books, edited three technical books, published more

than 190 technical papers and book chapters. His 1993

edited book Sparks of Innovation in Human-Computer

Interaction collects 25 papers from the past 10 years

of research at the University of Maryland.

Ben has been on the Editorial Advisory Boards of

nine journals including the ACM Transactions on Com-

puter- Human Interaction and the ACM interactions.

He edits the Ablex Publishing Co. book series on

“Human-Computer Interaction.” He has consulted and

lectured for many organizations including Apple,

AT&T, Citicorp, GE, Honeywell, IBM, Intel, Library of

Congress, NASA, NCR, and university research groups.

Page 17: BEN SHNEIDERMAN Manipulation - ACS Lab

58 i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

these principles it would go a long way inmaking me sympathetic. For example Pattiewrites “Make the user model available to theuser.” I don’t see that being done in most ofthe work about agents. Explanations shouldbe available and methods of operation shouldbe understandable to the user. So much of thework in agentry goes against the principles. Ilike the new Pattie. I am ready to be partnersand collaborate with the new Pattie.

Now, to focus on speech. We have heardfor 25 years the great hope and dream thatspeech is going to solve our user interfaceproblems. Dreamers prophecy that the StarTrek scenario is going to take over and we willtalk to our computers. I do not believe thatspeech will be a generally usable tool. It hasimportant niches: opportunities for disabledusers, for certain hands-busy, eyes-busy, andmobility-required applications. In preparingthe third edition of my book I worked hard tofind speech applications that do recognitioneffectively. I am quite happy with speechstore-and-forward applications by telephone,but the recognition paradigm is not beingwidely accepted, even for minor tasks such asvoice dialing. Speech output, except by tele-phone, is also a problem because speech isvery slow and disruptive of cognitive process-ing. I think what annoys me the most aboutthe devotees of speech, is their failure to takein the scientific evidence that speaking com-mands is cognitively more demanding thanpointing. Speech uses your short-term memo-ry and working memory. By contrast, hand-eye coordination can be conducted in parallelwith problem solving by another part of yourbrain and therefore does not degrade your per-formance as much as speaking.

Question: You can do both at the same time.

Ben: Yes, you can do hand-eye tasks in paral-lel with problem solving, more easily that youcan speak while problem solving. This fact isnot a barrier to use of speech, but it is a hur-dle that designers of speech systems must rec-ognize if they are to find ways to overcome it.Pattie: I must admit, I actually agree with a lotof what Ben says. I haven’t used speech at all

in my research, the main reason being com-pletely personal—that these systems oftendon’t understand my accent, but apart fromthat, I do agree that it is not a very high band-width kind of connection. There is also a lotof ambiguity.

So I personally would like to see speechbeing used just for situations where the handsare not available, like in your car or as an addi-tional channel, actually, for example, giving anagent some additional advice while you arealso pointing at something. For example,Henry Lieberman, sitting here in the audi-ence, did some interesting work where hetaught an agent a particular procedure, andwhile he was performing actions with themouse, he would give speech inputs to tell theagent what it had to pay attention to. Forexample, pay attention to this corner here thatI am dragging or to this side of the rectangle.So in that situation it is very useful becauseyour hands are already doing something else,and you need that additional channel to con-vey some more information in parallel.

Question: This question is for both of you. Bothof you seem to be concerned about protecting theuser’s control of the environment, but the onethings studies have shown time and time again isthat users are very good at making mistakes. Sohow do your positions relate to time-critical deci-sion-support environments, such as medical sys-tems or cockpit systems?

Pattie: I have actually been focusing on acompletely different kind of application, atype of application that is not as critical. Forexample, if your World Wide Web agent givesyou a wrong Web page to look at—it assumesthat you are interested in a Web page and youare not—that is not at all critical. It is not abig deal. I have been focusing on that kind ofsituation and those kinds of problems, theones where if there is an error it is not verycostly. I’ve been doing that because I believethat it will be very hard to make agents thatalways come up with the right answer, alwaysdo the right thing. I believe that there is a verylarge set of these kind of applications wherethings don’t have to be completely precise or

“User interfacesshould be predictable, sothat users trustthem. User interfacesshould be thor-oughly tested,and users shouldbe thoroughlytrained for allemergencies. In emergencysituations, people cannotsolve problems.They can onlydo what’srehearsed andpredictable.

Page 18: BEN SHNEIDERMAN Manipulation - ACS Lab

59i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

100 percent correct in order for the agent tobe very, very useful to a user.

Ben: I like your question. I think it is anextremely important research area. There is along history of work often called supervisorycontrol. Tom Sheridan is a key player in thearea for nuclear-reactor control rooms, cock-pits, and so on. I think the design of these sys-tems is most effective when the users have aclear predictive model of what their actionswill produce. If they do not know or areuncertain about what the results of theiractions are, they will disengage the automaticsystem, as is the evidence with cockpit systemsor nuclear-control rooms. So one danger isthat in complex control-room environmentswhen an emergency occurs, users are uncer-tain about its behavior. In these situationsthey are likely to disengage a potentially help-ful system and do what makes sense to them.Therefore, as Jim Foley said, keep it simple,very simple.

User interfaces should be predictable, so thatusers trust them. User interfaces should bethoroughly tested, and users should be thor-oughly trained for all emergencies. In emer-gency situations, people cannot solve problems.They can only do what’s rehearsed and pre-dictable. It is a good topic that I would love tosee more attention to it by this community.

Question: In the interest of brevity, I was goingto bring up several human limitations or con-straints on humans that make direct manipula-tion a little more interesting and wanted to askyou both for comments, not that I have all of theanswers, but I think I would just limit it to yourbasic kind of law that the more things you haveto scroll through or the more things you have tosearch through, the longer it takes you to search.The idea of how do you deal with this, particu-larly, for many people scrolling is not a particu-larly usable thing. I have particularly seen thiswith older people, and of course, if you want agood example just go to a fast food restaurant andwatch a new person at the cash register try to findhow to ring up your hamburger. It is pretty ter-rible. You will be there for days.

That kind of idea, and particularly—we are

talking about the wonderful visual processingand how important visualization is—this seemsto be assuming that everybody has perfect vision.What if I am blind, if I am over sixty-five, andI have a very small useful field of view and Idon’t notice things so much on the periphery, or ifI am, in my case, a person sitting right here whois having a problem with my contact lenses, howcould you, Ben you were saying—well speech isokay for disabled users—how are you going torender that diagram into speech for a blind per-son, and what is the role of that?

Ben: That is a legitimate concern. Directmanipulation does benefit from and dependsheavily on visual representations. For thosewho are vision challenged or blind, alterna-tives to visual displays are important. Whatsurprises me is there are great supporters ofdirect manipulation in the visually-challengedcommunity, because direct manipulationdepends on spatial relationships. Blind usersoften are strong at spatial processing. If youcan provide movement left, down, backward,forward, they can navigate fairly rich spaces inefficient ways.

I would say also you have been a little tooquick about criticizing menu selection. Thequestion is what would the alternative be, andhow might those menus be better designed? Ido believe that fast and vast menus are a greatbenefit in many applications.

Jim: Okay. There are two more questions ifyou could be very brief, please.

Question: This is directed more toward Benthan toward Pattie. Will this debate betweendirect manipulation and agency always exist ininterface design or will it eventually be replacedby some kind of fusion of the two approaches? Inother words, are we going to see new Bens andnew Patties every day, or is there going to be somekind of “Shneider-Maes”?

Ben: I think it has been interesting to see howthe debates evolve. I certainly will point youto the new Pattie, whom I am ready to cele-brate and be partners with, as I said, but Ithink the debate will move on. I think it has

d e b a t e

I have beenfocusing on thatkind of situationand those kindsof problems, theones where ifthere is an errorit is not verycostly. I’ve beendoing thatbecause Ibelieve that itwill be veryhard to makeagents thatalways come upwith the rightanswer, alwaysdo the rightthing. I believethat there is avery large set ofthese kind ofapplicationswhere thingsdon’t have to becompletely precise or 100percent correctin order for theagent to bevery, very usefulto a user.

vsdirect

manipulation

interfaceagents

Page 19: BEN SHNEIDERMAN Manipulation - ACS Lab

60 i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

matured in interesting ways from where wewere a year ago. I think we are all at the edgeof looking at new interfaces, and so, as wepush that envelope back, we are getting betterunderstanding of the territory, of the strengthsand weakness of direct manipulation, of thestrengths and weaknesses of agents, and wherethey are appropriate. I am pleased by theprogress in the discussion.

Pattie: I think we both have changed. Wouldyou agree to that or not?

Jim: I think before they start kissing let’s moveon. Last question.

Question: This starts out at least as a clarifica-tion question for Dr. Shneiderman, but it maygo places from there. To what extent does directmanipulation, in your definition and in yourview, admit autonomous system behavior?Because it seems to me that as soon as you admitanything unexpected, uncontrolled, potentiallyanthropomorphizable out of the system that youare interacting with, you have opened the door,you have taken a step down the slippery slopetoward agentness.

Ben: Yeah. I am in favor of increased automa-tion that amplifies the productivity of usersand gives them increased capabilities in carry-ing out their tasks, while preserving their senseof control and their responsibility, responsibil-ity, responsibility. I am sort of not answeringyour question because I don’t want to work inthe language you are dealing with. We shouldbe thinking about productivity improvementtools for users, whether they are graphicalmacros, dynamic queries, starfield displays orother things.

Question: Then what is it about agents that youdislike?

Ben: Can I go to my closing slide? I want toreassert the importance of scientific evalua-tions. We must get past the argumentationabout my system being more friendly thanyours or more natural or intuitive, and talk

about user performance. We can deal with sat-isfaction also, but please focus on user perfor-mance and realistic tasks. Please, please, pleasedo your studies—whether they are controlledscientific experiments, usability studies, orsimply observations, and get past the wishfulthinking and be a scientist and report on realusers doing real tasks with these systems. Thatis my number one take-away message.

I am here to promote direct manipulationwith comprehensible, predictable, and con-trollable actions. Direct manipulation designspromote rapid learning. It supports rapid per-formance and low error rates while supportingexploratory usage in positive ways.

Direct manipulation is a youthful conceptwhich is still emerging in wonderful ways.Our current work leans to information visual-ization with dynamic queries, but there arepeople doing fascinating things with enrichedcontrol panels style sheets, and end-user pro-gramming. Graphical macros would be myfavorite project to advance the design of gen-eral computing tools. It is embarrassing thatafter 15 years of graphic user interface beingwidely available, we have no graphical macrostools. What is going on? This is the greatestopportunity for visual programming.

Third—and I am answering your questionhere—I think the intelligent agent notionlimits the imagination of the designer, and itavoids dealing with interface issues. That’s myview of the agent literature—there is insuffi-cient attention to the interface. Maybe theway agents will mature is as Pattie is suggest-ing; that the agents take care of the processesbelow the table, and there is a nice directmanipulation interface that the user sees.

A leading AI researcher commented to methat the 30 years of planning work in AI isessentially down the tubes because of lack ofattention to the user interface. The designersdeliver a system and the first thing that theusers say is, “This is great but what we reallywant to do is change these parameters.” Thedesigners say, “Well, you know, we didn’t putthat in the interface.” They just haven’tthought adequately about the interface, nordone testing early enough.

I believe that this language of “intelligent,autonomous agents” undermines human

“I am here topromote directmanipulationwith compre-hensible, pre-dictable, andcontrollableactions. Directmanipulationdesigns promoterapid learning.It supports rapidperformance andlow error rateswhile supportingexploratoryusage in positiveways.

vsdirect

manipulation

interfaceagents

Page 20: BEN SHNEIDERMAN Manipulation - ACS Lab

61i n t e r a c t i o n s . . . n o v e m b e r + d e c e m b e r 1 9 9 7

responsibility. I can show you numerous arti-cles in the popular press which suggest thecomputer is the active and responsible party.We need to clarify that either programmers oroperators are the cause of computer failures.Agent promoters might shift some attentionto showing users what is happening so thatthey can monitor and supervise the perfor-mance of agents. I was disturbed that in theAutonomous Agents conference that Pattie isparticipating in, the organizers refused toinclude the topics of supervision of agents anduser interfaces for programming agents. Bycontrast, I like Pattie’s summary slide—I thinkher list is quite wonderful.

My closing comment is that I think thereare exciting opportunities in these visual inter-faces that give users greater control and there-fore greater responsibility in the operation ofcomputers.

Jim: Thanks. Okay, Pattie?

Pattie: I want to conclude by saying that Ibelieve that there are real limits to what wecan do with visualization and direct manipu-lation because our computer environments arebecoming more and more complex. We can-not just add more and more sliders and but-tons. Also, there are limitations because theusers are not computer-trained. So, I believethat we will have to, to some extent, delegatecertain tasks or certain parts of tasks to agentsthat can act on our behalf or that can at leastmake suggestions to us.

However, this is completely a complemen-tary technique to well-designed interfaces—visualization, and direct manipulation—not areplacement. Users still need to be able tobypass the agent if they want to do that. AlsoI should say that what we have learned thehard way really is that we have to, whendesigning an agent, pay attention to user-interface issues, such as understanding andcontrol. These are really very, very importantif you want to build agents that really workand that users can trust. The user has to beable to understand what the agent does.

The user has to be able to control things ifthey desire or to the extent that they want tocontrol things. I agree with Ben that the agent

field definitely has grown a lot in the past 10years or so. In fact, one of the ways I think inwhich a lot of this agent work distinguishesitself from traditional AI is that agent researchfocuses on building complete systems, systemsthat are tested, systems that really have towork, and those same principles and method-ologies can be seen in all of the agents work,whether it be robots, synthetic characters, orsoftware agents.

The field is maturing and paying moreattention to building things that really workand paying attention to important UI issues.As to people taking responsibility for theiragents, I think they indeed should. It is soft-ware that is running on your behalf and thatyou have delegated certain tasks to. So, per-sonally, I don’t see why that problem is specif-ic to agents as opposed to software in general.Thanks.

Jim: Okay, I would like to thank the Shnei-derman–Maes team for coming here todayand talking to us, and thank you all for par-ticipating.

References [1] Ahlberg, C. and Shneiderman, B., Visual Informa-

tion Seeking: Tight coupling of dynamic query filters

with starfield displays, Proceedings Of ACM CHI94

Conference (April 1994), 313-317 + color plates.

[2] Ahlberg, C. and Shneiderman, B., AlphaSlider: A

compact and rapid selector, Proceedings of ACM CHI94

Conference, (April 1994), 365-371.

[3] North, C., Shneiderman, B., and Plaisant, C., User

controlled overviews of an image library: A case study

of the Visible Human, Proceedings 1st ACM Internation-

al Conference on Digital Libraries, (March 1996), 74-82.

[4] Plaisant, C., Rose, A., Milash, B., Widoff, S., and

Shneiderman, B., LifeLines: Visualizing personal histo-

ries, Proceedings of ACM CHI96 Conference (April

1996), 221-227, 518.

PERMISSION TO COPY WITHOUT FEE, ALL OR PART OF THIS MATERIAL IS GRANT-

ED PROVIDED THAT THE COPIES ARE NOT MADE OR DISTRIBUTED FOR DIRECT

COMMERCIAL ADVANTAGE, THE ACM COPYRIGHT NOTICE AND THE TITLE OF

THE PUBLICATION AND ITS DATE APPEAR, AND NOTICE IS GIVEN THAT COPYING

IS BY PERMISSION OF THE ASSOCIATION FOR COMPUTING MACHINERY. TO COPY

OTHERWISE, OR PUBLISH, REQUIRES A FEE AND/OR SPECIFIC PERMISSION.

©ACM 1072-5520/97/1100 $3.50

d e b a t e

I believe thatthere are reallimits to whatwe can do withvisualization anddirect manipula-tion because ourcomputer envi-ronments arebecoming moreand more com-plex. We cannotjust add moreand more slidersand buttons.Also, there arelimitationsbecause theusers are notcomputer-trained. So, Ibelieve that wewill have to, tosome extent,delegate certaintasks or certainparts of tasks toagents that canact on ourbehalf or thatcan at leastmake sugges-tions to us.