Tom Chi  

I, Robot, You Jane.

December 12th, 2003 by Tom Chi :: see related comic

As I type this my Roomba rumbles in the room behind me. It negotiates a tambourine noisily, then gets caught in some guitar cables. I scoop him up and drop him a couple feet over. Even though it is obviously a machine, I often find myself speaking as if it wasn’t. I’ll tell people how “he sometimes gets confused” or how “diligently he vacuums under the table”. Sometimes I’ll even raise my voice when he is about to get caught in the Venetian blind strings.

None of this makes sense, yet it makes perfect sense. It feels perfectly natural for me to relate to my little autonomous vacuum cleaner in this way. After all, how else could I talk about it? Our entire language is built around words for emotions, words for folly, words for thoughts and ideas. That’s the raw material we have to work with, but these words refer to properties that machines do not have.I *could* sidestep this problem by talking about my machine in purely technical terms. Instead of simply saying that “he is confused”, I could speak at length about how his memory-less algorithm for floor coverage is based on simple feedback-driven local behavior with a couple global heuristics slapped on top… and that such an algorithm cannot negotiate terrain of a certain type. But really, it makes more sense to say: “he is confused.” It makes more sense because ultimately whenever I describe my robot, I am talking to another human being.

I think this is an important point. Just like an essay is a form of communication, products can be a form of communication as well. We tend to think of the user as interacting with the technology, but when you look one level deeper than this, he/she is also interacting with the designers of the technology. The designers want to communicate with their design, and aim to inspire certain thoughts and actions in the users of their product. Such an act of creation is not so different than painting a picture or writing an essay to communicate a set of ideas or possibilities.

So perhaps we anthropomorphize technology because in some designs we can feel the intention of the designer so distinctly, that it feels like a human presence. Or perhaps we simply don’t have the words to talk about technology in another way. Regardless of the why, we definitely do try to relate to technology in human ways, and this week’s comic is all about how much our technology should try to relate back. From Tomaguchi to Clippy to ELIZA to Aibo, agents provide a set of interaction problems and possibilities as complex as the social interactions they model.

I’d like to hear some reader experiences with agents, affective interfaces, robots, or any other anthropomorphized technology. Do they work for you?

18 Responses to “I, Robot, You Jane.”
Steve Nelson wrote:

We have a long way to come before it’s easy to teach technology what humans really want. It will take quite a bit of work to make technology understand our emotions and just general weirdness, but I’m convinced it is possible. I’m equally convinced nothing has come close yet. We need more things like the Loebner prize. Take a look at this:

Tom Chi wrote:

Ah, of course, the Loebner prize. This is the prize award to the first person who can create a machine which passes the Turing Test. Salon wrote a great article about the history of the prize wherein Loebner gets into a flame war with Marvin Minsky.

Steve is correct to say that we have a long way to go before our research is complete. I mean, even if we had a machine that passes a Turing Test, we still couldn’t do much with it.

Just think about how we’d apply it to even a simple task like taking orders at a fast food window or replacing those prerecorded phone menus voices. It would need to be mapped on to some knowledge base to be useful, and it would need to be able to handle a myriad of complex corner cases, as well as being able to deal with garbled words and strange requests. In both cases I would imagine users longing for the certainty of the old ways as soon as the system failed them once.

On top of all that humans would need to feel comfortable speaking to a machine as if it were human. Although I raise my voice at my Roomba, the entire dynamic would be wholly different if it talked back. The wrong response at the wrong time might make me switch the little bugger off. Now if this interface was in a critical path for me to get something done, and I couldn’t switch it off or flip back to an old menu-based system, I might start raising my voice at the system designers!

Kevin Cheng wrote:

Steve, why are you convinced that it is possible for technology to understand our emotions and weirdness? It seems that we don’t have the ability to understand other humans to that level so how could a machine, programmed by humans that don’t understand, have the ability to do so?

Jamie wrote:

This is a classic argument made forth by Ada Lovelace back in the 1800s. Specifically, she said (about her prototype computer):

“The Analytical Engine has no pretensions to originate anything. It can do whatever we know how to order it to perform.”

However, computers are already “learning” to do things we never thought about. For example, we used computers to find better ways of playing backgammon that previously no one had ever tried before. We simply told the computer to “play backgammon”, and it came up with very surprising strategies that beat the world champion.

Of course, this is a long way away from the Turing Test, but the point is that a computer is not limited to what we tell it to do; or rather, what we know does not limit what it discovers.

Tom Chi wrote:

I hadn’t heard of the Backgammon example before. It’s very interesting, but most games (Backgammon, Chess, etc) have very clear goals and clear ways of evaluate the position at each point in the game.

Compare this to interfaces which deal with emotion. If the goal of the system is to keep the user happy, then how do you measure progress? What is the optimal amount of happiness the system could expect to achieve?

Really these questions don’t make any sense. Even the owner of the emotion is not great at quantifying his or her happiness, and as Kevin mentioned, the happiness at any point is probably a function of much more than the interaction with the machine. So how can a system like this ever converge to optimum happiness induction?

And before anyone mentions “Go” in the gaming department, all I can say about Go is that modern computers are not fast enough to solve it *yet*. This would be true of any game that is built on an NP problem. Give it 20 years. Even without a breakthrough, our computers will be fast enough in 20 years to beat the world Go champion at his own game.

Jamie wrote:

Actually, it is not clear at all how to evaluate your position in chess at any given point of the game. Chess programs like X3D Fritz use miriads of heuristics to gauge how well they’re doing. But the key word here is heuristics. Even in chess, we don’t converge to an “optimal solution”. So in your scenario, the computer would simply *try* to bring about happiness, much like a human does. Isn’t that what we’re after anyway?

As for GO, I don’t believe we will *solve* it without some sort of major breakthrough. That is, under current methods of game playing — predominantly searching, we would need far more than Moore’s law to search through the entire space. Here’s a quick analysis:

Chess: ~10^40 positions.
Go: ~10^170 positions.

Let’s say we can search through all the positions in chess today (which we can’t). Let’s say we double the CPU power every year (which we don’t). That only amounts to being able to ~10^60 positions. That’s not even accounting for the fact that evaluating positions in GO is far harder than evaluating positions in Chess.

The fact is, we need either a major computing breakthrough, or a major AI game-playing breakthrough in order to beat the world champion. I believe at least one of these will happen, but I don’t think it’s plausible to assume computers will simply “catch up”.

Although having said all this, I’m not convinced computers will be able to deal well with emotions, either. =)

Jamie wrote:

I must’ve been smoking crack… the above number (after 20 years) should have been 10^40*2^20, which actually is much less than 10^60, which I suppose further proves my point. Whoops.

Tom Chi wrote:

The entire Chess and Go thread is a little off topic, but your comment is worth a comment. You are correct to say that it is currently impossible to calculate every possible chess position, and it might never be possible. But we already have programs which you can run on non-specialized machines which can draw the world champion. So to beat the human you don’t *need* to reach the optimal solution. Same is true with Go. Though the computational complexity of considering every possibility in Go is way more than anyone can imagine, you don’t *need* to be that good to beat the world champion.

Also I tend to take computational calculations like this with a grain of salt, because they remind me of this analogy: A baseball is made up of atoms. A frickin’ boatload of atoms, all bumping into each other, bumping into the surrounding air, etc. They also represent a complex collection of electrical and nuclear forces. Well if someone were to try to attack the problem of calculating a baseball trajectory by calculating all the tiny forces and atomic interactions they might soon conclude there is not enough processing power in the universe to do so! You don’t need a “major breakthrough” to find out that the problem can be modeled more sucessfully at a different level. The path of the ball can determined to many significant digits by starting with a parabolic trajectory then adjusting for air resistence. While the answer is not the “perfect” answer that you would get from modeling every atom, it’s as good as we need.

Fritz 8, which just drew Kasparov in a match uses this approach. It does agressive pruning on move trees — doesn’t even try to explore most of them. It has measures for activity, space, and grades for material advantage. All of these are used to consider chess from a higher level than simple move by move crunching. I’m expecting Go to follow a similar trajectory. Moore’s law will just a be a tiny part of it.

Steve Nelson wrote:

Chess and Go are child’s play compared to human emotions. It would be interesting to find out how many of those 10000000000000000000000000000000000000000000000000000000000 chess moves (the 0s are more dramatic) had to be taught to the computer to beat Kasparov? If it knew all of them, then he would have no chance of winning. But it doesn’t, that’s why he can still win.

A computer program doesn’t have to know every edge case to be useful, which is why i think it’s possible to beat the Turing test. The big question in my mind is how much information is necessary to pass the Turing test? i don’t know.

I think the place to begin with the Turing test is to create a grammar checker that, just like Kasparov, can beat any college English professor at editing a paper. The end result would be both useful and teach us quite a bit about human emotions.

Jamie wrote:

Tom: You mention calculating trajectory — when we figured out how to do that, I would classify that as a major breakthrough in physics. The same can be said of GO — we need a major breakthrough in the way we search for (or otherwise derive) strategies. My point about the numbers is not that we *have* to search through everything, but rather that we can’t go about it the same way as chess because it is immensely more complex. But I suppose we’re to a point wherewe actually agree, but it’s just the technicalities that are in the way.

Bob Salmon wrote:

How does the understanding of / display of emotions compare to adaptation or idiosyncrasy in terms of importance to usability or even attachment/anthropomorphism?

Sorry - I’ll decode that. Going back to a previous discussions about personality and character, many Mini owners name their cars and say things like “Don’t stall on me girl”. The car doesn’t display any emotion (despite Norman’s Turn Signals book), and yet the interaction involves human terms.

Is this because the car displays unpredictability, or uniqueness or because you’ve owned it for a long time etc? (If it’s length of ownership, why doesn’t e.g. an old pair of shoes inspire the same interaction?)

Would things be enhanced if the car started adapting itself to your behaviour etc? This would mimic a human relationship more, where each side is changed via the relationship.

I think what I’m saying is that while emotions would add a lot to interactions, a lot of ground can be covered without it. Trouble is, I can’t put my finger on what this other stuff is.

Tomas Lindroth wrote:

Well, a lot can be said about emotions, but I don’t think, at this time, it is so much about computers or other technical artifacts interpreting our emotions as much as they affecting ours. As Norman puts it “So if only for purely utilitarian reasons, devices and software should be designed to influence the mood of the user; they will be more effective because they are more affective.” (Scientific American), I think that this is where we are now and maybe some time in the future we may learn the little R2D2s out there how to interpret us. As I see it R2D2 is a quite good example of a simple way to affect our emotions, with only a little piip here and a piip there, it is quite easy to tell the mood of the droid. In the same way it was quite easy for the designer of Furby to affect our emotions when one held it up-side-down for a while (it started to scream, heartbreaking!). These are all really simple techniques to affect human emotions and if we could just begin to understand and start to integrate this into ICT we would be far better of then we are now. And yes, I think that the iMac and other neatly designed products is a first step towards making technology more personal and emotional, in that sense that they affect us in a positive way.

Kevin Cheng wrote:

One: I’ve never put a Furby upside down like that and didn’t realize they did that. Now I wish I want to go and try it.

Two: A lot of people seem to be talking about the way products affect the humans such as our previous thread ( see: “The One” ) about the magic of Apple and Minis. There’s clearly value in stirring emotional bonds with the user the way Tom has described (and Bob previously described regarding his former Mini-ownership status).

Does that imply that it would also be useful for the machines to learn to detect our emotions so they can be smarter about how to form such a bond?

Blackbeltjones Work wrote:

Coddling your Roomba.

Tom from OK/Cancel on the anthropomorphic tendency I’ve definitely felt towards my Roomba. “None of this makes sense, yet it…

Roomba Review wrote:

In much the same way the Furby screams when you hold it upside down, the Roomba actually calls for help when it get’s stuck as well.

I believe that the only way that the Turing test could be considered truely passed would be when holding a furby type object upside down in order to observe the results would be something that actually disturbed people, instead of amused them.

Kevin Cheng wrote:

That’s an interesting point. If we reached that stage, I can see extremists start to fight for “robot rights” … wrote:

Roomba reflections

OK/Cancel: I, Robot, You Jane, via Blackbelt Jones. Roomba Review, community site for Roomba. Looks just launched. Astroturf? Those are some good prices. Genius art-guy Gary Panter on Roomba. A bar. Takeapart walkthrough at Jake’s World (which has some…

Peter Centgraf wrote:

Just to throw a monkey wrench into this discussion of affectation and emotion . . . Recent work in social psychology has raised the interesting possibility that our outward manifestations are causes as well as effects of our emotional states. For example, subjects that were asked to smile while viewing a particular video clip rated it as more funny than subjects asked to frown. Extrapolating from this result, it could be very easy to induce emotional responses to interactions with devices — if you can induce behaviors that are similar to those of a compassionate person, you may induce the feeling of compassion. If you provide a feedback mechanism for talking to a device, you could foster a sense of attachment to it. No complex understanding of human behavior or emotions would be necessary.

Leave a Reply

OK/Cancel is a comic strip collaboration co-written and co-illustrated by Kevin Cheng and Tom Chi. Our subject matter focuses on interfaces, good and bad and the people behind the industry of building interfaces - usability specialists, interaction designers, human-computer interaction (HCI) experts, industrial designers, etc. (Who Links Here) ?