Take a look at this video:
You may have seen it before -- it's the work of a CGI animation studio that takes the motions of human actors and turns them into animated models, giving them the ability to put incredibly realistic figures in impossible situations, like on Mars, or swimming in lava, or whatever an animator can conceive of.
But the advent of realistic simulations such as this makes it clear that people need to be more aware than ever of the potential for digital fraud. We now have email spam, but in the future we might have similar computerized instant messaging spam, or even video messaging spam. What does it take to determine if a video image of a human represents a real person or a computer simulation? It can actually be easier than you might think to create a computer simulation that is convincing to humans.
Consider this simple experiment led by Jeremy Bailenson. Thirteen pairs of students were set up at computer terminals in different rooms, and given only one way to "communicate" with each other: pressing the space bar. Depressing the space bar caused a light on the other user's screen to change color, and releasing it caused it to return to the original color. Meanwhile, a second light on the other user's screen was controlled by one of several possible types of simple computer simulations. So the display looked like this:
The job of each participant was to convince the other participant that he or she was human, while simultaneously figuring out which flashing dot represented the actions of her/his partner.
Each trial lasted 60 seconds, after which the partners tried to guess which dot was "human" and which was computer-generated. They also rated their confidence in their guess.
The computer used five different possible tricks to fool the humans:
- Play a pre-recorded script generated by a human
- Random
- Play a pattern such as the morse-code "SOS" signal, at gradually increasing/decreasing speeds
- Mimic the human: play back her/his actions with a four-second delay
- Alternate between mimicry and a pattern
So how often were the humans correct? Here are the results:
For most of the patterns, humans did no better than chance (although the trend was to be accurate). But for the "mimic" condition, humans were fooled more than 60 percent of the time. All it took for the computer to reliably deceive was copying exactly what the humans did.
Next the researchers moved to a virtual environment. This time, a computer-generated character read a persuasive speech while viewers watched wearing a three-dimensional virtual reality headset. The headset detected head movements of the viewer in order to generate a more realistic virtual environment. However, this also allowed the computer-generated character to mimic the viewer's movement in one of three ways (again delayed by four seconds):
- Mirror: An exact-mirror-image of the viewer's movements
- Congruent: A mirror-image of right-left movements, but opposite of up/down movements
- Switch: left-right movements in viewer imitated in up-down movements by computer, and vice-versa
Viewers were asked if they agreed with the speech, as well as being asked several evaluative questions about the computerized speaker: how trustworthy, warm, and informative was he/she (the speaker's gender was matched to the viewer)? They were also asked if they noticed anything unusual about the speaker.
Here are some of the results:
In nearly all the conditions, if viewers didn't detect the computer was mimicking their own movements, they rated the computer higher, including even their level of agreement with the computer's statements. It's quite clear that a simple way for computerized spam engines to impress humans is to imitate their own actions -- as long as the computer isn't so obvious as to get detected. In this experiment, humans were much more likely to detect the mirror-image imitation by the computer, presumably because they see themselves in the mirror every day.
J BAILENSON, N YEE, K PATEL, A BEALL (2008). Detecting digital chameleons Computers in Human Behavior, 24 (1), 66-87 DOI: 10.1016/j.chb.2007.01.015
- Log in to post comments
I'm not really sure I understand the blue graphs. Also, the meaning of switch doesn't seem to be well-explained. Can we see examples of that?
This really looks interesting. I'd like to understand the implications a little better.
I'm not really sure I understand the blue graphs. Also, the meaning of switch doesn't seem to be well-explained. Can we see examples of that?
This really looks interesting. I'd like to understand the implications a little better.
Hmm. So the first comment was from Greg Padilla, and the second comment was from a computer simulating Greg Padilla? Or maybe it was vice versa, I dunno. This is too sinister.
Just a note about the video. Only the face is CG. It is superimposed on a live plate. The technology is still amazing, including the tracking, rendering and animation, but some of the verisimilitude is taken from the realism of the live plate.
Mark: heh. No, we're just having some problems with the commenting system.
You still need an Actor for this to work ie. human involvement. What I'd like to see is the same performance from a completely CG driven character. The "Uncanny Valley" is still too wide to bridge that gap (see Beowulf).
to me they are so distinctly different. and their voices are even more different... i don't understand people honestly confuse them. :) great blog btw. i added a link to you.