Interacting with Artifacts

I finished an experimental paper on people's disposition to adopt the intentional stance while interacting with artifacts such as computers in social dilemmas.

The findings are rather surprising: People are willing to trust and to punish computers and male subjects display more trust than female subjects.

Here is the abstract:

Abstract

Previous studies have found numerous behavioral and neuropsychological differences between people’s interactions with humans and their interactions with computers in social dilemmas. A common explanation of these differences is that we adopt the intentional stance when we interact with humans, but not when we interact with computers. Although this explanation is plausible, in this paper, we provide some evidence that in some contexts, people are willing to adopt the intentional stance in interacting with computers in social dilemmas. As a result, they interact with computers—cooperate with them and, surprisingly, punish them—as they would interact with humans. Using a single-round, extensive form trust game with a punishment stage, we compared the performance of participants who wrongly believed that they were interacting with a human partner and the performance of participants who knew that they were interacting with a computer. In both conditions, participants behaved differently from the predictions of the standard economic model: Players cooperated and punished their partner. Surprisingly, we found a strikingly similar pattern of cooperation and punishment when participants incorrectly believed that they were playing with a human partner and when they knew that they were playing with a computer. Additionally, gender was found to affect participants’ decision to trust their partner. We discuss whether and when people adopt the intentional stance in interacting with artifacts such as computers.

The paper can be read Here: Download punishing_artifacts_11_12_machery.pdf .  Comments are most appreciated, as usual!

Edouard Machery

Comments

[Chris]

Maybe it's just the stats teacher in me, but I'd really like to see error bars on those graphs, especially Figure 5.

[Edouard Machery]

Chris,

Good point. I will add them.
Edouard

[Anibal]

As a devil´s advocate.
Following your line of reasoning (people in interaction punish humans or computers alike because they use the intentional stance indiscriminately), what if all the participants are "strong reciprocators" and they punish every violation of an expectation to cooperate irrespective if the cooperative partner is a human fellow or an inanimate device (artificial intelligence program).Perhaps, too much coincidence.
But even if they are all strong reciprocators or not, and punish humans partners or computers in similar manner, what if the cyberpsycology in the absence of salient "humane feature" make them to treat always as if were a human partner "whoever" responds within the margins or instructions of the games in question. Because if the participants would have an image or idea about the program (visualitation of it) maybe they stop to use their intentional stance.

For example, developmental psychologists (Meltzoff A. [1995]"Understanding the intentions of others: Re-enactment of intended acts by 18 month-old-children" DEVELOPMENTAL PSYCHOLOGY,31, 838-850) deny that infants percive or atribute goals to mechanical devices; and is very probable that this remains within an adult psychology.

The intentional stance is a three-fold socio-psychological tool, which of the three parts is in use because the outcome (punishment) could be the same but the process arriving at it not. For example, participants can punish the computer via the design stance (they could think the program is nasty subsequently they feel outraged and then punish) or the physical stance (they could think the structure is defective subsequently they feel a negative emotion and then punish) but not necessarily via the "intentional stance" deploying belief or intentions to computers. In the rationalizations of any person always there is a mysterious factor so we´ll never know what process they first engage.
The psychology in virtual enviroments is more concern with the transference of our natural psychology to virual realities, and our natural psyhcology evolved to deal with that kind of staff that is animate or inanimate, the technological progress that make inanimate entities "as if" animate, or truly animate, pose many problems for the moral domain or social domain to be extrapolated outside the prefered ontology.

[Edouard Machery]

Anibal,

It is not the case that people are willing to "punish" any partner--independently of its status as an agent. Previous experiments show that subjects cooperate and punish less when they are interacting with a computer whose strategy has been described as "Random" than when they believe that they are interacting with a human partner (See the literature review in Section 1 of my paper).

[Dan]

Neat study. I have several rather technical comments about the paper.

First, I'd guess that all of the participants who offered zero would request zero, and that everyone who requested zero would punish zero, since there's nothing to request if you haven't offered anything, and there's no reason to punish if you didn't request anything. Did that happen? Even if that didn't happen in every single case, I think that you should exclude all of the people who offered zero when analyzing requests, and exclude all of the people who offered or requested zero when analyzing punishments, since their responses are already largely determined by earlier stages of the game. Including these people probably inflates your correlation between request and 1.5 x offer, as well as inflating your standard deviations for request amounts and for punishment amounts. If you do exclude these people, then when giving the results for each stage of the game, you should say your exclusion criterion and why you're using it, and what percent of the participants in each condition are being excluded, and whether there is a difference between conditions in percent excluded.

Second, about that correlation between the request and 1.5 times the offer: rescaling a variable (multiplying it by a constant) does not change the correlation. The correlation between the request and the offer must be the same as the correlation between request and 1.5 x offer. Here is what I think you should do here instead. First, report the correlation between offer and request (including only the participants who made a nonzero offer). Second, take the request divided by the offer as your variable, and report the mean, the SD, and what percent of the time it was exactly 1.5. If participants are doing what you say, then this variable should have a mean near 1.5 and a small standard deviation, and it should often be exactly 1.5.

Actually, given the strong correlation between request and offer, and the theoretical relationship between the two, I think that request divided by offer is probably a better measure of request size than the raw monetary amount requested. I'd recommend replacing all of your analyses of the request amount with analyses of this ratio. You could mention in a one or two sentence footnote that other ways of doing the analysis turn out the same (assuming that they do). If a lot of people are requesting exactly 1.5 times the offer, then you might want to do some kind of nonparametric analysis as well.

I don't know the citations, but previous research on dictator & ultimatum games has found evidence for a fairness norm, with lots of people choosing a 50-50 split. You might want to refer to that research when you discuss your request=1.5xoffer prediction. Another seemingly fair distribution for your game is for player 2 to return enough money so that both players end up with the same total. For instance, if player 1 offered $6, then player 2 would return $7 so that both players end up with $11. It would be good to report what percent of participants made requests according to this fairness rule (in each condition). (The two fairness rules are identical when player 1 offers the whole $10, so you might want to give that number separately).

For the punishment results, it would be nice to at least have a one sentence footnote mentioning that other ways of analyzing the amount of punishment (e.g. the raw amount instead of the percentage of income remaining) had the same pattern of results (again, assuming that they did).

Next, you should probably say something about how participants were debriefed and what payment you gave them. At least in most psychology journals that I know of, that would be expected. I hope that you gave everyone at least the 10 ecus!

Is the difference in number of comments referring to emotions like anger & disgust statistically significant? If you're claiming that there's a difference, you should report the statistical test.

I wouldn't emphasize the sex differences that you found, unless there was a reason to expect them based on previous research. Researchers test for sex differences in just about every study that they conduct, and so one in twenty studies will find a significant sex difference just by chance. You should mention them in the results, but I don't know if it's worth putting them in the abstract.

Finally, Figures 3 and 5 appear unnecessary, since the same information is already displayed in Figures 2 & 4. And, you should be consistent in your Figures about which color represents males and which females.

Again, I really like the study, or else I wouldn't have taken the time to make these comments. I doubt that these changes to the analyses would have much of an impact on your conclusions.

By the way, I'm a grad student in social psychology who has been reading this blog since its beginning, and commenting occasionally.

[Edouard Machery]

Dan,

Thanks for these great comments. I will make the changes you suggest this week-end.

Edouard