Games With A Purpose
Games With A Purpose
Definition and appeal
"Games with a purpose" are a category of computer game which convert the result of the play into useful computation or data. (The Wikipedia article is located here.) Games with a purpose are specifically designed to leverage humans to complete tasks at which computers are known to perform poorly. This has several important effects
- It sets a minimum rate of completion for a variety of tasks at which computers are bad that is itself quite fast.
- It provides a large quantity of priming data against which future machine learning models can be trained.
- It allows people to spend time in which they might be engaged in play that benefits only themselves engage in play that also benefits others. The most prominently known GWAP game is The ESP Game, described in a 2004 CHI paper.
Difficulties in implementing
There are a variety of difficulties associated with producing good games with a purpose. A basic problem is that humans do not actually find all activities fun - it may take a significant amount of abstraction to dissociate the task from its nature as labor and make it enjoyable. (DeOrio and Bertacco dealt with this in coming up with a game that doubled as a SAT Solver, as described here.) Furthermore, the audience playing your game may have a variety of demographic biases that keep your findings from being generally applicable. Dong and Fu have written a paper looking at this problem as seen in the different photo-tagging styles practiced by different cultures.
An alternate avenue to GWAP that is worth mentioning is the one taken by Amazon's Mechanical Turk, in which users are paid to complete small tasks. The pay rate comes out to approximately $8 per hour, according to Amazon's Best Practices Guide, but most tasks take substantially less than an hour to complete. While GWAP games rely on inherently ludic motivations, Mechanical Turk is financially motivated. The difference between these two motivations and the tasks which they can get people to perform is substantial. For example, researchers have used Mechanical Turk to find experimental test subjects or idividuals to take surveys, tasks which are not inherently thought of as "fun". Ahn and Dabbish reported quite high accuracy rates for the ESP Game, a single case suggesting the potential of ludus as a motivator for accurately completing simple tasks; a recent paper by Downs et al. looked specifically at how to vet Mechanical Turk workers for these slightly more sophisticated and boring tasks.
The "dark side" of GWAP and its cousins
As noted by Ahn and Dabbish, the GWAP process of human computation is inherently a technique, and one that can be abused by a variety of different actors. Just as GWAP can be used to solve computationally difficult but morally unambiguous tasks such as photo tagging, it can also be used for similarly difficult but much more ambiguous tasks, such as tagging photographs of political dissidents. Zittrain has presented on this potential, linking together GWAP and Mechanical Turk as existing on the same spectrum of activity; both essentially serve as digital Skinner boxes, providing different kinds of token rewards for completing simple tasks.
We have already seen innovative malicious examples of the leveraging of these services. On the financial side, and in a bizarre if expectable twist, the CAPTCHAs pioneered by Luis von Ahn as a method to make sure that you are dealing with a human are now being brute-force solved by humans employed by services such as Decaptcher in exchange for small financial rewards. On the ludic side, the 4chan community has created software to both help users easily game TIME magazine's TIME.com 100 Poll (thoroughly documented here), and to help people engage in small-scale DDoS attacks without using a botnet. (The DDoSing program is colloquially known as the Low Orbit Ion Cannon.)
Are there ethical issues associated with GWAP?
For Computer Scientists, there is little to be done regarding the "ethics" of GWAP. These games are unlike much of the rest of computer science because they effectively use a set of basic human impulses as black box replacement for much better understood -if less effective- computational algorithms. GWAP has brought us much more in touch with fundamental human thought processes than much conventional computer science research -including AI- has ever done before; if its introduction provides any kind of broader, philosphical lesson, it is that we ourselves can be reduced to poorly-understood cogs in a machine, and it behooves us to work with social scientists to better understand these human components.