What if this time and energy were also channeled toward solving computational problems and training AI algorithms?

People playing GWAPs^22–25 perform basic tasks that cannot be automated. The ESP Game, ²² a.k.a. the Google Image Labeler ( images.google.com/ imagelabeler/), is a GWAP in which people provide meaningful, accurate labels for images on the Web as a side effect of playing the game; for example, an image of a man and a dog is labeled “dog,” “man,” and “pet.” The game is fast-paced, enjoyable, and competitive; as of July 2008, 200,000 players had contributed more than 50 million labels; try it yourself at www.gwap.com. These labels can be used to improve Web-based image search, which typically involves noisy information (such as filenames and adjacent text). Rather than using computer-vision techniques that do not work well enough, the ESP Game constructively channels its players to do the work of labeling images in a form of entertainment.

Other GWAPs include Peekaboom, ²⁵ which locates objects within images (and has been played more than 500,000 human-hours); Phetch, ²³ which annotates images with descriptive paragraphs; and Verbosity, ²⁴ which collects commonsense facts in order to train reasoning algorithms. In each, people play not because they are personally interested in solving an instance of a computational problem but because they wish to be entertained.

The ESP Game, introduced in 2003, and its successors represent the first seamless integration of game play and computation. How can this approach be generalized? Our experience building and testing GWAPs with hundreds of thousands of players has helped us spell out general guidelines for GWAP development. Here, we articulate three GWAP game “templates” representing three general classes of games containing all the GWAPs we’ve created to date. They can be applied to any computational problem to construct a game that encourages players to solve problem instances. Each template defines the basic rules and winning conditions of a game in a way that is in the players’ best interest to perform the intended computation. We also describe a set of design principles that comple-

People play not
because they
are personally
interested in solving
an instance of a
computational
problem but
because they wish
to be entertained.

ment the basic game templates. While each template specifies the fundamental structure for a class of games, the general design principles make the games more enjoyable while improving the quality of the output produced by players. Finally, we propose a set of metrics defining GWAP success in terms of maximizing the utility obtained per human-hour spent playing the game.

Related Work

Though previous research recognized the utility of human cycles and the motivational power of gamelike interfaces, none successfully combined these concepts into a general method for harnessing human processing skills through computer games.

Networked individuals accomplishing work. Some of the earliest examples of networked individuals accomplishing work online, dating to the 1960s, were open-source software-develop-ment projects. These efforts typically involved contributions from hundreds, if not thousands, of programmers worldwide. More recent examples of networked distributed collaboration include Wikipedia, by some measures equal in quality to the Encyclopaedia Britannica. ⁶

The collaborative effort by large numbers of networked individuals makes it possible to accomplish tasks that would be much more difficult, time consuming, and in some cases nearly impossible for a lone person or for a small group of individuals to do alone. An example is the recent Amazon Mechanical Turk system ( developed in 2005, www.mturk.com/mturk/ welcome) in which large computational tasks are split into smaller chunks and divvied up among people willing to complete small amounts of work for some minimal amount of money.

Open Mind Initiative. The Open Mind Initiative^{18, 19} is a worldwide research endeavor developing “ intelligent” software by leveraging human skills to train computers. It collects information from regular Internet users, or Netizens, and feeds it to machine-learning algorithms. Volunteers participate by providing answers to questions computers cannot answer (such as “What is in this image?”), aiming to teach computer programs common-