The Proteus Paradox (22 page)

In 2007, the balance between data access and third parties changed yet again. Blizzard released the Armory for
World of Warcraft,
a website that allows you to look up any active character playing the game. For each character, the Armory catalogs thousands of statistics and achievements. And this list has kept growing with the updates over the years. Currently, the Armory provides over 3,500 variables for each active character, updated daily. These publicly accessible variables cover a broad range of behaviors in
World of Warcraft:
progress through high-level dungeons, combat skill specializations, number of vanity pets owned, frequency and nature of deaths (for example, by drowning or by falling from high places), and even the exact number of virtual hugs given over a character's career. It is an unbelievable treasure trove of data.

Certainly, there is commercial value in the server-side data, but the earlier experiment with add-on scripting created a thriving community of
modders
—gamers developing in-game tools, usually for free. Hundreds of add-ons were developed, many of which were updated diligently with each game patch. The modding community enhanced engagement by allowing gamers to tweak the game interface to their specific play styles. By observing the download counts of the add-ons, it also allowed Blizzard to understand the player community's needs. More than once Blizzard added game functionality that was previously available only with an add-on. If anything was
lost with sharing the game data, it was more than made up for with the free programming labor and marketing research generated by the player community. The Armory seemed to expand on this philosophy. The website enhances player engagement by making it easy for new players to understand how elite players are optimizing their equipment and abilities, quickly check a guild candidate's résumé, or figure out what upgrades are available for a character's current equipment. Different websites use the Armory data to create detailed census reports, prevalence of specific classes or specializations in player-versus-player rankings, and the progress of elite guilds in the high-end content. The data sharing created a thriving player community around the game as well as increased the engagement with the game itself.

The Armory also dramatically changed how we could collect data from
World of Warcraft
. Now we could collect data from characters regardless of what server they were on and whether they were currently online. And instead of 7 variables, we now had access to over 3,500. The Armory also solved the problem of linking in-game data with survey data on demographics and personality. We could track census data from an entire server, but there was no easy way to survey those players afterward. The flipside also wouldn't have worked. Say we surveyed players and asked for their active characters. When we were using the old census add-on, there was no easy way to track a random set of characters because they were spread across over two hundred available
World of Warcraft
servers. The cycling time, required to read information on all the servers before repeating, would be so high that, during peak hours, we would miss many characters. With the Armory, it was trivial to access data from a random set of characters, no matter what server they were on. This meant that we could
collect survey data from players first and then go to the Armory to access their character data.

In 2009, with the help of a government grant, this was exactly what my colleagues and I at PARC, where I was now a full-time researcher, set out to do. Our team first created an automated collection tool for the Armory. With the infrastructure in place, we collected survey data from more than a thousand
World of Warcraft
players. Apart from demographic and personality information, we asked the players to list their active characters in the game. We then let the Armory data collection accumulate in-game metrics for six months.

Somewhere in those six months it dawned on the team that we were about to be deluged with data. Even when we had just 7 variables in the early PlayOn study, it took us two years to analyze the data. In the new study, we had 1,040 players, each with on average three characters, each of which generated over 3,500 variables per day over a six-month period. In the field of data mining, one well-known saying is “Garbage in, garbage out.” Researchers have many tools that allow them to click a few buttons to create attractive graphs and charts, but unless they are feeding these tools meaningful data, the pretty graphs are useless. In 2012, Dmitry Nozhnin, head of analytics at online games publisher Innova, described his experience analyzing data from the online game
Aion
. He was interested in understanding why some players left the game early. Using analytic tools to build complex models, he was able to predict whether a player was about to quit the game, but Nozhnin lamented that even “knowing with very high accuracy when a player will leave, I still don't have a clue why she will leave.”
⁶

As the PARC team peered into the growing database of Armory data, we were aware of the risk of falling into an abyss of uninterpretable
metrics. If Extraversion were correlated with owning a particular set of swords or visiting a particular zone, what does that actually mean? The statistical connection between different variables does not come with an explanation for why they are connected. This is the issue Nozhnin encountered. Moreover, a correlation between two variables does not imply that one caused the other. In fact, the correlation may be caused by a third variable that isn't measured or accounted for. For example, people with bigger feet have bigger brains, but bigger feet do not cause bigger brains. This is because as children get older, both their feet and brains grow larger. The connection between foot size and brain size is coincidental.

A more problematic issue concerned the game variables: many were hopelessly confounded with frequency of gameplay and character progression. Imagine that Frank and I both play
World of Warcraft
. I have a level 80 character—that is, a very high-level character—and Frank has a level 80 and a level 1 character. If we take the average of all the variables for each player, then Frank is unfairly penalized for having a level 1 character. His level 1 character's combat and achievement metrics drag down his overall average. Even if we both have just one character each—say I have a level 40 character and Frank has a level 80 character—we'd still have a problem. Metrics in
World of Warcraft
do not progress linearly; there are sudden, uneven, and exponential gains at certain levels in the game that make it difficult to compare our metrics. If Frank has made 800 player kills at level 80 and I have made 10 player kills at level 40, is Frank really more aggressive than I am, or is it just that level 80 characters have easier access to large numbers of player kills?

We began to develop strategies to create meaningful variables from this morass of data. One was to create conceptually meaningful
aggregates. It is impossible to interpret what stepping into any one particular zone means, but the percentage of all zones visited maps to a psychologically meaningful concept of exploration. In short, we tried as much as possible to create variables that came with explanations built in. For the problem of variable confounds, we used multiple normalization strategies. Consider the fact that a player's total amount of healing done in the game is hopelessly confounded with their character level and frequency of playing, but the ratio of healing done against damage done produces a preference metric for healing.

We developed a dozen such strategies to create meaningful variables from the data. And when we examined the connections between real-world personality and in-game behaviors, we found that personality was indeed expressed in
World of Warcraft
. More important, these in-game behaviors aligned with the personality trait definitions.

Players with high Extraversion prefer group activities and are more likely to participate in large dungeon raids. Players with low Extraversion prefer solitary activities such as cooking, fishing, and questing. Players with high Agreeableness give out more virtual hugs, cheers, and waves, preferring noncombat activities such as exploration and crafting. Players with low Agreeableness prefer the combat and antagonistic elements of the game. They enjoy killing other players, die more often, and participate in more duels and arena matches. Players with high Conscientiousness enjoy collecting things in the game, whether this is accumulating vanity pets or travel mounts (for example, horses, griffons). They also enjoy the self-discipline required to advance in the cooking and fishing professions. On the other hand, players with low Conscientiousness are more likely to die from falling from high places. Players with high
Openness to Experience have more active characters, are more likely to have characters on multiple servers, and spend more time exploring the game world. Players with low Openness to Experience prefer the traditional combat elements of the game, focusing on dungeons and raiding.

Neuroticism was the only trait that did not have a clear correspondence with in-game behaviors. Nevertheless, the predominant pattern of correspondence between real-world personality and in-game behavior is striking. Neither the overt fantasy nor the constant invitations to reinvent ourselves drown out our personalities. Even when we take on virtual bodies, our personalities are expressed in online games.
⁷

We also used machine learning tools to see if we could extract simple rules that could predict someone's demographic attributes based on his or her in-game behaviors. The rules for predicting gender had surprisingly high accuracies. If you play a male character for more than 61 percent of your total playing time, there is a 94 percent chance that you are male in real life. And if you have no male characters and have given out more than eighty-nine hugs, there is a 93 percent chance that you are female in real life. As Charles Duhigg reports in his book
The Power of Habit,
the retail chain Target used similar predictive tools to infer whether female shoppers were pregnant based on their shopping behavior. By accurately making this inference before a baby is born (and before the public birth record invites a flood of advertisements from retailers), Target would be able to capitalize on a moment in a woman's life when her routines and habits change because of a significant life event. By sending
advertisements at this critical juncture, Target hoped to create new shopping habits. Target's algorithm was so successful that they figured out that a teenage girl was pregnant before she had told her parents; they were tipped off when Target addressed congratulatory coupons for baby clothes and cribs to the pregnant teen. The Proteus Paradox reveals psychological tools in virtual worlds that have a great synergy. Virtual worlds not only provide novel methods for psychological control, they also provide the means to perfectly tailor those manipulations based on an individual's attitudes, demographics, attributes, and personality.
⁸

Peter Steiner's 1993 cartoon in the
New Yorker
captured the promise of freedom and anonymity that the Internet once offered: “On the Internet, nobody knows you're a dog.” But our era of big data—especially as more data are captured and at finer-grain resolutions—flips this premise around. Our behaviors online, in virtual worlds, and when using smart mobile devices allow others to make accurate inferences about who we are and what we like. On the Internet, everybody knows you're a dog. And as marketers and advertisers compete over this growing flood of data, the facts they learn about each of us are likely to become more and more unsettling. Pregnancy may be a celebratory occasion, but what about the onset of diabetes, an impending divorce, or impending unemployment? These retailers not only have data from you, they also have data from your friends, children, spouse, and employer. What happens when a retail store knows that your spouse is planning to leave you before you do? Furthermore, the true risk may not be in accurate predictions but, rather, in inaccurate predictions. Target may send you shopping coupons if they think you are pregnant, but what might law enforcement agencies do if they think you are committing a crime? On an April morning in 2012, a Kansas family found out when a police
squad armed with assault rifles stormed their house. The family believes they were targeted because they had purchased hydroponic supplies to grow plants indoors. The police squad did not find marijuana even after a drug-sniffing dog was brought in to help, but they did find six plants in the basement: three tomatoes, one melon, and two butternut squash.
⁹

We are living in a world in which a digital escapist fantasy and a surveillance state refer to the same thing. Media scholar Mark Andre-jevic uses the phrase “digital enclosure” to refer to this rapidly growing phenomenon of users freely submitting to enhanced surveillance in order to gain access to a digital network or community—whether it is Gmail, Facebook, or virtual worlds. Not only do virtual worlds provide unprecedented powers of surveillance, they also provide unprecedented executive powers. In January 2005, many
World of War-craft
players became disgruntled with the announced changes to the warrior class. In addition to complaining on web forums, they staged an in-game protest. These disgruntled players created level 1 gnome warriors on the Thunderlord server, stripped down save for their loincloths, and protested on the bridge in the Dwarven city of Ironforge. Soon afterward, many of those players found their game account locked, with the following message displayed on their login screen: “This World of Warcraft account has been suspended—Please check the registered email address of this account for further information.” It is easy to create and customize virtual characters. It is equally easy to delete virtual characters; our digital bodies do not decompose and do not require elaborate disposal methods. Existence comes and goes at the click of a button.
¹⁰