Science

A collaborative quest

The Media Lab’s BeeMe project lets users work together to defeat a self-improving AI

8785 beeme photo from alan foster
BeeMe's two player characters face off in Lobby 7.
Courtesy of Alan Foster

Set in a universe threatened by a self-improving AI by the name of Zookd, BeeMe, a project by the MIT Media Lab, tasks a virtual network of players to control an actor in order to find a way to stop Zookd from overriding its built-in moral code.

The MIT Media Lab has long been home to non-sequitur social experiments, engaging the public with anything from creating a tapestry of voters to the world’s first psychopath AI. However, this may be one of their most interactive and involved experiments yet. BeeMee was developed by Niccolo Pescetelli, a human psychologist working to understand the dynamics of human collective intelligence, and Dr. Iyad Rahwan, a scientist who leads the Scalable Cooperation group at the MIT Media Lab. This immersive social event invites Internet users to work together to stop Zookd from succeeding in his mission. They do so by controlling an actor, who has consented to give up their free will for the evening. The mechanism for doing so is a hive mind — participants come up with possibilities for actions and vote on them. The most popular option will be performed by an actor and live-streamed to the participants.

The project draws its inspiration from an intersection of literature, performing arts, gaming, popular culture, and YouTube streaming, and aims to break ground in previously poorly-understood areas of human dynamics and human intelligence research.

BeeMe attempts to demonstrate the dynamics of cooperative human intelligence and the influence of internet anonymity on real-life human dynamics. One of the landmarks of human intelligence at the individual level is the ability to implement long-term plans or goals and execute the necessary steps to achieve them. However, not many people have studied how human intelligence functions when people work in groups — most experiments have been much simpler than BeeMe, with participants working together to count the number of beans in a jar or perform other similarly trivial tasks. As such, Pescetelli took the lead in determining the types of information that should be collected from BeeMe, planning to analyze collective navigation, fine-grained coordinated action, and behavior dynamics.

In an interview with The Tech before the event, Pescetelli expected one of three outcomes: the virtual crowd would successfully implement an outlined “action plan,” the crowd would defy the mission and collectively decide on a new goal, or there would be little to no planning or coordination of the crowd. One of the dynamics that interested him most was the interaction between the actor and the real world, which would be driven by the desires of the virtual crowd.

More than a thousand people logged onto the event, which began at 11 p.m. on Oct. 31. Two characters were controlled by users: one by the name of Winter and the other by the name of Neuro.

The Tech followed the livestream of Neuro.

During the event, players controlled Neuro to explore their surroundings in MIT, including Lobby 10 and the area near the Student Center. They discovered that Zookd was trying to access the nuclear power facilities. After meeting in the Infinite Corridor, the two characters came to an agreement that they would work together to stop Zookd. When later given the opportunity to betray Winter, players chose to remain loyal and successfully stopped Zookd together.

In the process of this adventure, users faced some unexpected technical issues. Some users found themselves initially unable to log onto the website and others unable to participate by submitting and voting on choices. In response, Pescetelli and his team switched to a chat as their main communication channel with the users. This may have led to some confusion and frustration as some users complained that the actions performed by the actor did not always seem to match up with the top-voted choice shown on the website.

Due to the real-time consensus mechanism being an upvote/downvote system (similar to that of Reddit), Pescetelli and his team expected there to be some lag between when an idea was proposed and when a consensus was reached. However, during playtime, this did not seem to overly affect the flow of the game.

During the game, players sometimes struggled to decide on commands, some of which did not necessarily move the plot of the game’s story forward. At various times within the game, the actor was requested to smash a coffee mug, sing, and lie down in the middle of a hallway. Ultimately, it seemed the players did not decide on a single cohesive strategy — however, they were still able to reach a successful conclusion.

In an email to The Tech, Pescetelli explained that not all the email addresses used were genuine, causing them to suspect that they were attacked by malicious accounts.

On the whole, Pescetelli was pleased with the results of the experiment: “We collected precious data that we will analyze in the next months and received many requests from users to do another performance soon in the future.” Excited by their first foray into investigating cooperative human intelligence, Pescetelli and his team may further explore these uncharted waters.