Gamers, rejoice: you can now tell your parents that gaming is a productive use of your time. Thanks to computer science, you can help process genomic data and train Artificial Intelligence (AI) to aid real-world medical research and one day cure diseases – all with a controller in your hand.
It’s happening in the virtual world of Borderlands 3, an online action game developed by 2K Games. It launched in 2020 and continues to welcome around 3000 active online players per day. Players are free to explore the in-game world as they see fit and can choose to participate in any number of high-intensity challenges. And when they need a break, they can wind down with a simple puzzle mini game, designed to help process genomic data.
Jérôme Waldispühl, a professor of computer science at McGill University, co-designed the mini game known as Borderlands Science. The goal is simple: players must organize a group of coloured tiles on a grid, similar to Tetris or Candy Crush, by sliding them around. Each tile corresponds to genomic data from a massive DNA sample set. As they complete puzzles, players are teaching AI to recognize genomic abnormalities.
This kind of crowdsourcing solves a big problem in AI training. AI models are only as good as the data they’re built on, and they require huge datasets to learn from. It’s a particular challenge, because many applications require human input to label data or identify errors made by the AI to help train it. This is normally an arduous task. But if millions of gamers can help through a gamified version of the task, it reduces this bottleneck in AI development.
Gamification is perfect for non-scientists who want to participate in the scientific process. But in this case, the genomic mini game is competing for attention against high-octane boss fights, alien car chases, and action-packed side quests. How do you convince gamers to choose tile-organizing puzzles instead?
That’s where Professors Setareh Farajollahzadeh and Rob Glew come in, both experts in operations management from McGill University. They co-authored a research paper observing why people play and – this is important – keep playing the game. Their findings offer valuable insight into how gamification can be used as part of the supply chain of data for AI models, everywhere from medicine to mapping.
Show me the loot
For Farajollahzadeh and Glew, Borderlands Science provided the perfect environment to study why people volunteer their time to complete certain tasks. In this case, the task was to complete puzzles linked to specific genomic sequences. In other contexts, volunteers might label mapping data to help humanitarian relief efforts (e.g. OpenStreetMap) or identify photos of birds from their back yard or local park to help train ecology AI (e.g. Birda). If managers can figure out why volunteers give their time and energy, they can design tasks that are both enjoyable and have a wider benefit to society.
When players first enter the Borderlands Science minigame to start solving DNA data puzzles, they watch a short explainer video. The narrator – Mayim Bialik, made famous by The Big Bang Theory — describes where the data came from, what it’s being used for, and how players’ participation can help advance important scientific research.
“They really framed it around the science,” said Farajollahzadeh.
The video is an appeal to people’s intrinsic motivation, she said – the drive to engage in an activity for its own sake, stemming from their enjoyment of the game and connection to the scientific mission. The hope was for players to complete the puzzles because they truly wanted to help.
To sweeten the deal, 2K Games also awarded in-game rewards to players who completed enough puzzles, such as special outfits for their avatars.
In this case, intrinsic motivation and the promise of a prize were enough to draw in players, said Farajollahzadeh. But, as she and Glew discovered, boredom can outweigh these motivations, even if volunteers believe in the mission (medical research) and receive some kind of reward (in-game loot). This means managers must find additional ways to keep volunteers engaged.
This is particularly challenging in settings like Borderlands Science, because the best training data for AI models requires both a high quality and quantity of annotations. So you need people who can complete increasingly complex puzzles without losing motivation.
AI-powered flow state
Perhaps unsurprisingly, the gamers in Borderlands 3 were very much motivated by fun and a good challenge, discovered Farajollahzadeh and Glew. Grouped with other motivators, this helped increase players’ engagement – eventually improving the quality of the data.
Farajollahzadeh and Glew were most interested in the minimum scores players had to reach before submitting a puzzle. These targets themselves were set by another AI model. Farajollahzadeh and Glew wanted to answer a simple question: just how hard should the targets make the gamers work to solve the puzzles?
If the AI makes the targets impossibly high, the gamers will become too frustrated and give up. But making the targets too easy also pushes players away, even more than higher difficulty levels.
Glew attributes this, in part, to flow—or the lack thereof. Flow is the mental state in which a person becomes fully engaged in an activity. It happens when a person’s task captures their interest and sufficiently matches their skill level.
“We want the AI to set the targets so that tasks are just hard enough that the gamers are working, but it’s not so hard that it’s unpleasant for them,” said Glew.
For instance, as Farajollahzadeh and Glew discovered, if the difficulty of the puzzles doesn’t progress at the same speed as the players’ skill level, their engagement drops off. Even when players complete every level in the game and earn every reward, if the targets are set just right, they may keep playing many 100s more puzzles, simply because they’re enjoying it.
For Farajollahzadeh and Glew, this was a critical finding: setting the right targets (using AI) helps people enjoy solving the tasks, leading to more and higher-quality training data for a different, complex scientific AI model. The right difficulty allows gamers to find their flow state and help develop new technology—all while having a good time.
Managers of these platforms can carefully apply this approach to help improve both the experience for the gamers and the quality of the AI training data they produce.
In this case, the gamers in Borderlands 3 shine a light on what motivates people to participate, turning their own gaming pastime into a productive output for science. As new, increasingly niche applications of AI develop, this research shows how gamers could help push the tech forward.
“It’s about being innovative with how you set expectations for people,” said Glew.
This article was written by Eric Dicaire, Delve’s Managing Editor. Inspired by the research paper, “Raising the Bar: Motivating Contributors in AI Assisted Crowdsourcing,” by Setareh Farajollahzadeh and Rob Glew.
Featured experts
Setareh Farajollahzadeh
