Georgia Tech researchers have created software with a built-in AI agent that works alongside human designers in real time to create game levels. The software, dubbed MorAI Maker in a nod to Nintendo’s game Mario Maker, uses new machine learning techniques for game content generation that allows humans and an AI agent to work in a turn-based fashion on the same digital canvas. This is the first such tool of its kind.

Through two studies with more than 100 game hobbyists and practicing game developers, the Georgia Tech team found that people varied significantly in how they used the AI.

“We did not explicitly structure any roles into our machine learning models, but we still found that users naturally projected different roles onto the same AI and took corresponding roles,” said Matthew Guzdial, Ph.D. student in computer science and lead researcher.

According to researchers, after refining the machine learning model, the AI agent was capable of picking up on users’ preferences for level structures. A majority of game developers reported that they would use the AI co-designer in the software, which was developed in Unity.

Researchers observed four major categories of roles that people assigned their virtual partners.

Some participants viewed the AI as a friend. One participant prompted the AI to begin the level design, forfeiting her own turn and stating, “Let’s see what my friend comes up with.”

Some participants wanted an equal design partner (collaborator), others seemed to expect the AI to adhere to their specific design beliefs or instructions (student), and some designers followed the AI’s lead or expected to be evaluated on their design (manager).

“Human designers in the study demonstrated a willingness to adapt their own design practices to the AI, sometimes as a means of attempting to determine how best to interact with it,” said Guzdial.

Conversely, every participant had at least one interaction where the AI adapted to the human designs. For some, this was the exception rather than the rule. “The [AI] agent placed objects fairly arbitrarily, in places where it didn’t really affect gameplay, just looked weird,” said another participating professional designer.

The AI agent embedded in the game design software was trained on implicit feedback from the user. If a user kept the AI’s game level additions, the AI received a “reward,” and if the user removed them a “penalty” was given to the AI. The AI was not allowed to remove human-generated elements.

One designer said, “It was nice to be surprised by the AI partner. It prompted conversation/discussion in my head.” Another said, “I was running out of ideas, then prompted the AI for help, and I said, ‘Oh yeah I forgot about these things!’”

Despite mostly positive feedback, not everyone found the tool to be consistently valuable. As one participant put it, “I could see using this tool as a way to give myself inspiration. But, if I had more specific goals in mind... I would have found it more inhibiting than useful.”

Guzdial says MorAI Maker is intended as a design aide, not as a replacement for designers.

“The AI system is developed in favor of augmenting, not replacing, creative work,” he said.

The full research, Friend, collaborator, student, manager: How design of an AI-driven game level editor affects creators, is published in the 2019 Proceedings of the ACM Conference on Humans Factors in Computing Systems.

The research is based upon work supported by the National Science Foundation under Grant No. IIS-1525967. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.