NYU researchers have created an AI model that mimics how humans generate playful goals. This breakthrough could revolutionize AI’s ability to understand and align with human intentions and improve game design.
In a groundbreaking development, researchers from New York University have designed a computer model capable of generating human-like goals by learning from the way people create games. This innovation, published in the journal Nature Machine Intelligence, holds the promise of advancing AI systems to better understand and replicate human intentions.
The study provides a new perspective on goal-setting, a fundamental aspect of human behavior often beginning in childhood.
Lead author Guy Davidson, a doctoral student at NYU, emphasized the significance of this research.
“While goals are fundamental to human behavior, we know very little about how people represent and come up with them — and lack models that capture the richness and creativity of human-generated goals,” Davidson said in a news release. “Our research provides a new framework for understanding how people create and represent goals, which could help develop more creative, original and effective AI systems.”
The challenge in the AI field has long been to replicate the subjective and nuanced nature of human goals. Previous experimental and computational efforts fell short, leaving a gap that the NYU team aimed to fill.
The team began their research with a series of online experiments where participants were asked to generate various playful goals within a virtual environment filled with objects.
The participants invented nearly 100 different games, describing scenarios such as bouncing a ball into a bin or building towers from wooden blocks. These descriptions, guided by common-sense principles and creative recombination of gameplay elements, formed the dataset from which the AI model learned.
An intriguing finding from the study was the simplicity behind seemingly complex goal-setting. Despite the variety of games, all goals were fashioned from a limited set of principles: plausibility (physical feasibility) and recombination (mixing basic gameplay elements in novel ways).
To evaluate the AI’s ability to replicate human creativity, the researchers asked a different group of participants to rate both human and AI-generated games on fun, creativity and difficulty.
The AI-generated games, which included scenarios like throwing dodgeballs to land on a high shelf, received ratings comparable to those created by humans.
This compelling evidence indicates that the AI model successfully captured the essence of human goal creation. Moreover, it opens up new avenues for developing AI systems that can assist in designing more human-like games and potentially other creative endeavors.
Such advancements could not only enhance the field of AI but also offer practical applications in game design, providing tools for creating engaging and realistic games that resonate with human users. As AI continues to progress, understanding and replicating human goal-setting will remain a vital component in making these systems more intuitive and aligned with human behavior.