AI System Foretells Human Actions with Remarkable Precision, According to Scientists
In a groundbreaking development, researchers at New York University have unveiled Centaur, an AI model designed to predict and simulate human thought and behaviour with unprecedented accuracy and generality. This breakthrough in cognitive modelling and psychology research could potentially reshape our understanding of human cognition.
The purpose of Centaur is multifold. It aims to accurately predict human decisions and response times across a wide range of psychological scenarios, generalize predictions to completely novel or unfamiliar decision tasks, serve as a "virtual laboratory" for understanding human cognition, and aid research in areas such as clinical psychology and understanding the effects of various mental health conditions on decision-making.
Centaur's capabilities are impressive. Trained on the world's largest known dataset of human behaviour—over 10 million decisions from 60,000 individuals across 160 different psychological experiments (Psych-101 dataset)—it predicts human choices with about 64% accuracy, surpassing previous AI and traditional cognitive models. It also accurately forecasts human response times based on cognitive complexity, capturing more variance in reaction times than both base language models and classical cognitive models.
Remarkably, Centaur's internal representations naturally align with actual human neural activity, despite not being explicitly trained on brain data. It can generalize to new domains, predicting behaviour under modified cover stories, altered problem structures, and completely novel tasks. Built on the LLaMA 3.1 70B language model, fine-tuned with parameter-efficient techniques like quantized low-rank adaptation (QLoRA), Centaur leverages a rich knowledge base and adapts efficiently.
Centaur's significance lies in its ability to model the remarkable generality of human cognition, unlike specialized models such as AlphaGo (which focus on a single domain). It offers a unified approach to understanding how people think and behave across diverse situations, promising advances in psychological theory, mental health research, and AI-human interaction.
The future of Centaur looks promising. Researchers plan to expand Psych-101 to include demographic and psychological data, such as age, socioeconomic status, and personality traits, to strengthen Centaur's training capabilities. However, the unknown factor is whether Centaur just predicts human thought or whether it reproduces the underlying cognitive processes.
Centaur could have numerous applications, including psychology and health research, predicting human behaviour in clinical contexts, prototyping and analysis of psychology experiments, and figuring out how to design experiments that demonstrate a phenomenon most clearly or require the fewest participants.
Brenden Lake, who studies psychology and data science at New York University, stated that Centaur is a step beyond in its predictive capabilities. He is most excited about the potential educational applications of Centaur, suggesting that if they can predict how a student learns and reasons through a problem, they could simulate the effects of different teaching strategies, which could be a potential game changer.
In essence, Centaur is a cutting-edge AI model with the unique ability to simulate human thought and behaviour both accurately and flexibly, making it a powerful tool for cognitive science and practical applications that require understanding or predicting human decision-making.
Science and technology have converged with artificial intelligence in the form of Centaur, an AI model designed to simulate human thought and behavior. This advancement, capable of predicting human decisions, generalizing predictions to new tasks, serving as a virtual laboratory for understanding human cognition, and even reproducing the cognitive processes under the guise of research in clinical psychology, signifies a significant leap in our understanding of human cognition and AI-human interaction.
