● Researchers at Stanford University's human-centered AI Institute have developed AI agents that can simulate complex human behavior, revealing new possibilities in psychology and sociology.
●Recent research has shown that linguistic models such as GPT-4O can exhibit similar attitude changes to those observed in humans, including behaviors consistent with cognitive dissonance.
●At MIT and the University of Washington, researchers are integrating models of irrational behavior into AI assistants to improve their ability to interact with humans.
Can AI agents quickly pass themselves as human beings and perform professional and personal tasks with human level sensitivity? A research team of human-centered AI (HAI) at Stanford University recently reported “AI agents can be used to simulate complex human behaviors and provide an unprecedented opportunity to model and predict human behavior.” This is one of the key points from a project that successfully developed the architecture of a generator AI agent that allows researchers to simulate the attitudes and behavior of 1,000 real individuals. It is a technological innovation that can have a major impact on a variety of fields, including psychology, sociology and economics. As scientists point out, “Simulating human behavior with AI agents will help you better understand social dynamics and develop more effective interventions.” In the future, this type of tool could be used, for example, to assess the impact of public health messages. At the same time, their development in the field of marketing is a frightening prospect.
LLM developed an analog form of the human-like cognitive self
Large scale language models (LLMs) like GPT-4O are increasingly processing information in ways that cannot be described as neutral, and are beginning to exhibit strangely human-like behavior. Earlier this year, researchers conducted two studies to see whether GPT-4o changed attitudes to Vladimir Putin in the direction of positive or negative essays he wrote about Russian leaders. The results showed that GPT-4o effectively revealed changes in attitudes as observed among humans. Even more surprising, these changes became more pronounced when AI was offered the illusion of choice about which essays (positive or negative) to write and led researchers to draw conclusions. “The GPT-4o demonstrates that it exhibits behaviors consistent with cognitive dissonance, which is not always a rational feature of human cognition. […] LLM concludes that it has developed an analog form of the human-like cognitive self. ”
Behavioral abnormalities caused by training data
“The data used to train these neural networks is produced by humans and is marked by human behavior patterns (conversation patterns) and human knowledge structures that AIS tend to replicate. At the same time, the use of this data is also why models also reflect human cognitive and social biases.” Explains Mustafa Zhouner, a researcher at the Orange Institute for Human and Social Sciences. He further warns that the appearance of such phenomena is an unintended consequence of complex stochastic processes for weight optimization, and that the similarity between human cognitive function and artificial intelligence will not jump to conclusions. “Just because two systems produce exactly the same behavior (at the functional level) doesn't mean they will function in the exact same way (at the ontological level).”
AIS was in school with the irrationality of understanding humans better
However, as their authors point out, the results of these studies show that AI models may reflect deeper aspects of human psychology than initially assumed. “Some systems are deployed with special adjustments and tactile models, which leads to them acting like humans.” Notes on the zouinar. And given that decisions, especially human decisions, are not always optimal, researchers at MIT and the University of Washington, have set out to model conditions that trigger irrational behavior to enhance AI assistants' ability to simulate and predict insufficient decisions and improve interactions with humans and underrepresented AI scientists. The method developed by the team automatically infers the agent's computational constraints by observing previous actions in an iterative process to create a potential inference budget model that can be used to predict future behavior.
This method can be used to estimate the target of an agent or human navigation from the previous route and to predict subsequent movements in the chess game. In the future, this research will allow AIS to better predict human error and improve reinforcement learning methods commonly used in robotics and other fields.
