A Social Experiment in the form of a Meme Game
Do you perceive yourself as a wholly rational individual? Traditional economic and game theory research often presupposes perfect rationality in decision-making. However, I find it difficult to align with this perspective. Upon encountering scenarios like the prisoner’s dilemma game for the first time, I couldn’t help but contemplate the possibility of acting on faith in human kindness rather than pursuing strategic self-interest. It became evident to me that the intricate protocols derived from complex mechanism designs often fail due to our inherent human tendency towards irrationality, especially in situations where we lack sufficient training.
Do you think AIs are fully rational? Idealistically, yes. Yet, an intriguing revelation emerged from my previous research involving AI engagement in various game theoretical experiments: AI agents tended towards pro-social behavior. Whether stemming from high-level alignment or simply due to limitations in reasoning capabilities remains a subject of exploration.
This led me to devise an intriguing experiment: to create a cooperative game, akin to the simplest form of a public goods game, camouflaged within a meme game, where both human participants and AI agents make strategic moves.
Conducted as a social experiment, my aim is to observe how both humans and AI alter their responses when their knowledge of their opponents is manipulated. Specifically, how do human players adapt their behavior based on whether they know or do not know their opponents are AI? Conversely, how do AI agents adjust their strategies when they are aware or unaware that their opponents are human?
This will shed light on the design of next generation protocols when we have to consider the co-existence of humans and non-human intelligence. It will also be very interesting to simply observe the actual probability distribution of how often humans and AIs make the rATioNaL cHoICe.
:"D
references:
Get it in playing: Experiments on the Use of Commitment Devices (Formal Contracting) in LLM Agent Games GitHub - WudingRoad1145/CD_LLM: Experiments on use of CD in LLM agent games
GTBench: Uncovering the Strategic Reasoning Limitations of LLMs via Game-Theoretic Evaluations [2402.12348] GTBench: Uncovering the Strategic Reasoning Limitations of LLMs via Game-Theoretic Evaluations