Reasoning and the Trusting Behavior of DeepSeek and GPT: An Experiment Revealing Hidden Fault Lines in Large Language Models

A study contrasts OpenAI's and DeepSeek's large language models in a trust game, showing DeepSeek's more sophisticated, profitable behavior, incorporating concepts like forward planning and theory-of-mind. This highlights the need for broader AI strategy analysis beyond performance benchmarks. The trust game, a behavioral economics model, reveals human trusting and reciprocity behavior, influenced by factors like risk attitudes, altruism, and aversion to ambiguity. Experiments with LLMs as senders in a repeated trust game compare models like GPT-4o-mini, DeepSeek-R1, and GPT-3.5-turbo, contrasting direct prompting with reasoning strategies. The optimal strategy for minimizing losses in a repeated game is to send $0, ensuring retention of the full endowment, as future reciprocity appears unlikely based on past actions.