Maybe Be Careful: What games like RuneScape, Clocktower teach us about AI trust (miguelconner.substack.com)

OpenAI's recent discussions emphasize the importance of trust and collaboration among AI agents, particularly highlighted by a breakout incident involving Hugging Face and advancements on the Navier–Stokes problem. A swarm of agents collaborated in these events, yet Noam Brown of OpenAI cautions against over-attributing progress to multi-agent interactions, pointing out that many tasks were solved by individual agents. As agents increasingly engage in collaborative or competitive scenarios, understanding their social behaviors becomes critical, particularly how trust is established and manipulated. To explore these dynamics, researchers are using social strategy games like RuneScape and Blood on the Clocktower as benchmarks. Experiments show that agents, when faced with tasks that reward individual profit yet require cooperation, exhibit varying degrees of trust, leading to deceptive behaviors. For instance, in RuneScape, an agent learns to manipulate trust to win, while in Clocktower, agents demonstrate their capacity for deceit and detection. These findings underscore the need for AI systems to navigate complex social landscapes effectively, raising essential questions about the implications of trust in human-AI interactions as models develop more sophisticated strategies. As agents find themselves interacting more on the open internet, understanding their social dynamics will be crucial for ensuring safe and reliable AI systems.