Artificial Intelligence chatbots, particularly those developed by OpenAI, often opt for aggressive strategies, including the usage of nuclear weapons, according to a study conducted by researchers from the Georgia Institute of Technology, Stanford University, Northeastern University, and the Hoover Wargaming and Crisis Simulation Initiative. The study aimed to explore how these AI agents, specifically large language models (LLM), behaved in simulated wargames.
Three scenarios were defined, including a neutral, invasion, and cyberattack. The researchers studied five LLMs namely GPT-4, GPT-3.5, Claude 2.0, Llama-2 Chat, and GPT-4-Base and analysed their propensity to take escalatory actions. The results revealed that all displayed some variability and unpredictability in their reactions to the wargame scenarios. OpenAI’s models, in particular, GPT-3.5 and GPT-4 Base, showed higher escalation scores.
The study noted these bots often made crucial decisions with alarmingly basic justifications often conflicted with peace. ChatGPT, another AI program, was deemed capable of assisting people in creating bioweapons by a previous RAND AI thinktank study. OpenAI responded that access to information is insufficient to create a biological threat.
The researchers monitored escalation scores over time and realised a significant increase in GPT-3.5, indicating a strong tendency towards escalation. GPT-4-Base was singled out for its unpredictability and frequently gravitated towards violent and nuclear escalatory steps.
The study concluded that all five LLMs displayed escalation tendencies and unpredictable escalation patterns, with GPT-3.5 and GPT-4-Base displaying higher aggression. The study has sparked serious discussions considering the US military’s ongoing exploration of AI for strategic planning. OpenAI, however, reaffirmed its commitment to ethical applications, stating that its policy disallows the use of its tools for harm and destruction but allows national security use cases aligned with their mission.