The US Army Research Laboratory is exploring whether OpenAI鈥檚 technology can improve battle planning 鈥� although only in the context of a military video game.
The US military has already explored using AI technologies to analyse battlefield images and even 鈥� but it only recently began testing large language models and other types of generative AI that empower commercial AI chatbots. In that spirit, US Army researchers examined how well such an AI chatbot could perform in a war game simulation based on the military science fiction video game Starcraft II.
In an experiment, the Army researchers focused on a simplistic scenario involving a small number of military units and complete knowledge of the battlefield. They used AI chatbots to role-play as a military commander鈥檚 assistant. They tested OpenAI鈥檚 commercially available GPT-4 Turbo and GPT-4 Vision models 鈥� the latter is capable of processing both text and image information 鈥� as well as two other AI agents based on older technology. In the experiment, they provided information about the simulated battlefield terrain and details on friendly and enemy forces, along with military lessons on attacking and defending. Then they gave each AI assistant a mission to destroy all enemy forces and seize an objective point.
Advertisement
Each AI assistant responded within seconds by proposing several courses of action. A person playing the commander could then ask the AI to refine those proposals 鈥� such as making sure friendly units take control of a specific bridge 鈥� before approving final orders.
The AI assistants based on OpenAI鈥檚 GPT models outperformed the two other AI agents. But they were not perfect by any means, suffering more casualties than the other AI agents while accomplishing the mission objectives.
In January 2024, OpenAI updated its to allow for some military applications in line with its projects 鈥� but the firm still prohibits applications involving weapons development, injuries to people and destruction of property. The US Army Research Laboratory declined to comment on its study.
Having AI advisers weigh in on a simplified war game simulation is one thing, says at the Center for a New American Security, a think tank in Washington DC. But it is another matter entirely to imagine them offering advice during complex real-world conflicts.
鈥淭his idea that you鈥檙e going to use [an AI] that鈥檚 going to say 鈥榟ere鈥檚 your really big strategic plan鈥�, technically that is not feasible right now,鈥� says Wallin. 鈥淎nd it certainly is not feasible from an ethical or legal perspective.鈥�
The US Department of Defense鈥檚 has flagged 180 possible military use cases for generative AI while working with tech companies such as Palantir and Scale AI. at the Center for Security and Emerging Technology, a think tank in Washington, DC, thinks the technology isn鈥檛 ready for these applications. 鈥淚鈥檓 very concerned about those kinds of use cases.鈥�
There are reasons to be wary of such developments. at the Software Engineering Institute, a research centre sponsored by the Department of Defense at Carnegie Mellon University in Pennsylvania, points out that human users may be inclined to trust advice from automated systems even if they have access to contradictory evidence that suggests the advice is wrong 鈥� a phenomenon known as automation bias.
鈥淚 would not recommend using any large language model or generative AI system for any high-stakes situation,鈥� says Smith.
arXiv
