Using artificial intelligence (AI) for warfare has been the promise of science fiction and politicians for years, but new research from the Georgia Institute of Technology argues only so much can be automated and shows the value of human judgment.

“All of the hard problems in AI really are judgment and data problems, and the interesting thing about that is when you start thinking about war, the hard problems are strategy and uncertainty, or what is well known as the fog of war,” said Jon Lindsay, an associate professor in the School of Cybersecurity & Privacy and the Sam Nunn School of International Affairs. “You need human sense-making and to make moral, ethical, and intellectual decisions in an incredibly confusing, fraught, scary situation.”

AI decision-making is based on four key components: data about a situation, interpretation of those data (or prediction), determining the best way to act in line with goals and values (or judgment), and action. Machine learning advancements have made predictions easier, which makes data and judgment even more valuable. Although AI can automate everything from commerce to transit, judgment is where humans must intervene, Lindsay and University of Toronto Professor Avi Goldfarb wrote in the paper, “Prediction and Judgment: Why Artificial Intelligence Increases the Importance of Humans in War,” published in International Security.

An example Lindsay and Goldfarb highlight is the Rio Tinto mining company, which uses self-driving trucks to transport materials, reducing costs and risks to human drivers. There are abundant, predictable, and unbiased data traffic patterns and maps that require little human intervention unless there are road closures or obstacles.
War, however, usually lacks abundant unbiased data, and judgments about objectives and values are inherently controversial, but that doesn’t mean it’s impossible. The researchers argue AI would be best employed in bureaucratically stabilized environments on a task-by-task basis.

“All the excitement and the fear are about killer robots and lethal vehicles, but the worst case for military AI in practice is going to be the classically militaristic problems where you’re really dependent on creativity and interpretation,” Lindsay said. “But what we should be looking at is personnel systems, administration, logistics, and repairs.”