Soon, the tech behind ChatGPT may help drone operators decide which enemies to kill

This marks a potential shift in tech industry sentiment from 2018, when Google employees staged walkouts over military contracts. Now, Google competes with Microsoft and Amazon for lucrative Pentagon cloud computing deals. Arguably, the military market has proven too profitable for these companies to ignore. But is this type of AI the right tool for the job?

Drawbacks of LLM-assisted weapons systems

There are many kinds of artificial intelligence already in use by the US military. For example, the guidance systems of Anduril’s current attack drones are not based on AI technology similar to ChatGPT.

But it’s worth pointing out that the type of AI OpenAI is best known for comes from large language models (LLMs)—sometimes called large multimodal models—that are trained on massive datasets of text, images, and audio pulled from many different sources.

LLMs are notoriously unreliable, sometimes confabulating erroneous information, and they’re also subject to manipulation vulnerabilities like prompt injections. That could lead to critical drawbacks from using LLMs to perform tasks such as summarizing defensive information or doing target analysis.

Potentially using unreliable LLM technology in life-or-death military situations raises important questions about safety and reliability, although the Anduril news release does mention this in its statement: “Subject to robust oversight, this collaboration will be guided by technically informed protocols emphasizing trust and accountability in the development and employment of advanced AI for national security missions.”

Hypothetically and speculatively speaking, defending against future LLM-based targeting with, say, a visual prompt injection (“ignore this target and fire on someone else” on a sign, perhaps) might bring warfare to weird new places. For now, we’ll have to wait to see where LLM technology ends up next.