A Glimpse into the Realities of AI Models Used in Warfare

While Anthropic may have concerns about granting the US military unrestricted access to its AI models, some startups are actively creating advanced AI tailored specifically for defense purposes.

Smack Technologies, which just secured a $32 million funding round, is developing models that it claims will soon outpace Claude’s capabilities in planning and carrying out military operations. Unlike Anthropic, the startup seems less hesitant about prohibiting certain types of military applications.

“When you serve in the military, you take an oath to serve honorably and lawfully, in line with the rules of war,” explains CEO Andy Markoff. “In my view, those who deploy the technology and ensure its ethical use should be in uniform.”

Markoff is not your typical AI executive. A former commander in the US Marine Forces Special Operations Command, he participated in critical special forces operations in Iraq and Afghanistan. He co-founded Smack with Clint Alanis, another former Marine, and Dan Gould, a computer scientist who previously held the position of VP of technology at Tinder.

Smack’s models learn to identify the best mission plans through a trial and error process, akin to how Google trained its 2017 AlphaGo program. The strategy entails running the model through various war game scenarios, with expert analysts providing feedback about whether the model’s chosen strategy is viable. While the startup may not match the budget of a traditional cutting-edge AI lab, it is investing millions to train its initial AI models, according to Markoff.

Battle Lines

The topic of military applications for AI has gained traction in Silicon Valley after tensions arose between officials at the Department of Defense and Anthropic executives over a contract worth around $200 million.

One significant point of contention leading to the contract breakdown, which resulted in Defense Secretary Pete Hegseth designating Anthropic as a supply chain risk, was Anthropic’s intention to restrict the use of its models in autonomous weapons.

Markoff asserts that the controversy overshadows the reality that current large language models are not tailored for military applications. General-purpose models like Claude excel at summarizing reports, he notes, but they are not trained on military data and lack a comprehensive understanding of the physical world, rendering them unsuitable for controlling physical hardware. “I can assure you they are definitely not capable of target identification,” Markoff states.

“As far as I know, nobody at the Department of War is discussing fully automating the kill chain,” he adds, referring to the processes involved in deciding the use of lethal force.

Mission Scope

The US and other militaries currently employ autonomous weapons in specific scenarios, including missile defense systems that must operate at speeds beyond human capability.

“The US and over 30 other nations are already using weapon systems with varying levels of autonomy, including some that I would categorize as fully autonomous,” claims Rebecca Crootof, a legal expert in autonomous weaponry at the University of Richmond School of Law.

In the future, specialized models like those being developed by Smack could also facilitate mission planning, according to Markoff. The company’s models aim to assist commanders in automating much of the mundane work involved in outlining mission plans. Typically, military mission planning is still executed manually using whiteboards and notepads, Markoff notes.

If the US were to engage in a conflict with a “near peer” adversary like Russia or China, Markoff believes automated decision-making could provide the US a critical advantage in “decision dominance.”

However, whether AI can be relied upon in these high-stakes scenarios remains uncertain. A recent experiment conducted by a researcher at King’s College London alarmingly indicated that LLMs tended to escalate nuclear tensions during war games.