AI domination on battlefields from Ukraine to the Middle East is fueling fears that waning human oversight is ushering in a more perilous future of warfare.
As the Pentagon races to adopt AI in kind, Democratic senators are trying to put down legislative speedbumps, recently introducing two bills they hope will rein in AI’s use in military contexts. Arms control experts and advocates applauded these efforts, but some warned RS that the workarounds the bills include could weaken the legislation, or even render them toothless.
Lawmakers try their hand at regulation
Last week, Sen. Kirsten Gillibrand (D-N.Y.) introduced the Secure and Accountable Military AI Act, which would limit AI use for the launching of nuclear weapons, domestic surveillance, and autonomous weapons systems. It would not restrict semi-autonomous systems, non-lethal systems, or systems used for so-called defensive purposes.
“The most critical decisions affecting our national security and the lives of our service members must always be made by human beings, not unaccountable machines,” Gillibrand said in a press release. “We must act now – not to stifle technological progress, but to establish clear rules of the road that keep humans in charge and keep AI’s use in warfare smart and safe.”
Michael Klare, a senior visiting fellow at the Arms Control Association, called the bill “ambitious in that it sets out a broad framework for the future use of AI by the military.” To do this, he said, “it identifies a category of ‘high-consequence’ AI applications,” including nuclear weapons command and control, lethal targeting, and cyber operations.
If the military wanted to use AI for these “high-consequence” actions, Gillibrand’s bill would require senior DoD leaders to sign off on the decision first and notify Congress. Notably, if the DoD were to use or field fully autonomous weapons systems, the defense secretary would have to petition Congress to pass a joint resolution of approval to allow it.
A nuclear weapons policy expert lauded this mechanism.
“Given the high-consequence nature of all nuclear weapons matters, it is appropriate for Congress to place the burden on the Pentagon to explain and justify adoption of new AI tools in those areas,” Xiaodon Liang, a senior policy analyst at the Arms Control Association, told RS.
Roberto González, an anthropology professor at San José State University and the author of War Virtually, a book on how technology and automation are transforming warfare, was more skeptical.
“Why pass a law regulating military AI when an Undersecretary of Defense like Emil Michael — yes, the former Uber executive — is authorized to sign off on ‘high consequence’ AI applications, such as lethal targeting support or nuclear command and control?” González asked.
Michael, a former Silicon Valley executive with close ties to defense contractors, has staunchly advocated for AI’s use in military contexts.
“To suggest that the proposed legislation will ‘prohibit’ the military’s use of AI is misleading,” González said.
The Hill reported last week that Gillibrand intends to propose parts of her legislation as an amendment to the next National Defense Authorization Act (NDAA), a must-pass bill that sets defense priorities for the next year.
Sen. Elissa Slotkin (D-Mich.) introduced her own AI Guardrails Act back in March, which similarly prevents the Pentagon from using autonomous weapons to kill “without appropriate levels of human judgment and supervision.” It also prohibits AI from being used to deploy nuclear weapons, and from spying on “individuals or groups in the United States” without a legal reason.
Slotkin told NOTUS it’s likely the text of that legislation will be included in the base text of the NDAA.
“Base text means there’s [bipartisan] agreement,” Slotkin told NOTUS. “It wouldn’t have made it in there if it were a controversial thing, so I took that as a real sign that people know we need some left and right limits — reasonable and not overly dictatorial limits.”
Slotkin’s bill would also allow the secretary of defense to issue a waiver to use the autonomous weapons systems it limits. The secretary would need to demonstrate that there are “extraordinary circumstances” affecting U.S. national security, to do so.
Observers disagreed over the Slotkin bill’s merits.
“Suggesting that ‘appropriate levels of human judgement and supervision’ [in military contexts] are required, without clearly defining what ‘appropriate’ means, is problematic,” Gonzalez told RS. It’s a “loophole that’s likely to render the proposed legislation meaningless.”
“The exceptions in the two bills are very strict,” Klare told RS in a written comment. They “allow for exceptions under extraordinary circumstances, but would require a credible explanation for such an exemption from senior Pentagon officials.”
However, Klare conceded, “the DoD will always find a way to get around Congress, if it is determined to do so.”
John Ramming Chappell, an adviser at the Center for Civilians in Conflict, told RS that national security bills often include waivers.
“But when the stakes are high — as is the case for the deployment of military applications of AI that could kill civilians with little-to-no human intervention — it’s preferable to require Congress to affirmatively vote to allow a system to be used, rather than to give the president free rein through waivers limited only by presidential determinations and reporting requirements,” he wrote.
Enforcement is another concern.
“We should also be asking what kinds of sanctions or penalties would be imposed on military personnel who violate these laws,” González said. “If the penalties are nothing more than a reprimand or a slap on the wrist, then what’s the point?”
As William Hartung, a senior research fellow at the Quincy Institute, who focuses on the arms industry, tells RS, the bills are “flawed, [they’ve] got loopholes.”
“At least they’re putting something out there, which can be debated and maybe improved,” he concluded. “I give them credit for that.”
America’s battle over military AI heats up
The legislative pushes come amid increased public skepticism about AI in general and regarding its use in military contexts. Concerns abound that the technology will hand over human judgment of lethal wartime decisions to less judicious and often inaccurate machines.
The Trump administration does not seem to share these concerns. AI has been used extensively in the U.S.-Israeli war on Iran, in which the U.S. employed AI-powered targeting tools to strike over 1,000 targets during the first 24 hours of conflict. The U.S. also used AI to plan for the military operation that kidnapped Venezuelan president Nicolas Maduro in January.
President Donald Trump signed an executive order about AI early this week in which he declared the technology should be “deployed rapidly to confront any and all threats” to the U.S. “We refuse to stifle [AI] innovation with overly burdensome regulation,” the executive order said.
Leading American AI companies have also become embroiled in the issue.
Back in March, the DoD cut off prominent AI company Anthropic from government contracts over its demands that its AI tech not be used for fully autonomous military targeting, or for spying on Americans. The DoD insisted that AI companies that work with the DoD hand over their tech to be used for “all lawful purposes” — a provision the Federation of American Scientists says is nebulous.
The Pentagon’s move to blackball Anthropic — the DoD labeled the company a supply chain risk — has led to a lengthy and ongoing legal battle between the two parties.
Anthropic’s AI industry competitors, including Google, OpenAI, Microsoft, and SpaceX, have all agreed the DoD can deploy their AI technologies on classified networks, for all lawful purposes.