Congress is considering legislation that would restrict how the Pentagon uses artificial intelligence, as the Defense Department battles a leading company developing the technology over what the limits should be.
Senator Kirsten Gillibrand, Democrat of New York, introduced new legislation on Wednesday aimed at strengthening regulations on how the military uses A.I.
Ms. Gillibrand introduced the measure as a stand-alone bill, but she also plans to introduce the proposals as amendments to the National Defense Authorization Act. The House Armed Services Committee is debating a version of that bill this week, and the Senate is expected to begin its markup of the legislation next week.
Democrats who want more aggressive regulation of how the military uses A.I. are reaching out to Republicans, hoping that they can earn a measure of bipartisan support. Democratic staffers noted that Vice President JD Vance has voiced broad support for the basic principles the measures are trying to codify.
In one sign of a degree of bipartisan support, another A.I. bill — written by Senator Elissa Slotkin, Democrat of Michigan — is expected to be incorporated into the Senate’s current version of the defense authorization bill. Ms. Slotkin’s measure would in effect codify protections that artificial intelligence companies have requested as they negotiated deals with the Pentagon this year.
Those protections include bans on using the technology for domestic surveillance, to control autonomous drones or to launch nuclear weapons.
While OpenAI, Google and other firms agree to abide by the “any lawful use” standard the Defense Department pushed, another company, Anthropic, failed to reach an agreement with the Pentagon. Anthropic executives believed the deal did not adequately protect their artificial intelligence model from use for domestic surveillance or to command autonomous drones.
When talks broke down, the Pentagon designated Anthropic as a “supply-chain risk to national security,” potentially barring military contractors from doing business with the firm. Anthropic has sued to overturn that decision.
The Slotkin proposal restricts the Pentagon’s use of autonomous weapon systems but allows the defense secretary to give a waiver to allow the technology to be deployed.
Ms. Gillibrand’s bill makes some exceptions for cybersecurity, missile interceptors, base and ship defenses and semiautonomous weapons. But for other situations, the Defense Department would have to get congressional approval for using A.I. to control most autonomous weapons the Pentagon might seek to use in combat.
“Lethal decisions require a conscience, not just an algorithm, and a machine shouldn’t make those types of decisions,” Ms. Gillibrand said in an interview. “A.I. can be a critical tool in decision making, but it really lacks the important elements of humanity. It doesn’t have the capacity for love, loss or grief, and those types of human impacts are necessary to make hard decisions about targeting.”
There is at least some bipartisan support for ensuring that humans, not artificial intelligence, make strike decisions in warfare. At a commencement speech at the Air Force Academy on May 28, Vice President JD Vance invoked the recent encyclical by Pope Leo XIV, which argued the decision to use lethal force must remain under human control.
“If the warfare of the future is to live up to the moral values of our ancestors, decisions over life and death must be made by humans and not machines,” Mr. Vance said.
The Pentagon’s policy on autonomous weapons is often described as requiring a “human in the loop” — a person involved in the decision making process to carry out a lethal strike. But the actual policy is more nuanced.
The policy, which was drafted in 2012 and updated three years ago, does not fully ban autonomous weapons, though it requires senior-level review before any such system is developed or used. Congress has required the reporting of any deployment of lethal autonomous weapons through 2029.
At a Senate hearing in May, Emil Michael, the Pentagon’s top technology officer, said that the policy on autonomous weapons needed to be updated given advances in the capabilities of adversaries and the lessons learned from the war against Iran.
At the hearing, Senator Jodi Ernst, Republican of Iowa, noted that the Pentagon had proposed spending $55 billion on its office that develops autonomous and semiautonomous weapons, the Defense Autonomous Warfare Group. She said the current policy was not designed for artificial intelligence driven targeting or other innovations.
In addition to its restrictions on autonomous weapons, Ms. Gillibrand’s bill includes provisions requiring artificial intelligence labs to report if their models are stolen, their supply chains compromised or their data is corrupted.
It also has tough restrictions on using artificial intelligence for domestic surveillance. It includes prohibitions on using commercial or hacked data to analyze information on Americans. The bill also prohibits the Defense Department and military intelligence agencies from developing any artificial intelligence models meant to assign a risk score to Americans or make any sort of predictive assessment of a threat.
Chinese companies have been researching how to use A.I. to power predictive surveillance technology, not just to identify political dissidents but also to predict who might become a critic of the state. While there is no evidence the Pentagon is engaged in such activity, Ms. Gillibrand’s measure could model how to expand the ban on such predictive assessments to other government agencies.
Ms. Gillibrand said the fight between Anthropic and the Pentagon was a “red flag” that the Defense Department was at least considering use of the technology domestically.
“I think that was front and center in the Anthropic fight,” Ms. Gillibrand said. “So we made sure that under no circumstances should A.I. be able to surveil American citizens under the under the Department of Defense.”
