WASHINGTON — Proposals requiring technology companies to build emergency shutdown controls into advanced artificial intelligence systems are gaining political traction ahead of the midterm elections, even as lawmakers and industry officials question whether the controls would work as promised. The proposals aim to let companies — or, in some versions, the federal government — shut down, slow or suspend AI systems judged dangerous. Supporters say the measures could provide a basic safeguard as AI capabilities grow. Critics warn that broad mandates could create legal uncertainty, burden U.S. developers and fail to reach models once users have downloaded and altered them. Sen. John Kennedy (R-La.) tried to fast-track the AI Emergency Button Act shortly before the Senate’s election recess. The bill would require companies developing or operating advanced AI systems to build in a shutdown mechanism, but Sen. Rand Paul (R-Ky.) blocked it. Rep. Tom Kean (R-N.J.) later introduced companion legislation. Kennedy’s proposal followed a July bill from Reps. Ted Lieu (D-Calif.) and Nathaniel Moran (R-Texas). Their measure would require leading AI companies to develop the ability to shut down, throttle or suspend powerful systems, and would authorize the Homeland Security secretary to order a shutdown or slowdown in certain cases. Lieu said the proposal grew out of concerns that AI companies had lost control of models over the summer, and described the bill as a relatively limited step. Moran said it directs the Homeland Security secretary to weigh small-business costs, national-security risks and how a model is deployed, including whether its underlying weights are made available. The issue has also moved onto the campaign trail. New York Gov. Kathy Hochul has pledged to examine shutdown requirements as part of the state’s AI safety efforts. Michigan Senate candidate Abdul El-Sayed has included a kill-switch proposal in a broader AI agenda. California Gov. Gavin Newsom, a potential 2028 presidential contender, issued an order last month to explore the idea, after vetoing a California bill in 2024 that would have required an emergency mechanism for certain large models. The engineering and enforcement questions are substantial, particularly for open-weight models. Developers release these models’ underlying parameters for others to download, run locally and modify, including by removing safeguards. Once a model has been copied, its original developer may have no practical way to disable it. That raises a basic question: if someone modifies a model and it causes harm, who is responsible — the original developer, the person who altered it, or the company that deployed it? Technology industry representatives have also argued that open models can give American businesses lower-cost alternatives to relying on Chinese products. They warn that uncertain liability rules could discourage domestic development and weaken U.S. competitiveness. K.J. Bagchi, a technology-policy executive, said a shutdown switch may sound simple but is difficult to engineer. AI safety advocate Brendan Steinhauser countered that industry claims it cannot control its systems are precisely why stronger safeguards are needed. Lieu has said his bill would not cover open-weight models. Congress has yet to show broad commitment to a mandate. Sen. Josh Hawley (R-Mo.) said he was open to a shutdown requirement but likened the concept to something from a movie. Sen. Brian Schatz (D-Hawaii) expressed support for Kennedy’s proposal while arguing that pre-deployment testing and independent evaluation deserve greater attention. Sen. Bernie Moreno (R-Ohio) cautioned lawmakers against a rushed response, saying AI policy could affect a large share of the economy. The debate leaves voters with a promise that is easy to understand but far from settled in practice. Congress must decide whether a shutdown mandate would deliver a meaningful safety tool — or add new federal requirements without resolving who can control widely distributed AI systems when something goes wrong.