State and federal agencies are now using artificial intelligence to help decide a host of life-altering issues such as who gets investigated for a crime and how law enforcement and the military direct their immense power. In most cases, no legislature voted to allow its use, no list tells the public what AI systems are being employed to help make these decisions, and no one knows for sure who can turn them off.

Bill Gates said last month that lawmakers need to establish mandatory “safeguards and monitoring” for AI use. A consensus on comprehensive AI regulation, though, seems far away, and rules for the companies building these tools will take time. But there is no reason the government cannot begin policing its own use of AI today. Lawmakers do not have to wait to safeguard and monitor government AI use. They can do that now as they debate other ways to respond to AI.

Hundreds of new lawmakers will take office in Congress and the states in January, many of whom campaigned on doing something to protect Americans from the risks of AI. We recommend they use their oversight and legislative powers to ensure that any AI system government uses to shape decisions about people’s lives and liberty is transparent, approved and stoppable. “Transparent” means its use and the underlying specifics appear on a list available to anyone.

“Approved” means no AI system in government can go live unless it has been authorized through legislation or tested for safety and security using criteria in law. “Stoppable” means a named official answers for its use and that the decisions it makes can be undone.

Transparency matters, because government cannot see what AI tools it is using. These tools get purchased, picked up for free, bundled into software an agency already owns, or run by a contractor. Even older, non-AI systems fail and, without scrutiny, can cause widespread harm. The risks with AI’s almost unimaginable abilities are exponentially greater and demand a much higher level of public visibility.

So, agencies need to start with a public list showing where AI is being used to make decisions that affect individuals. Agencies should then be required to provide notice to any person when AI was used in a decision about them and the agency, not the person, should pay to remedy any mistakes AI makes. Legislators can require this now and use their oversight powers to ensure government agencies comply.

Specific legislation should be required to approve the use of AI systems, and it should set rules for vetting AI system capabilities, limiting their use, and ending their deployment unless they are certified to be safe. An AI bot will not necessarily refuse to carry out immoral or illegal orders, as was revealed when Nikkei Asia found 17 research papers with hidden white text telling AI reviewers to give them positive reviews. Cyber criminals or dishonest public officials can do far more harm than academics by corrupting critical government systems and decisions. The fix is to only allow AI systems that have been authorized by both the legislature and government agencies and to let legislative and agency staff see the AI logs and run their own test cases.

AI systems must be “stoppable” because flawed systems can run bad programs for years before anyone catches them. Michigan’s automated unemployment system, though not AI, falsely accused thousands of people of fraud, seizing wages and tax refunds that took years to return. AI could make such cases rarer or far more common, depending on whether a government official is accountable for identifying and stopping a failing system.

Texas, Indiana, California and Connecticut have already ordered their agencies to inventory the automated systems they use, and a pending U.S. Senate defense bill would make contractors disclose what is behind the AI they sell the Pentagon. These instances show that legislators are thinking about government’s use of AI, but they should raise their ambitions.

Legislators have the authority and responsibility to lead on this, and voters have every right to demand — before and after November — that the AI our government uses is transparent, approved and stoppable.

Jim Townsend is director of the nonpartisan Levin Center for Oversight and Democracy at Wayne State University Law School in Detroit. He is a former member of the Michigan House of Representatives and a former congressional staffer. Patrick Carter spent more than a decade in civil servant roles within Minnesota state government, including in the state budget office, and now advises state officials across the country and political spectrum on how public money is budgeted and overseen.