In 2018, the New York City Council created a task force to study the city’s use of automated decision systems (ADS). The concern: Algorithms, not just in New York but around the country, were increasingly being employed by government agencies to do everything from informing criminal sentencing and detecting unemployment fraud to prioritizing child abuse cases and distributing health benefits. And lawmakers, let alone the people governed by the automated decisions, knew little about how the calculations were being made.
Rare glimpses into how these algorithms were performing were not comforting: In several states, algorithms used to determine how much help residents will receive from home health aides have automatically cut benefits for thousands. Police departments across the country use the PredPol software to predict where future crimes will occur, but the program disproportionately sends police to Black and Hispanic neighborhoods. And in Michigan, an algorithm designed to detect fraudulent unemployment claims famously improperly flagged thousands of applicants, forcing residents who should have received assistance to lose their homes and file for bankruptcy.
Report Deeply and Fix Things
New York City’s was the first legislation in the country aimed at shedding light on how government agencies use artificial intelligence to make decisions about people and policies.
At the time, the creation of the task force was heralded as a “watershed” moment that would usher in a new era of oversight. And indeed, in the four years since, a steady stream of reporting about the harms caused by high-stakes algorithms has prompted lawmakers across the country to introduce nearly 40 bills designed to study or regulate government agencies’ use of ADS, according to The Markup’s review of state legislation.
The bills range from proposals to create study groups to requiring agencies to audit algorithms for bias before purchasing systems from vendors. But the dozens of reforms proposed have shared a common fate: They have largely either died immediately upon introduction or expired in committees after brief hearings, according to The Markup’s review.
In New York City, that initial working group took two years to make a set of broad, nonbinding recommendations for further research and oversight. One task force member described the endeavor as a “waste.” The group could not even agree on a definition for automated decision systems, and several of its members, at the time and since, have said they did not believe city agencies and officials had bought into the process.
Elsewhere, nearly all proposals to study or regulate algorithms have failed to pass. Bills to create study groups to examine the use of algorithms failed in Massachusetts, New York state, California, Hawaii, and Virginia. Bills requiring audits of algorithms or prohibiting algorithmic discrimination have died in California, Maryland, New Jersey, and Washington state. In several cases—California, New Jersey, Massachusetts, Michigan, and Vermont—ADS oversight or study bills remain pending in the legislature, but their prospects this session are slim, according to sponsors and advocates in those states.
The only state bill to pass so far, Vermont’s, created a task force whose recommendations—to form a permanent AI commission and adopt regulations—have so far been ignored, state representative Brian Cina told The Markup.
The Markup interviewed lawmakers and lobbyists and reviewed written and oral testimony on dozens of ADS bills to examine why legislatures have failed to regulate these tools.
Report Deeply and Fix Things
We found two key through lines: Lawmakers and the public lack fundamental access to information about what algorithms their agencies are using, how they’re designed, and how significantly they influence decisions. In many of the states The Markup examined, lawmakers and activists said state agencies had rebuffed their attempts to gather basic information, such as the names of tools being used.
Meanwhile, Big Tech and government contractors have successfully derailed legislation by arguing that proposals are too broad—in some cases claiming they would prevent public officials from using calculators and spreadsheets—and that requiring agencies to examine whether an ADS system is discriminatory would kill innovation and increase the price of government procurement.
Lawmakers Struggled to Figure Out What Algorithms Were Even in Use
One of the biggest challenges lawmakers have faced when seeking to regulate ADS tools is simply knowing what they are and what they do.
Following its task force’s landmark report, New York City conducted a subsequent survey of city agencies. It resulted in a list of only 16 automated decision systems across nine agencies, which members of the task force told The Markup they suspect is a severe underestimation.
“We don’t actually know where government entities or businesses use these systems, so it’s hard to make [regulations] more concrete,” said Julia Stoyanovich, a New York University computer science professor and task force member.
In 2018, Vermont became the first state to create its own ADS study group. At the conclusion of its work in 2020, the group reported that “there are examples of where state and local governments have used artificial intelligence applications, but in general the Task Force has not identified many of these applications.”
“Just because nothing popped up in a few weeks of testimony doesn’t mean that they don’t exist,” said Cina. “It’s not like we asked every single state agency to look at every single thing they use.”
In February, he introduced a bill that would have required the state to develop basic standards for agency use of ADS systems. It has sat in committee without a hearing since then.
In 2019, the Hawaii Senate passed a resolution requesting that the state convene a task force to study agency use of artificial intelligence systems, but the resolution was nonbinding and no task force convened, according to the Hawaii Legislative Reference Bureau. Legislators tried to pass a binding resolution again the next year, but it failed.
Legislators and advocacy groups who authored ADS bills in California, Maryland, Massachusetts, Michigan, New York, and Washington told The Markup that they have no clear understanding of the extent to which their state agencies use ADS tools.
Advocacy groups like the Electronic Privacy Information Center (EPIC) that have attempted to survey government agencies regarding their use of ADS systems say they routinely receive incomplete information.
“The results we’re getting are straight-up non-responses or truly pulling teeth about every little thing,” said Ben Winters, who leads EPIC’s AI and Human Rights Project.
In Washington, after an ADS regulation bill failed in 2020, the legislature created a study group tasked with making recommendations for future legislation. The ACLU of Washington proposed that the group should survey state agencies to gather more information about the tools they were using, but the study group rejected the idea, according to public minutes from the group’s meetings.
“We thought it was a simple ask,” said Jennifer Lee, the technology and liberty project manager for the ACLU of Washington. “One of the barriers we kept getting when talking to lawmakers about regulating ADS is they didn’t have an understanding of how prevalent the issue was. They kept asking, ‘What kind of systems are being used across Washington state?’ ”
Lawmakers Say Corporate Influence a Hurdle
Washington’s most recent bill has stalled in committee, but an updated version will likely be reintroduced this year now that the study group has completed its final report, said state senator Bob Hasegawa, the bill’s sponsor
The legislation would have required any state agency seeking to implement an ADS system to produce an algorithmic accountability report disclosing the name and purpose of the system, what data it would use, and whether the system had been independently tested for biases, among other requirements.
The bill would also have banned the use of ADS tools that are discriminatory and required that anyone affected by an algorithmic decision be notified and have a right to appeal that decision.
“The big obstacle is corporate influence in our governmental processes,” said Hasegawa. “Washington is a pretty high-tech state and so corporate high tech has a lot of influence in our systems here. That’s where most of the pushback has been coming from because the impacted communities are pretty much unanimous that this needs to be fixed.”
California’s bill, which is similar, is still pending in committee. It encourages, but does not require, vendors seeking to sell ADS tools to government agencies to submit an ADS impact report along with their bid, which would include similar disclosures to those required by Washington’s bill.
It would also require the state’s Department of Technology to post the impact reports for active systems on its website.
Led by the California Chamber of Commerce, 26 industry groups—from big tech representatives like the Internet Association and TechNet to organizations representing banks, insurance companies, and medical device makers—signed on to a letter opposing the bill.
“There are a lot of business interests here, and they have the ears of a lot of legislators,” said Vinhcent Le, legal counsel at the nonprofit Greenlining Institute, who helped author the bill.
Originally, the Greenlining Institute and other supporters sought to regulate ADS in the private sector as well as the public but quickly encountered pushback.
“When we narrowed it to just government AI systems we thought it would make it easier,” Le said. “The argument [from industry] switched to ‘This is going to cost California taxpayers millions more.’ That cost angle, that innovation angle, that anti-business angle is something that legislators are concerned about.”
Number of industry groups that signed on to a letter opposing California's ADS bill.
The California Chamber of Commerce declined an interview request for this story but provided a copy of the letter signed by dozens of industry groups opposing the bill. The letter states that the bill would “discourage participation in the state procurement process” because the bill encourages vendors to complete an impact assessment for their tools. The letter said the suggestion, which is not a requirement, was too burdensome. The chamber also argued that the bill’s definition of automated decision systems was too broad.
Industry lobbyists have repeatedly criticized legislation in recent years for overly broad definitions of automated decision systems despite the fact that the definitions mirror those used in internationally recognized AI ethics frameworks, regulations in Canada, and proposed regulations in the European Union.
During a committee hearing on Washington’s bill, James McMahan, policy director for the Washington Association of Sheriffs and Police Chiefs, told legislators he believed the bill would apply to “most if not all” of the state crime lab’s operations, including DNA, fingerprint, and firearm analysis.
Internet Association lobbyist Vicki Christophersen, testifying at the same hearing, suggested that the bill would prohibit the use of red light cameras. The Internet Association did not respond to an interview request.
“It’s a funny talking point,” Le said. “We actually had to put in language to say this doesn’t include a calculator or spreadsheet.”
Maryland’s bill, which died in committee, would also have required agencies to produce reports detailing the basic purpose and functions of ADS tools and would have prohibited the use of discriminatory systems.
“We’re not telling you you can’t do it [use ADS],” said Delegate Terri Hill, who sponsored the Maryland bill. “We’re just saying identify what your biases are up front and identify if they’re consistent with the state’s overarching goals and with this purpose.”
The Maryland Tech Council, an industry group representing small and large technology firms in the state, opposed the bill, arguing that the prohibitions against discrimination were premature and would hurt innovation in the state, according to written and oral testimony the group provided.
“The ability to adequately evaluate whether or not there is bias is an emerging area, and we would say that, on behalf of the tech council, putting in place this at this time is jumping ahead of where we are,” Pam Kasemeyer, the council’s lobbyist, said during a March committee hearing on the bill. “It almost stops the desire for companies to continue to try to develop and refine these out of fear that they’re going to be viewed as discriminatory.”
Limited Success in the Private Sector
There have been fewer attempts by state and local legislatures to regulate private companies’ use of ADS systems—such as those The Markup has exposed in the tenant screening and car insurance industries—but in recent years, those measures have been marginally more successful.
The New York City Council passed a bill that would require private companies to conduct bias audits of algorithmic hiring tools before using them. The tools are used by many employers to screen job candidates without the use of a human interviewer.
The legislation, which was enacted in January but does not take effect until 2023, has been panned by some of its early supporters, however, for being too weak.
Illinois also enacted a state law in 2019 that requires private employers to notify job candidates when they’re being evaluated by algorithmic hiring tools. And in 2021, the legislature amended the law to require employers who use such tools to report demographic data about job candidates to a state agency to be analyzed for evidence of biased decisions.
This year the Colorado legislature also passed a law, which will take effect in 2023, that will create a framework for evaluating insurance underwriting algorithms and ban the use of discriminatory algorithms in the industry.