Last year, Nevada gambled that it could improve how it allocated funding for “at-risk” students by using artificial intelligence (AI). To do so, the state gave student data to a private contractor, which promised that its technology would identify the students most at risk of academic trouble using dozens of indicators of student performance. The gamble only made matters worse.

Rather than identifying more at-risk students across the state, the AI program dramatically reduced the number of students who met the definition, from 288,000 to just 63,000. Although the total amount of money stayed the same, which meant more funding for many schools, the change also left school districts that previously received funds with students in need of support but no funding to provide it.

Nevada’s experience has prompted important questions. Which students are considered at risk? How should states define at-risk students? Should AI play a role in those decisions? If so, what guardrails are needed?

So far, few answers have emerged. The company in question has provided little information about its methods beyond the measures considered. This lack of transparency, despite the use of student data and the allocation of public funds, perhaps poses the largest barrier to implementing similar technology in other states.

But if used correctly, AI can help states improve student outcomes. As other state policymakers, school districts, and local leaders test new AI models, they should ensure transparency with private contractors, pilot the models, and determine whether the model actually meets their education goals better than existing funding mechanisms before taking a similar roll of the dice.

Which students are considered at risk of dropping out?

Decades of research has shown that students from families with low incomes tend to perform worse academically. These students have fewer resources than their wealthier peers, which can mean fewer support systems to engage with academic material, less access to technology and other materials, and even a higher risk of missing school because of health issues or lack of transportation. As such, students from families with low incomes are generally considered at risk. But states can counteract these trends—boosting test scores, graduation rates, and earnings—by spending more on students from families with low incomes.

Most states take steps to spend more on these students. But each state’s funding formula is different, and how well it progressively allocates funds to students from families with low incomes varies.

Under its old funding formula, Nevada—like many states—identified at-risk students using the share of students receiving free and reduced-price lunch, which limits eligibility to students from families who make less than 185 percent of the federal poverty level (FPL), as an easy-to-implement substitute for income.

Recent work has suggested, however, that free and reduced-price lunch might not accurately represent at-risk students, given the rise in universal free meal programs. Urban Institute analysis found that students receiving free and reduced-price lunch exceeds the share of students expected to qualify because of their income by 20 percentage points in Nevada.

Using this measure, Nevada allocates roughly 14 percent more funding to students receiving free and reduced-price lunch than students who didn’t. But when using a lower threshold for poverty than free and reduced-price lunch (130 percent of the FPL compared with 185 percent), low-poverty school districts in Nevada outspent high-poverty districts per pupil by 33 percent, the largest gap in the nation.

How should schools define “at-risk” students?

Given the continuing gaps in per pupil funding and the growing challenges with using free and reduced-price lunch as a proxy, states might need a new measure to define at-risk students.

Part of Nevada’s motivation for turning to AI was to integrate other variables into its definition. Students who are English language learners, require special needs education, have a family history of neglect or violence, or are parenting while in school are also at a higher risk of poor performance. Some states already set aside supplemental funding for students who meet these characteristics in different areas of their funding formula. Nevada does not.

Infinite Campus, the company Nevada contracted, feeds its AI 75 indicators (PDF) of student performance, including grade point average, attendance, and even the number of parental logins to the student portal. The AI analyzes student data across all these inputs to create a “grad score” between 50 and 150, with those scoring below 72 considered at risk.

Proponents of the measure point out that students identified under the new system receive much more funding, about $2,900 compared with $303 previously. The system redirects money to those with the highest needs, they argue.

But the system also leaves thousands of students and many schools with shortfalls. Because Nevada uses a binary definition of at risk (as do many other states with or without AI assistance), students just above a cutoff are not allotted any additional funding, meaning directing funding to those with the highest need leaves those with just-above-average needs without above-average support. Consider Somerset Academy: more than 250 students are from families with low incomes and at least a dozen are experiencing homelessness, but none were deemed at risk, according to the New York Times, so the school received no extra funding.

Previous Urban research has also explored alternative definitions of “at risk.” Contracting with the Colorado Legislative Interim Committee on School Finance, we found that if the state used direct certification of students’ familial income and considered neighborhood socioeconomic status, the state could more reliably identify at-risk students. But in implementation tests, we’ve learned that this method might necessitate additional data collection that could be difficult for some states, particularly those with large rural populations.

Unlike Infinite Campus, we did not include academic progress makers in our analysis, in part because allocating funds based on poor student performance could cause districts to lose funding if outcomes improve, disincentivizing student success.

What role can AI play in state education funding?

The desire for a more comprehensive measure of students’ academic trajectories is understandable, but Nevada’s use of AI has created more confusion than clarity. Without transparency about how Infinite Campus weighs the inputs it considers, Nevada’s definition of at-risk students has only become more opaque.

As Nevada’s experience indicates, any definition of “at risk” can dramatically change which students and schools qualify for additional funding. When pursuing new AI mechanisms to identify and support at-risk students, states should consider the following:

• Transparency requirements. States should ensure that whatever definition they use clearly delineates why certain funding decisions are made. Without transparency in allocation, states risk exacerbating existing funding gaps.
• Sensitive data risks. Already, leading AI companies have faced data attacks and breaches. States can follow the recommendations presented in the Biden administration’s Blueprint for an AI Bill of Rights, which calls for enhanced restrictions for any data pertaining to youth.
• Piloting programs before implementation. States should run pilots using AI predictive modeling to see how funding changes before integrating such modeling into allocation processes. By testing new funding thresholds, conversing with local officials and subject-matter experts, and iterating on the modeling, states can better meet their intended goals.

Designing a better income proxy than free and reduced-price lunch is possible but might require states to collect more data or change other processes. Further, as states consider new measures to use, they should also consider following Michigan’s example of allocating funding along a spectrum, so students with all levels of need can receive additional support. As AI continues to improve and slot into our existing processes, state policymakers should consider how they can responsibly yield its benefits for student outcomes.