Public education systems around the world are under strain. Teachers bear the burden of doing more with fewer resources. In Brazil, limited resources for public education lead to persistent performance gaps: Public high school students tend to score lower than their private school peers on the writing section of the national college entrance exam. These gaps can translate into unequal access to higher education. Individualized support for students can help, but is time-consuming for public school teachers to provide.

An AI-powered feedback platform helped teachers spend more time on individual feedback and instruction. A randomized evaluation of an AI platform that provided essay feedback for high school students in the Brazilian state of Espirito Santo found notable gains in students’ writing. While the AI tool, Letrus, didn’t perfectly match teachers’ abilities to provide feedback, it flagged grammatical mistakes and sped up other routine editing tasks. With AI handling these basic elements, teachers spent more time discussing substantive feedback with students one-on-one, where their expertise has the greatest impact.

Students who used the platform wrote more and wrote better, narrowing performance gaps between public and private school students. With instant feedback and more time to engage with teachers, students drafted more practice essays and achieved higher scores on the national college admissions exam. After adopting the tool state-wide, the state of Espírito Santo jumped from 11th to 1st place in national writing rankings, and public school students in Espírito Santo closed nine percent of the test score gap with private schools. These results led to a nationwide expansion of the platform.

Education administrators and policymakers should explore how AI can support teachers by taking on time-consuming tasks. The success in Brazil shows that, with the right support and attention to measuring real-world impacts through randomized evaluation, AI can be a valuable tool for improving education delivery.

Human oversight plays a key role in catching errors and preventing unintended harm from AI tools, and should be balanced against the time and cost savings these tools are meant to provide. In Brazil, adding trained graders to review AI-generated feedback on student essays didn’t improve student outcomes. Teachers were able to understand and use the feedback effectively on their own. Still, in settings where teachers or other key users are less equipped to judge AI outputs, closer human review may be worth the extra cost.

At-scale deployment can make AI tools more cost-effective by lowering the cost per student. In Espírito Santo, the state education department decided to fully integrate the program into its high school senior curriculum and included it in the state’s education budget. With support from a foundation and other state governments, the platform is now scaling to reach over one million students across Brazil.