Two studies led by Jing Liu, director of the Center for Educational Data Science and Innovation at the University of Maryland, produced conflicting results about artificial intelligence in education. Liu released them together in October to demonstrate that AI tools are neither universally beneficial nor harmful.
In the first study, a math practice website called Goblins was deployed to 6,000 middle schoolers in Deer Valley, Arizona, during the 2025-26 school year. An AI tutor provided step-by-step help through text chat, virtual whiteboard, and spoken conversation. Students who solved at least two problems a week scored significantly higher on a year-end standardized math assessment than similar students who used Goblins less frequently. Accelerate, the nonprofit research organization that commissioned the study, calculated the statistical difference of 0.11 standard deviations as equivalent to about three additional months of learning.
The study cannot prove Goblins caused the gains, however. More motivated students may have used it more and improved regardless. Accelerate plans to commission a randomized controlled trial.
Half of the 6,000 students used Goblins at the dosage associated with higher scores. That stands out because education technology often suffers from low engagement. Studies on Khan Academy and IXL have documented that only about 5 percent of students use instructional software consistently as directed.
Goblins appeared to reach students often difficult to engage, including English learners and students with disabilities. Liu suspects the cartoon avatar teacher that could speak and understand Spanish and six other languages made the difference. "It was like talking to your own teacher," Liu said. "It made the students want to use it." Teachers' control over which problems to assign to which students may have also helped engagement.
A second experiment tested an AI assistant developed at the University of Maryland. In fall 2025, Liu randomly assigned 15 of 30 professors to add the assistant to their classes across subjects ranging from English and math to biology and social sciences. The other 15 taught identical courses without AI.
Students in courses with the AI assistant ended the year with lower grades, reduced by an average of 0.11 points on a four-point scale, compared with 1,200 control students. They also interacted less with course materials and were less likely to attend office hours. The negative effects were more pronounced for first-generation students. Only 15 percent of students in AI-enhanced classes ever used the assistant. Those who did mostly asked it for answers rather than using tutoring features.
Liu and Provost Jennifer King Rice suspect that the AI assistant's constant presence sent an unintended signal: AI use is encouraged. Students may have interpreted this as permission to use ChatGPT, Claude, or Gemini outside the university platform to complete assignments. The researchers cannot establish whether that happened, but if external AI tools did the work, students may have learned less.
Rice told a University of Maryland publication that the results were troubling. "AI tools could exacerbate inequities in learning outcomes," she said. Liu stressed that colleges should not avoid AI but should understand that carefully designed tools can have consequences beyond their intended use.