What happened
Two separate research projects conducted by the University of Maryland’s Center for Educational Data Science and Innovation were published in October. The first evaluated Goblins, an AI‑driven math practice website, with 6,000 middle‑school students in Deer Valley, Arizona. Students who solved at least two problems per week (about 50 over the year) scored 0.11 standard deviations higher on a year‑end math assessment, a gain the researchers equated to roughly three months of additional learning. Half of the participants reached this usage level, but only 6 % logged at least half an hour per week. The second study tested a university‑wide AI assistant integrated into Canvas for 30 courses during the 2025 fall semester. Professors were randomly split, with 15 courses receiving the AI chatbot and 15 serving as controls. Among the 2,300 students, those in AI‑enhanced sections earned an average of 0.11 points lower on a four‑point GPA scale and interacted less with course materials, attended fewer office hours, and posted fewer discussion comments. Only 15 % of students in the AI‑enabled classes used the assistant, primarily to request answers rather than for tutoring.
The Goblins study was commissioned by Accelerate, an independent nonprofit research organization. Researchers compared students who used the platform at a minimum of two problems per week with peers who used it less or not at all. The statistical analysis showed a 0.11‑standard‑deviation advantage for the higher‑usage group, which Accelerate translated into roughly three extra months of learning. Goblins offered text‑based step‑by‑step help, a virtual whiteboard, and spoken conversation, and could converse in Spanish and six other languages. Teachers retained control over problem assignment and could upload custom items, a factor the researchers believe contributed to the positive effect.
In the university AI assistant experiment, the chatbot was fed syllabi, readings, and assignment details for each course. Students accessed it through Canvas and could ask about deadlines, concepts, or assignment instructions. The study’s random assignment of professors ensured comparable course content across treatment and control groups. At semester’s end, the AI‑enhanced cohort’s average GPA was 0.11 points lower, and engagement metrics—discussion board comments, office‑hour attendance, and overall interaction with course materials—were all reduced. The negative impact was most pronounced among first‑generation students. Only a minority (15 %) of students used the assistant, and those who did tended to request direct answers rather than engage in tutoring dialogues.
Both studies were reported by Jill Barshay of The Hechinger Report, who interviewed Jing Liu, director of the Center for Educational Data Science and Innovation, and University of Maryland Provost Jennifer King Rice. Liu emphasized that the two studies were released together to illustrate that AI in education is neither uniformly beneficial nor uniformly harmful, and that ongoing empirical testing is essential.
Why it matters
The contrasting outcomes highlight how AI tools can produce divergent educational impacts depending on design, implementation, and student behavior. Goblins’ modest gains suggest that AI‑mediated step‑by‑step tutoring, especially when paired with teacher‑assigned problems and multilingual support, can boost learning for engaged users. However, the low overall engagement mirrors the “five‑percent problem” common in ed‑tech, where most students do not adopt the tool sufficiently to realize benefits. Conversely, the AI assistant’s association with lower grades and reduced participation raises concerns that ubiquitous AI presence may unintentionally signal permissive AI use, prompting students to rely on external generative models for assignments and thereby diminishing learning. The findings also reveal equity risks: first‑generation students experienced the steepest declines, indicating that AI deployment could exacerbate existing achievement gaps. For policymakers, educators, and ed‑tech developers, the studies underscore the need for rigorous, randomized evaluations before campus‑wide rollouts and for careful messaging about appropriate AI use.
The Goblins results suggest that AI‑driven tutoring can be an effective supplement for students who are motivated to practice, especially when the tool is multilingual and integrates teacher oversight. This points to a potential pathway for scaling personalized math support in under‑served populations, provided that engagement barriers can be overcome.
The AI assistant findings raise red flags about unintended consequences of AI chatbots in learning management systems. If students interpret the presence of an AI as an endorsement to use external generative tools for completing work, the technology could undermine learning outcomes and widen achievement gaps. This underscores the importance of clear usage policies and monitoring mechanisms.
Equity implications are evident: first‑generation students suffered larger grade declines, indicating that AI tools may amplify existing disparities if not carefully designed and supported. Policymakers and institutions must consider differential impacts when approving AI deployments.
Methodologically, both studies illustrate the value of randomized designs in educational AI research. While Goblins’ observational findings cannot confirm causality, the university’s randomized trial provides stronger evidence of effect—albeit a negative one. Future work should aim for larger, multi‑site RCTs to validate these early signals.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Why can ethical evaluation not be reduced to one model score?
What to watch next
Future research should track whether randomized controlled trials of Goblins confirm causal learning gains and whether higher dosage thresholds improve outcomes. Universities planning AI assistants must monitor usage patterns, clarify permissible uses, and consider safeguards against over‑reliance on external AI services. Stakeholders should also watch for broader adoption of AI tutoring platforms and any regulatory guidance on AI in K‑12 and higher education, especially regarding equity and data privacy.
Whether Accelerate proceeds with a randomized controlled trial of Goblins, which could clarify the causal relationship between AI tutoring dosage and learning gains.
University of Maryland’s response to the assistant study, including any policy changes, tool redesigns, or additional monitoring of AI usage in courses.
Adoption trends of AI tutoring platforms in K‑12 districts, especially those targeting multilingual learners and students with disabilities.
Potential regulatory or guidance documents from state education agencies or the U.S. Department of Education addressing AI integration, data privacy, and equity safeguards.