Research Confirms AI Builds Student Skills One Way, Undermines Them Another

What the Middlebury research found

The study comes in two parts, both led by Middlebury College assistant economics professor Germán Reyes and colleague Zara Contractor. The first was a survey conducted between December 2024 and February 2025. They reached 634 students — more than 20 percent of Middlebury College’s undergraduate body. What they found: 80 percent of students already use AI for coursework. Among those, 61 percent reported augmentation use — asking AI to explain a concept, generate feedback, brainstorm possibilities they then evaluated themselves. Forty-two percent reported automation use: having AI produce essays, generate code, or complete tasks on their behalf.

The second part was an experiment. In spring 2025, students were randomly split into two groups: one with standard online tools, one with those tools plus AI. Both groups researched CRISPR gene-editing technology and wrote essays — first in week one, then again in week two. The results split cleanly along usage type. Automation users got a short-term grade lift in week one. By week two, they performed worse on average than students who used no AI at all. Augmentation users showed only a modest improvement in week one — and a significantly larger one by week two. “The effects of AI largely depend on how students use the tools,” Reyes said. The research is available as a working paper through IZA (Discussion Paper No. 18055).

The frame the science actually supports

Coverage of this study tends to highlight the comfortable finding: most students use AI responsibly. That framing lets everyone off the hook and misses the harder lesson, which is the one parents of struggling learners actually need.

AI automation produced short-term grade boosts and long-term learning loss. That pattern has a precise explanation in decades of cognitive science. The brain builds durable skill through active engagement — through struggle that requires the learner to retrieve, connect, and construct. Brain-imaging studies led by Sally Shaywitz at Yale and by researchers at Stanford show that the neural pathways reading and reasoning rely on are physically built through practice. When a student has AI generate the essay, those pathways get no exercise. When they use AI to challenge their own draft, explain what confused them, or think through why an argument fails — they are doing exactly the kind of effortful, engaged processing that builds lasting understanding. The science on how the brain builds new pathways has said this for years. The Middlebury experiment now shows it in a classroom with an outcome measure.

The system getting this wrong is not the students — the majority of them are already reaching for augmentation. It is the binary framing that dominates the adult conversation: either ban AI (it is cheating) or celebrate any AI use (it is personalized learning). A school policy that bans all AI forecloses the augmentation that helps. An ed-tech platform that celebrates any AI use hides the automation that costs learners. Neither camp is asking what the student’s brain is doing. For children who already find the cognitive work harder — those struggling with reading, math processing, or attention — the automation trap runs deeper. These are the children already looking for ways to make the work easier. AI automation is the most efficient shortcut yet: it produces the output that satisfies the assignment while the brain misses the practice it needs to build the underlying skill. The IDA 2025 definition’s acknowledgment that reading and learning draw on multiple processing systems — language, attention, working memory, processing speed — means there are multiple systems that develop through active engagement and get bypassed through automation.

What this means for your child

None of this makes AI a problem to avoid. The augmentation users in the Middlebury study outperformed the no-AI group by week two. A student who uses AI to test their understanding, evaluate options they have generated, or improve a draft they already wrote is using a cognitive scaffold that works. The research says so.

The honest tradeoff: AI automation is efficient in the short run and corrosive in the long one. AI augmentation asks more of the student in the short run and builds more skill by week two. A few questions worth asking when your child uses AI for schoolwork: Did they read what the AI produced and explain it back to you in their own words? Did they use AI to generate a first draft, or to improve a draft they had already written? Could they answer a follow-up question on the topic without AI? If the answer to all three trends toward no, the AI did the thinking. That is the usage pattern the Middlebury experiment showed costs learners. If the answers trend toward yes, that is augmentation — the pattern that beat the no-AI group. The question to ask of any tool is the same one that governs all of it: is this building my child’s skill, or doing the work in its place?