Common Sense Media’s 2026 census on tween and teen AI use found that 85 percent of teen AI users turn to it for schoolwork, and 16 percent say they now struggle to start or finish an assignment without it—a figure that climbs to 28 percent among students who use AI for homework daily. Those numbers describe a dependency, not a tool preference, and dependency is what happens when a technology performs the cognitive work a student was supposed to perform.
The distinction that matters for a school evaluating classroom AI is not whether a tool uses AI. Nearly everything on the market does by now. The distinction is whether the task wrapped around the model requires the student to generate the reasoning, or hands the reasoning to the model instead. A tool built to finish work and a tool built to teach work can run on identical technology and still produce opposite outcomes, because the difference lives in task design, not in the model itself.
Why the Friction Is the Point, Not the Bug
Robert Bjork’s 1994 concept of desirable difficulties supplies the mechanism. Bjork distinguishes performance, doing well in the moment, from learning, retaining and transferring a skill later. Conditions that slow a learner down (withheld answers, delayed feedback, required retrieval) produce worse short-term performance but stronger long-term learning. An AI tool that removes that friction by producing a finished paragraph on request is focused on task completion, a metric that runs opposite the one that actually matters: skill transfer.
Graham and Perin’s 2007 meta-analysis of adolescent writing instruction, conducted for the Carnegie Corporation, examined more than 140 studies and found the same pattern from the instructional side. Explicit strategy instruction—teaching students how to plan, draft, and revise—produced effect sizes well above passive approaches like studying model essays. The strategies worked because students had to apply them, not because a model or a text explained them well. An AI tool that skips the applying and supplies the finished plan removes the exact mechanism Graham and Perin identified as effective.
What Instructional AI Feedback Actually Requires
Hattie and Timperley’s 2007 review of feedback research in Review of Educational Research proposed that effective feedback answers three questions: where am I going, how am I doing, and where do I go next. Most AI writing tools answer only the first question, by producing a version of where the student is going, the finished draft, skipping the other two. Likewise, a tool that scores a paper without addressing what to revise next has delivered evaluation, not feedback. Evaluation without a next step does not move a student’s writing forward any more than a grade alone does.
I built Guided Scholar’s Coach Me mode around that third question specifically. It responds to a student’s draft with recommendations of what to revise next, not a rewrite, which keeps the feed-forward step that Hattie and Timperley identified as the one most classroom AI tools skip entirely.
The practical test for a school piloting a classroom AI tool is not whether the output is accurate or the interface is polished (though best-in-class AI tools often provide both). It is whether a demonstration can show a moment where the tool withholds an answer and requires the student’s move first. If every demo scenario ends with the AI producing the finished product or giving the answer before the student struggles, the tool belongs in the same category as the chatbot most students are already leaning on for homework, regardless of what the sales deck calls it.
In a digital environment that constantly fragments students’ attention, schools should protect the moments of struggle that learning requires. Students need opportunities to attempt the work themselves, receive targeted feedback, and revise based on that feedback rather than leaning on AI for an immediate answer. That cycle—attempt, feedback, revision—is how students build confidence, transfer skills, and learn that success is something they can repeat.
Further Reading
- Bjork, Robert A. “Memory and Metamemory Considerations in the Training of Human Beings.” In Metacognition: Knowing about Knowing, edited by Janet Metcalfe and Arthur P. Shimamura, 185–205. MIT Press, 1994. The original formulation of desirable difficulties—the case that conditions which slow learning down produce stronger long-term retention.
- Graham, Steve, and Dolores Perin. Writing Next: Effective Strategies to Improve Writing of Adolescents in Middle and High Schools. Washington, DC: Alliance for Excellent Education, 2007. A meta-analysis of more than 140 studies identifying which writing-instruction elements move student outcomes, with explicit strategy instruction among the strongest. Link.
- Hattie, John, and Helen Timperley. “The Power of Feedback.” Review of Educational Research 77, no. 1 (2007): 81–112. The three-question feedback model—feed up, feed back, feed forward—used here to separate evaluative AI output from instructional AI output. Link.
- Common Sense Media. The Common Sense Census: AI Use by Tweens and Teens, 2026. San Francisco: Common Sense Media, 2026. The survey data behind the homework-dependency figures cited above. Link.