A few months ago, I watched a student in the back corner of my classroom work through a complicated source analysis for about forty-five minutes without once raising his hand. He was quiet, focused, and — I assumed — doing fine. He wasn’t. He’d been stuck on the same paragraph the entire period and hadn’t wanted to bother me.
That moment bothers me more now than it did then, because new research is suggesting that AI tools in classrooms might be making this kind of invisible struggle worse, not better.
The “Sticky Help” Problem
A study out of NC State, presented this week at the Learning Analytics & Knowledge Conference, analyzed 1.4 million student interactions in math classrooms using AI-powered tutoring systems. The researchers found something that sounds technical but hits hard when you think about it in practice: teachers who use AI tutoring tools are significantly more likely to help the same students over and over again.
They called it “sticky help.” Once a teacher has stepped in to assist a particular student, they’re more likely to return to that same student in future sessions — even when other students are struggling just as much. The intervention creates a kind of gravitational pull toward the already-served.
Here’s the part that stings: this is the opposite of what we hoped AI tools would do for equity. The promise was always that data dashboards and intelligent systems would help teachers see the students they were missing — the quiet kid in the back corner, the one who never raises their hand, the one who looks fine but isn’t. Instead, the tools may be reinforcing existing patterns of teacher attention rather than disrupting them.
The researchers were careful to note this isn’t a character flaw. It’s a structural one. The AI systems weren’t designed to actively redistribute teacher attention. They flag struggle. They don’t redistribute help. And so teachers, operating on gut instinct and prior history, keep gravitating toward the students they already know need support.
The Five Percent Problem
This dovetails with another uncomfortable finding that’s been circulating in education research circles. A Gray DI analysis of AI tutoring adoption found that only about 5% of students actively use AI tutoring tools in ways that produce measurable academic gains. A Los Angeles Pacific University study found students who used AI assistants at least three times saw an average 7.5% GPA increase. Good news, right?
Except the students generating those gains are overwhelmingly high-performing students from higher-income backgrounds. The remaining 95% — including most of the students who would benefit most from personalized support — show little to no measurable improvement. The tools that were supposed to democratize access to high-quality tutoring are, in practice, giving another advantage to students who were already ahead.
This isn’t inevitable. But it’s where we are right now, and I think most teachers aren’t hearing this part of the conversation.
What Eight Middle Schoolers Got Right
Meanwhile, a group of eighth graders at Percy Julian Middle School in Oak Park, Illinois spent a year investigating AI tools under their teacher Ashley Kannan’s guidance and then presented their findings to the faculty. I love this for a lot of reasons, but one thing they said stuck with me: they stressed the importance of “creating systems and mindsets that respect boundaries,” and they placed responsibility on both developers and users.
Fourteen-year-olds said that. They understood intuitively what a lot of ed-tech marketing materials gloss over: AI tools don’t come pre-loaded with equity. The equity has to be built in, on purpose, by people.
One of them, Viv Rowell, noted that her generation would “be the first kids who have AI grading our college resumes.” That’s a line worth sitting with. Because if the tools doing that grading are trained on data that reflects existing advantages — if the systems themselves encode who gets attention and who doesn’t — then we’re not disrupting inequality. We’re digitizing it.
The Real Promise (And How Not to Squander It)
None of this means AI doesn’t belong in classrooms. A peer-reviewed review in Artificial Intelligence in Education found that AI tools can genuinely support personalized learning, expand access, and improve feedback quality — but the strongest gains happen when AI augments human teaching rather than substituting for it. The teacher still has to be in the loop, actively redirecting their attention based on what the data shows.
The U.S. Department of Education finalized new grant priorities this month that emphasize the “appropriate and ethical use of AI in education.” That’s a good signal. But policy frameworks don’t change what happens in a classroom on a Tuesday afternoon. Teachers do.
And a recent eSchoolNews piece put it in a way I keep returning to: “In an AI-rich classroom where ideas are abundant and answers are cheap, the scarce resource is not information — it is ownership.” Who owns the learning? Who’s driving it? And critically: who’s being left to navigate it alone?
What This Means for My Classroom (and Maybe Yours)
I’ve been thinking a lot about how to be more intentional about attention distribution, with or without AI tools. A few things I’m actively trying:
Map your attention, not just your students’ work. At the end of a class period, try to recall: who did you talk to? Who did you check on? You might be surprised. The NC State research suggests our instincts about who needs help are shaped heavily by who we’ve helped before — which means the quiet, self-sufficient-looking student in the corner might be invisible not because they’re fine, but because we’ve never established the pattern of checking.
Use AI dashboards as redistribution tools, not confirmation tools. If your school uses an AI-powered system that shows student engagement data, look for the students with low flags, not just the ones with high flags. A student who isn’t generating alerts might be coasting, or might be so lost they’ve stopped trying. The absence of struggle signals isn’t the same as the presence of understanding.
Design AI entry points that work for hesitant students. The research on adoption gaps points to something real: AI tutoring benefits tend to flow to students who are already comfortable taking initiative. That means teachers have to lower the floor. Structured AI prompts, class-time (not homework-time) AI tasks, and explicit modeling of how to use a tool effectively can all help bring reluctant or overwhelmed students into the conversation.
The goal isn’t to use AI less. It’s to use it with our eyes open — aware of where it’s likely to help, where it’s likely to reinforce old patterns, and where the human part of teaching still has to do the work that no algorithm can.
David Jacobson is an AP World History teacher at an international school in Shanghai. He writes about AI and education at shouldiuse.ai. Find him on LinkedIn or reach him at dawidio@gmail.com.
