In just the past few years, artificial intelligence has infiltrated nearly every industry, is accessed daily on billions of smartphones, and now seems poised to transform the way children learn in school. Teachers say they’re caught in the middle, exasperated by the ubiquity of chatbots and large language models—and the temptation students have to use them as a crutch—but fearful of being left behind in the AI race. A question many teachers, parents, and administrators are asking: What do teaching and learning mean in an AI world?
The question is so central for BU Wheelock researchers that the college made it the topic
of its 2026 BU Wheelock Forum in March. The highlight was a faculty panel, moderated by Dean Penny Bishop and featuring some of the college’s brightest researchers in the field: Naomi Caselli, an associate professor of Deaf education; Michael Alan Chang, Margaret McGuire Earl Career Development Professor, an assistant professor, and a faculty fellow in BU’s Faculty of Computing & Data Sciences; Nermeen Dashoush, a clinical associate professor; and TJ McKenna, a clinical assistant professor of science education.
These researchers—along with Ola Ozernov-Palchik, a research assistant professor and director of Boston University’s Evidence-Based AI in Learning (EVAL) Industry Collaborative—are not naive about the ubiquity of AI in schools. Teachers and administrators are using it for curriculum planning, students are using it to research essays and papers, and many computer-based education applications have chatbots built in. BU Wheelock researchers even see AI’s upsides. But successfully harnessing AI will require a commitment to centering human relationships and to introducing only evidence-based technologies, they say.
Dashoush, who left the University in May 2026 to become head of educational impact at LEGO, adds that before we integrate artificial intelligence meaningfully into American education, we must first consider it in the context of educational inequity, where educational outcomes are still too often determined by a child’s race, zip code, and family’s socioeconomic status.
McKenna directs BU Wheelock’s new AI & Education master’s and certificate programs. Much of his research and school-based work deals with how children figure out and understand the scientific world, which he calls sensemaking. He says, for example, a student who’s learning how clouds can be so heavy but can float in the sky could simply ask Google or a chatbot. “A chatbot can hand you the answer, but it can’t hand you the aha when you figure it out,” McKenna says. “The aha is the signal. The sensemaking is the learning.”
He points to a 2025 Proceedings of the National Academy of Sciences (PNAS) study of nearly a thousand high school math students in Turkey, where researchers tested two versions of GPT-4: one that worked like a standard chatbot and one designed with guardrails to support learning. Students using the standard chatbot scored 48 percent higher on practice problems, but when access was taken away, they performed 17 percent worse on the exam than students who never had AI at all. The guardrailed version largely eliminated that learning loss. McKenna sees the takeaway as a design question, not a verdict on AI.
“A chatbot can hand you the answer, but it can’t hand you the aha when you figure it out. The aha is the signal. The sensemaking is the learning.”
The risk, he says, is dependency. “Once you start leaning on those crutches, it becomes really hard,” says McKenna, whose work centers on building AI tools with the kind of curricular guardrails that distinguish learning support from learning replacement. McKenna built the Next Generation Science Standards AI Coach, a free tool grounded in the curriculum teachers are using. High-quality instructional materials like OpenSciEd contain enormous amounts of pedagogical wisdom, he says, but it’s buried across hundreds of pages of teacher editions, unit overviews, and lesson-level guidance. The Coach surfaces that wisdom on demand, drawing from the exact unit a teacher is using that week. Teachers, instructional coaches, and curriculum coordinators across 31 states use it to get answers in seconds rather than mining hundreds of pages of teacher guides, he says.
Building better tools is half of the equation. The other is teaching the next generation to use them well. That’s where AI literacy, starting before children enter school and continuing in the earliest grades, is crucial, Dashoush says. The multimedia tools she designs for children, families, and educators build AI literacy by helping them see that the computational thinking and computer science skills behind artificial intelligence are core problem-solving skills they already use in everyday life. “By demystifying AI, my goal is to ensure they have the skills to dissect and understand whatever AI tools they encounter and become creators rather than simply consumers of these technologies,” she says. Her vision for responsible AI development centers on co-design—meaningfully involving children in the process of developing a technology—and ensuring access to any technologies is distributed equitably.
“I design AI literacy apps and multimedia tools with the goal of making them accessible to all children—not just those whose families can afford private camps, tutors, or after-school programs,” Dashoush says. “By partnering with public media and integrating these resources into school-based curricula, I work to ensure that high-quality AI literacy learning experiences are available to children wherever they already learn every day.”
According to the EdTech Top40 Report, released by Instructure, the average American student uses 42 educational technology tools each year in school. Grade schoolers spend, on average, 98 minutes each day on school-issued digital devices. But even as the value of the educational technology, or edtech, industry ballooned to more than $200 billion in 2025, Ozernov-Palchik says her research shows that less than 10 percent of the technology kids use in school is backed by studies showing they actually work.
Students with special needs—the most vulnerable—are often the heaviest users of unproven apps, precisely because schools don’t have the specialists to serve them. “Those are the students who need the most support,” says Ozernov-Palchik, who is also the associate director of translational research for BU’s AI & Education Initiative. “And those are the students who are the first line of defense for using these tools. We’re just dumping these technologies on kids who need human support, effective support, evidence-based support in order to thrive.”
She thinks edtech, including AI tools, ought to be held to the same causal standards as pharmaceuticals: fully blinded, randomized controlled trials (RCTs). “I think about it like we need FDA for education,” she says. “Basically, we don’t integrate a tool into classrooms unless we have the causal evidence that it works.”
Through the center she runs, Ozernov-Palchik launched the EVAL Challenge, an open call to edtech companies to apply for a free RCT conducted by BU faculty. After receiving 35 applications for its first RCT, EVAL selected Amira Learning, a widely used AI-based reading tutoring tool that has served millions of students but has no published causal evidence of effectiveness. The Amira RCT will compare three groups: students using Amira, students receiving no supplemental instruction, and students receiving live tutoring from trained BU undergraduates—the third control group designed to test whether a scalable, minimally trained human can rival an AI tool. She hopes the RCT results build a little friendly competition between edtech companies to show, with evidence, that they improve educational outcomes.
“Let’s say we find that [Amira] works. Amira is now getting an edge above other tools that have not been evaluated,” she says. “That puts pressure on those other tools now to get that research evaluation.”
Chang, a computer scientist by training, whose education research explores ethical, participatory approaches to AI-supported teaching and learning, says there’s a lot of excitement about harnessing AI to make learning faster and more efficient. He urges, however, that there be a collective pause when it comes to thoughtlessly saturating our classrooms with AI.
Oftentimes pausing is seen as a fearful reaction, Chang says. “In another framing, it can totally be about courage. It’s the courage to resist this urge to just chug forward. We’re in the space of computer science. There’s often this thing of making things really fast, and then breaking them—which can work in a lot of contexts. But in social contexts like education, it doesn’t often work out that way.”
Pausing AI’s takeover of education requires asking foundational questions, like what work is this technology doing within the broader system? he says. AI schools like Khan Academy, he says, make big promises, but if the underlying system produces inequitable outcomes for Black and brown students, is AI just scaling inequity?
“We need to imagine schooling first,” Chang says, “and then think about the role of technology afterwards.”
Teaching and Learning in an AI World
Steve Holt Profile
Justin Metz Profile
Boston University moderates comments to facilitate an informed, substantive, civil conversation. Abusive, profane, self-promotional, misleading, incoherent or off-topic comments will be rejected. Moderators are staffed during regular business hours (EST) and can only accept comments written in English. Statistics or facts must include a citation or a link to the citation.
Your email address will not be published. Required fields are marked *
About the Author
