Students do not learn solely through logic. Frustration, curiosity, hesitation, and confidence can shape whether a lesson lands or slips away, yet most AI tutors still respond as if every learner were in the same state. Chenyu Zhang is part of a group of researchers asking whether educational technology can do more than deliver answers, whether it can sense emotion, respond with care, and make learning feel more attentive to the people using it.

That question has guided Zhang’s work across research and teaching. Based in Cambridge, he has held roles linked to Harvard’s Berkman Klein Center, MIT Media Lab, Stanford HAI, and the University of Georgia, while developing GlowingStar, a startup focused on emotion-aware tutoring. His academic path, from computer science at the University of Toronto to a Master of Education at Harvard, helps explain why his work sits so naturally between engineering and the lived realities of the classroom.

Where Feeling Meets Code

For years, AI tutors have excelled at speed. Feeling, though, remains the blind spot. Zhang’s ACII 2025 paper tackled that problem by studying affective dynamics in student-tutor dialogue across 16,986 turns from 261 learners at three U.S. institutions.

Rather than relying on a single large language model, the study used an ensemble, allowing several systems to weigh the same exchange and read mood with greater care. Curiosity, confusion, frustration, and recovery mattered because a wrong reply from a tutor can be more than a mistake; it can become the moment a learner decides a subject is no longer for them. Zhang’s premise is simple: a machine that misses emotion may miss the lesson, too.

“Dialogue is a dance of affect and intent. When AI joins that dance, we must teach it when to lead, when to listen.” That line carries the heart of his research. A tutor, human or digital, does more than deliver information. A tutor sets the pace, reads hesitation, and senses when a student needs a simpler way into the same idea.

Older tutoring software tends to treat every stalled answer in the same way. It sees the error, then serves another hint. Human teachers know the difference between a learner who is confused, one who is embarrassed, and one who is simply tired. Zhang wants AI to notice that difference, too.

Later papers from NeurIPS 2025 workshops pushed at another hard truth: emotion rarely flows through a single clean channel. A face may look calm while the voice tightens. A quick answer may hide panic. Zhang’s work on multimodal reasoning examined what happens when one signal sabotages the others, a problem with real weight once a tutor starts reading text, tone, and expression at once.

A Researcher Shaped by the Classroom

Long before conference titles and startup decks, Zhang knew the quieter side of learning. He grew up in China, studied computer science at the University of Toronto, and later earned a Master of Education in Learning Design, Innovation, and Technology at Harvard. That path gave him two traits that do not often live in one person for long: the discipline of an engineer and the patience of a teacher.

Teaching kept pulling him back to the human question. His record runs through natural language computing at Toronto, Stanford’s Code in Place, AI4ALL, MIT Media Lab classes on affective computing, Chengdu University, and graduate teaching at Northeastern University. Across those rooms, one lesson kept returning: students rarely stop because content is hard alone; many stop when confusion turns private and hardens into shame.

Work in the industry sharpened that lesson. Before his recent run in research, Zhang built large software systems at Manulife and ROSS Intelligence, where he worked on reliability, search, accessibility, and large-scale user products. That chapter matters because emotion-aware learning tools will live or die on quiet virtues such as stability, speed, and clear user experience, not grand claims alone.

“Every learner is a glowing star. Our job is to help them shine—by noticing when they’re stuck, encouraging them when they’re close, and challenging them when they’re ready.” GlowingStar grew from the same belief, with public material describing an alpha tutor called Glowy that aims to act less like a cold answer machine and more like a steady learning coach.

​Plenty of AI stories lean on access, funding, or market heat. Zhang’s story carries a different pulse. Attention sits at the center of it. He keeps returning to the small classroom moments software often misses: the extra beat before a reply, the flat tone after a setback, the silence that says a learner is still there but slipping away.

Why the Next Tutor May Feel Different

Glowy is still in alpha, and that early stage matters. AI can generate excitement quickly, but trust in the classroom takes longer to earn. Early testing with users connected to Harvard and MIT has helped shape the tutor, grounding its development in Zhang’s research on emotion-aware learning.

What makes the idea compelling is its modesty. Zhang is not arguing that software can love a student or replace a teacher. He is betting that better noticing can change the tone of learning: a gentler reply after frustration, a pause before more difficulty, a nudge that respects a learner’s pride. That kind of tutor feels less like a machine barking answers and more like a companion that knows when to lower its voice.

Trouble shadows the promise. Material on GlowingStar’s production work points to privacy, cultural bias, manipulation risk, and weak generalization across users as live problems rather than solved ones. A system that reads emotion badly can embarrass a student, push too hard, or mistake performance for well-being. Human-centered AI sounds noble until it fails in a real classroom.

​Zhang’s recent roles suggest he knows that. At Harvard’s Berkman Klein Center, he has studied ways to steer agentic conversational AI through interpretable control dimensions tied to safety and controllability. At the same time, his current University of Georgia work examines how generative AI materials can be tested for clarity, engagement, and classroom usefulness.

GlowingStar’s ambition reaches well past schoolchildren. A recent talk description framed the tutor as a companion for working-class students and lifelong learners, drawing on signals from speech, facial microexpressions, response delay, and silence. That breadth matters. Adult learners carry different fears from teenagers. Many return to study after years away, already braced for failure. A system that can sense strain without shaming them could change whether they stay the course.

​Schools have chased speed for years: faster grading, faster feedback, faster content. Zhang is chasing something slower and, for that reason, more daring: the pause before a learner quits. Overall, students learn best when they feel heard.