Researchers at the University of Pennsylvania discovered that small changes in how AI tutors assign homework can produce outsized learning gains. In a five-month study across 10 high schools in Taipei, students who received personalized problem sequences scored significantly higher on final exams than peers following standard curricula, even though both groups used the same AI chatbot and course materials .

Why Do Students Stop Asking for Help When AI Is Available?

The shift toward AI tutoring has created an unexpected problem for educators. Before generative AI became ubiquitous, students would visit office hours or raise their hands in class when confused. Now, many skip those interactions entirely and ask ChatGPT instead. This leaves instructors flying blind about what students actually understand until exam day arrives .

"One of the things that's happened with generative AI is that, as a professor, I now have no idea what's going on with most of my students. They don't come to office hours anymore," explained Hamsa Bastani, Professor at Wharton.

Hamsa Bastani, Professor at Wharton School of Business

This visibility gap matters because it forces educators to infer student struggles from indirect signals rather than direct conversation. The Wharton team's solution was to leverage the data trails students leave behind as they interact with AI tutors, using those signals to proactively adjust what comes next.

How Does Personalized Problem Sequencing Improve Learning?

The research team built an AI tutoring platform for a Python certification course and randomly assigned students to two groups. The control group received problems in a fixed sequence, progressing from easy to medium to hard, similar to traditional homework. The treatment group received problems dynamically adjusted based on their performance and interactions with the AI tutor .

The algorithm monitored multiple signals to determine readiness: how students engaged with the chatbot, their code submissions, and their solution attempts. If a student demonstrated mastery at one difficulty level, the system moved them forward. If they struggled, they stayed at the current level until ready to advance .

Students in the personalized group improved exam scores by 0.15 standard deviations, which researchers estimate translates to roughly six to nine months of additional learning. Critically, this gain came without increasing instructional time or teacher workload .

"The key point is personalization. If you don't personalize, slower students can't catch up. But if you move too slowly, more advanced students aren't going to learn much either. With AI, you can tailor instruction to each student's individual learning trajectory," noted Angel Tsai-Hsuan Chung, PhD Candidate at Wharton.

Angel Tsai-Hsuan Chung, PhD Candidate in Organizational Information Design at Wharton

What Makes This Different From Simply Using an AI Chatbot?

Many educators assume AI tutoring is already personalized because chatbots respond to individual questions. But that misses a fundamental challenge in learning: students often don't know what to ask. They lack awareness of their own knowledge gaps or where they should push themselves to improve .

The Wharton approach flips this dynamic. Rather than waiting for students to formulate questions, the system proactively identifies what each student needs to practice next based on their demonstrated understanding. This "proactive personalization" removes the burden from students to self-diagnose their learning needs.

The researchers also built guardrails into their chatbot to preserve what educators call "productive struggle." The AI doesn't give answers directly; instead, it provides guidance that requires students to show effort. This design choice prevents the over-reliance problem that can occur when AI makes learning too easy .

How Can AI Tutoring Help Anxious and Multilingual Readers?

Beyond problem sequencing, AI tutoring addresses an emotional barrier that traditional classroom reading often creates. When students read aloud in front of peers, teachers, or parents, anxiety can interfere with learning itself. For emerging readers, students with dyslexia, and multilingual learners, this public performance can feel like a high-stakes evaluation rather than a learning opportunity .

Research on the "affective filter" in language learning shows that emotional stress makes it harder for learners to process and internalize new information. An AI reading tutor removes the social judgment from practice, allowing students to attempt, fail, and retry without fear of evaluation .

A study from the University of Chicago's Human-Robot Interaction Lab found that children reading aloud to a robotic companion showed lower observable anxiety than when reading to an adult. The key difference: the robot did not display evaluative facial expressions. Similarly, research published in the Arab World English Journal found that learners reported lower anxiety when interacting with AI tutors, largely because the system provided feedback without social judgment .

Ways to Implement AI Tutoring Effectively in Your Classroom

• Preserve Productive Struggle: Design AI tutors with guardrails that guide students toward answers rather than providing them directly. This maintains the cognitive effort necessary for deep learning and prevents over-reliance on AI assistance.
• Use Interaction Data to Personalize: Monitor signals from student behavior, such as chatbot interactions, code submissions, and solution attempts, to dynamically adjust the difficulty and sequence of practice problems based on demonstrated mastery.
• Create Low-Stakes Practice Environments: Use AI tutors as a "batting cage" before live performance. Allow students to practice reading, problem-solving, and other skills without social judgment, building confidence before they demonstrate skills to teachers, peers, or family members.
• Leverage AI as a Capacity Multiplier: Deploy AI tutoring to increase practice opportunities within limited intervention time, reduce friction during homework routines at home, and support multilingual learners navigating language anxiety, while keeping teachers as the primary relationship and decision-makers.

What Do Teachers Need to Know About AI Tutoring Limitations?

While the Wharton study shows meaningful gains from personalized problem sequencing, researchers emphasize that AI tutoring is not a replacement for human instruction. Teachers remain essential for building relationships, providing encouragement, and making professional judgments about individual student needs .

The most effective learning environments combine AI-powered practice with teacher-led instruction and responsive teaching. AI handles the repetitive, low-stakes practice that builds foundational skills and confidence. Teachers handle the nuanced feedback, motivation, and real-world application that transforms practice into mastery .

For administrators considering AI tutoring adoption, the research suggests focusing on tools that personalize based on student performance data, rather than simply providing access to a chatbot. The difference between a generic AI assistant and a thoughtfully designed tutoring system can mean the difference between marginal gains and learning improvements equivalent to months of additional instruction .