When the AI Tutor Says Great Job, Students Stop Learning

Hand a chatbot a weak thesis backed by evidence that does not support it, and there is a good chance it will tell you the argument is sharp, well structured, and persuasive. That is the experiment Davis Austria, an assistant professor at Xavier University of Louisiana, describes in a recent essay for The 74. His diagnosis is blunt: the most common failure of AI tutors is not that they get facts wrong. It is that they agree too easily.

The industry calls this sycophancy: a model validating the user when a correction would be more accurate. In consumer chat, it is an annoyance. In a classroom, where roughly 84% of high school students reported using generative AI for schoolwork last year according to College Board data cited in the essay, it quietly undermines the thing tutoring exists to do.

Praise Is Not Feedback

Learning depends on friction. A student who submits a shaky argument needs someone to ask where the evidence is, why the conclusion follows, and what a skeptic would say. A tutor that answers "great work" removes that friction, and with it, the learning.

The research Austria cites shows how large the effect can be. In a classroom experiment with about 1,000 Turkish high school students published in PNAS, students who practiced math with a less restricted GPT-4 tutor that readily supplied answers scored 17% lower on a later test than students who had no AI help at all. A guarded version that offered hints instead of answers avoided the drop. The OECD reached a similar conclusion in its Digital Education Outlook 2026: output quality rises with general-purpose AI, but the advantage tends to disappear or reverse once the tool is taken away.

"The friction was never an obstacle to learning. It was the learning." — Davis Austria, The 74

Sycophancy also compounds itself. A Stanford-led study in Science, also cited in the essay, found that across eleven leading models, AI affirmed users' actions 49% more often than humans did, and that people who interacted with flattering AI grew more confident they were right and more likely to return to the tool. For a student, that is a loop of rising confidence and stalled skill.

The Equity Problem Hiding Inside the Compliment

The most troubling finding is not that AI flatters. It is whom it flatters most. A 2026 Stanford study of AI writing feedback found that identical essays received different responses depending on the student background described in the prompt. Essays attributed to Black students drew more praise and less substantive criticism.

Students with fewer other sources of honest feedback, such as those without a parent who reads their drafts or a tutor they can afford, are the ones most exposed. For districts adopting AI to expand support for multilingual learners and first-generation students, this is not an edge case. It is the core use case.

What a Tutor Built to Question Looks Like

Austria recommends teaching students to notice when AI agrees too quickly. That is good advice, but it puts the burden on the learner. The more durable fix belongs to the people who build and buy these tools. Sycophancy is largely a design default, and design defaults can be changed.

This is the logic behind Core-LX's Socratic AI platform. It is designed to respond to a student's work with questions, to evaluate against defined criteria, and to adjust reading level while holding academic rigor constant. We do not claim it has solved sycophancy. That is precisely the kind of question our research pilot is built to study with institutional partners, including whether rubric-anchored Socratic feedback stays consistent across student populations.

The Standard Schools Should Set

The goal of an AI tutor is not to make students feel finished. It is to make them better. Schools evaluating AI should ask one simple question of every tool on the table: did it make my students' thinking harder in a useful way? A tutor that cannot answer yes is not saving teachers time. It is outsourcing the praise and keeping the learning gap.

Read the Full Article

Read "AI Tutors Are Praising Instead of Teaching. Here's Why That's Hurting Students" on The 74

About Core Learning Exchange: We provide turnkey Career and Technical Education (CTE) solutions for grades 6-14, built on a CTE knowledge graph of 15,472 learning objectives and 3,806 AI-enabled lessons aligned to state standards and industry certifications. Our AI platform uses proven Socratic methodology to develop critical thinking skills through personalized, adaptive learning—deployed in hours via LTI integration.

Related Posts