Education is one of the domains where AI's potential has been discussed the longest and delivered the least — until now. The combination of capable language models and the organizational willingness to deploy them is producing the first generation of AI tutoring systems that behave substantially differently from the knowledge-search tools that came before.
Khan Academy's Khanmigo, built on GPT-4, has been the most visible example of an educationally-principled AI tutor. Rather than answering student questions directly, Khanmigo is designed to guide students toward understanding through Socratic questioning — responding to "what is the derivative of x²?" not with "2x" but with "What rule do you remember for differentiating powers of x?" This design choice, contentious among some educators who want instant answers, reflects decades of learning science showing that effortful retrieval and productive struggle produce better long-term retention than passive receipt of correct answers.
Early outcome data from a small pilot study published in early 2026 suggests the approach is producing measurable effects. Khan Academy reported that students using Khanmigo for math problem-solving showed 17% higher performance on subsequent assessments compared to students using the platform without AI, though the study was small and not peer-reviewed. Independent verification is needed before strong conclusions are drawn.
Duolingo's AI features have taken a different approach, using AI primarily for conversational practice in language learning. The company introduced "Roleplay" scenarios in 2024, where learners can practice conversations with AI characters in their target language — a first date, a job interview, ordering at a restaurant. The feature has been one of Duolingo's most-used additions in recent memory, because it addresses the practice gap in language learning: textbooks teach grammar and vocabulary, but learners have historically had few accessible ways to practice actual conversation before traveling to a target-language country.
Personalized pacing is the third major application gaining traction. Traditional classroom instruction moves at the pace of the average student, which means the fastest students are bored and the slowest are lost. AI-driven adaptive systems can modulate the difficulty and pacing of instruction in real time based on student performance signals. Carnegie Learning's MATHia platform, which has been in schools for longer than most AI tutors, has published the most rigorous outcome data in the field: a meta-analysis of 14 studies found students using MATHia learned the equivalent of an additional 2.5 months of content in a school year compared to control groups.
The adoption barrier is not technological — it is institutional. Integrating AI tutoring into existing curriculum, professional development for teachers who must work alongside these tools, and ensuring the tools do not entrench existing inequities are challenges that require sustained organizational effort beyond what any software product can provide.