I spent two years at Edxcare International as a Growth PM, working on a platform that served students preparing for professional certification exams. The problem we were trying to solve was deceptively simple: every student comes in with different knowledge gaps, different learning styles, and different constraints on their time. A one-size-fits-all curriculum wastes time for students who already know the material and loses students who have foundational gaps.
We tried to solve it with sophisticated question banks, performance tracking, and rule-based adaptive paths. It worked reasonably well. But the rules were brittle - they captured the learning patterns we'd anticipated, not the ones that actually emerged. And they couldn't adapt to the emotional and motivational dimensions of learning, which are often the real barriers to progress.
What I saw in edtech then, and what I see increasingly now, is that AI is solving the personalization problem in ways that rule-based systems fundamentally couldn't. Not perfectly - the overhype in educational AI is real and worth discussing. But the progress in the last two years is genuine and consequential.
The Personalization Problem in Education
Before exploring what AI is doing, it's worth understanding why personalization has been so hard.
True personalization in education requires at least four things happening simultaneously:
- Accurate diagnosis: Understanding what a student knows, what they don't know, and why they don't know it
- Content mapping: Knowing which content will address each specific gap most efficiently
- Engagement optimization: Presenting content in a format and at a pace that works for this specific student at this specific moment
- Motivation management: Keeping the student engaged and confident through the inevitable plateaus and setbacks
Human tutors do all four intuitively. Great teachers do it at classroom scale with imperfect tools. Digital platforms have historically been able to do rough versions of (1) and (2), but almost nothing with (3) and (4).
AI is changing the calculus on all four dimensions.
Adaptive Learning: Beyond Rule-Based Systems
Khan Academy's integration of GPT-4 into their platform - now called Khanmigo - is the most widely known example of AI-powered adaptive learning, and it's worth studying carefully because it illustrates both the potential and the current limitations.
Khanmigo operates as a Socratic tutor. Rather than giving students answers, it asks guiding questions designed to lead them to understanding. When a student is stuck on a math problem, Khanmigo doesn't just explain the solution - it probes what the student understands, identifies the specific misconception, and guides them to the insight themselves.
This is qualitatively different from earlier adaptive learning systems. Earlier systems adapted content selection based on performance - if you got question type X wrong, serve you more question type X. Khanmigo adapts the teaching interaction itself - it responds to how the student is thinking, not just what they answered.
Duolingo's approach is complementary. Their AI model, Birdbrain, has become the backbone of their adaptive learning engine. It models each user's knowledge of each language concept as a probability distribution - tracking not just whether you know a word, but how likely you are to remember it tomorrow, next week, and next month. The spacing repetition logic is AI-optimized per learner rather than fixed by curriculum design.
Their more recent AI Max subscription adds conversational practice powered by GPT-4 - letting learners practice open-ended conversation in the target language, with the AI adapting to their vocabulary level and correcting errors in context.
AI Tutoring: The Promise and the Reality
AI tutoring is the application that gets the most breathless coverage, and the one where I'd urge the most calibrated optimism.
The promise: personalized, patient, always-available tutoring at near-zero marginal cost. For the hundreds of millions of students globally who have no access to quality tutoring, this is a transformative opportunity.
The reality as of 2026: AI tutors are genuinely useful for well-defined academic content - mathematics, grammar, factual recall, procedural problem-solving. They struggle with open-ended creative tasks, with cross-domain synthesis, and with the kind of Socratic depth that a great human tutor brings to ambiguous problems.
They also struggle with motivation. A student who is bored, anxious, or frustrated doesn't just need better content - they need encouragement, context-appropriate challenge, and sometimes the acknowledgment that learning is hard. Current AI tutors handle these moments awkwardly. They can detect frustration signals in text, but their responses lack the genuine empathy of a human who has themselves struggled and succeeded.
Companies working seriously on this problem:
- Synthesis: Originally built for SpaceX employees' children, now broadly available. Their AI tutor focuses on collaborative problem-solving and critical thinking rather than content delivery.
- Carnegie Learning: The academic heritage of cognitive tutoring systems, now integrating LLMs into their long-established adaptive learning platform.
- Socratic (Google): AI-powered homework help that focuses on explaining concepts rather than giving answers.
Assessment Automation: The Underappreciated Revolution
While AI tutoring gets the headlines, AI-powered assessment may be creating more immediate, widespread value.
The traditional assessment model is deeply inefficient: teacher creates assessment, students take it, teacher grades it (often days later), students receive feedback when the learning moment has passed. For open-ended assessments - essays, coding problems, design exercises - grading is slow, expensive, and inconsistently applied.
AI assessment tools are transforming this:
- Immediate feedback: AI can give students feedback on their work within seconds, while they're still in a learning mindset
- Consistent rubric application: AI applies evaluation criteria consistently across all submissions, eliminating grader variability
- Formative assessment at scale: Teachers can assign frequent low-stakes assessments knowing AI handles the feedback, enabling more frequent course corrections
- Detailed diagnostic insights: Rather than "73% - needs improvement," AI assessment can identify specifically which concepts are weak and suggest targeted practice
Turnitin's AI assessment tools and ETS's e-rater for essay scoring have been in production for years. The newer generation - built on LLMs - can handle more complex open-ended responses and provide richer feedback.
At Edxcare, the highest-use thing we ever did to improve student outcomes was shorten the feedback loop from weekly assessments to daily micro-assessments with immediate AI feedback. Completion rates went up. Exam pass rates went up. Students who would have fallen behind for weeks before anyone noticed were now getting help within 24 hours.
Teacher Copilots: AI as a Force Multiplier
The most overlooked AI application in education - and perhaps the one with the highest near-term ROI - is AI tools that amplify teacher capability rather than replacing teacher instruction.
The teacher's job is wildly complex: lesson planning, differentiated instruction, assessment creation, parent communication, administrative documentation, student support. A significant fraction of this time is spent on tasks that are valuable but not uniquely human - generating practice problems, writing rubrics, drafting parent communications, tracking attendance patterns.
AI copilot tools for teachers:
- Lesson plan generation: Tools like MagicSchool and Diffit generate standards-aligned lesson plans from a curriculum objective, saving hours per week
- Differentiated materials: Automatically generating versions of the same content at different reading levels for different students
- Communication assistance: Drafting parent emails and progress report comments
- Early warning systems: AI analysis of engagement and performance patterns to identify students at risk before they fall significantly behind
The teacher copilot use case is strategically important for another reason: it sidesteps the "AI replacing teachers" narrative that creates legitimate resistance to edtech adoption. Teachers don't resist tools that make their jobs easier. They resist tools that threaten their role or add burden without benefit.
Framing matters enormously in edtech adoption. At Edxcare, we learned that teachers were our most important champions - not students, not administrators. A tool that helped teachers look good in front of their students spread virally through teacher networks. A tool that made teachers feel replaced or surveilled died on arrival. Design for teacher empowerment first.
What's Overhyped
In the interest of intellectual honesty, some things that are frequently oversold:
- "AI will replace human teachers": Not in any meaningful timeframe. The relational, motivational, and community-building functions of great teachers are not close to being replicated.
- "Personalized learning will solve the achievement gap": The gap is driven by structural inequities - poverty, nutrition, housing instability, access to resources - that personalized learning software cannot address.
- "AI tutors work for all content types": Strong for structured academic content, much weaker for creative, interdisciplinary, and socio-emotional learning.
- "AI can replace the social dimension of learning": Students learn from each other. Peer interaction, collaborative problem-solving, and healthy competition are irreplaceable.
The Infrastructure Challenge
One thing that rarely makes the edtech AI coverage: the infrastructure reality of deploying AI in educational settings is difficult.
Many school districts - particularly in developing markets, rural areas, and low-income urban areas - have unreliable internet connectivity and a wide range of device capabilities. AI-powered tools that require low-latency LLM inference don't work well on 2G connections or decade-old Chromebooks.
The edtech companies that will win in global markets will be the ones that solve for offline capability, low-bandwidth operation, and device heterogeneity. This is a much harder engineering problem than building for high-bandwidth users on modern hardware.
AI in education is past the demo stage and into early deployment at scale. The results are encouraging: better engagement, faster feedback loops, more granular diagnostic insights, and genuine personalization that goes beyond rule-based content selection.
The ceiling for impact is enormous. A system that gives every student access to patient, personalized, always-available academic support - at the cost of a few dollars per month rather than the hundreds of dollars per hour of human tutoring - is one of the most leveling technologies ever developed.
We're not there yet. But the trajectory is unmistakably toward it.
