Progress in AI has been exceptionally fast over the past few years. Since the initial release of ChatGPT, there has been an emergence of new capabilities, such as AI systems that can process and interpret information from multiple modalities, including images and speech, as well as new, sophisticated ways for users to interact with an AI system, like deep research tools and agentic AI. Yet, despite these leaps, these innovations have yet to reach their full potential in education.
While AI has rapidly advanced, its integration into education doesn’t always align with the specific needs of today’s classrooms. Vision models, for instance, are highly capable at visual common sense reasoning, acquiring knowledge from real-world images to answer questions. However, they often struggle with tasks that require interpreting angles or comparing visual lengths, skills that are essential in subjects like math. This limitation reduces their practical usefulness in certain educational contexts. Similarly, audio models have seen substantial progress in automatic speech recognition, but most of these systems have been trained primarily on adult voices. As a result, they tend to perform poorly when it comes to detecting children’s speech, an ability that is crucial for developing tools that effectively support early literacy in young students.
Even the most promising approaches for broadening the utility of AI systems, such as tool calling, where AI can access external resources like the web or calculators to enhance its responses, face limitations in education. This is largely due to the underrepresentation of educational tools designed for this kind of integration. In contrast to software development, where coding co-pilots have become a staple, education lacks similarly AI-powered tools, such as interactive whiteboards (e.g., Excalidraw) or mathematical plotters (e.g., GeoGebra), that could support dynamic learning experiences.
But this isn’t just about AI falling short in education because its current capabilities aren’t fully suited to the field’s unique challenges. It’s also about how education itself can drive AI innovation. While education presents real challenges for AI, such as addressing bias and safeguarding student privacy, it also offers a rich source of inspiration to push AI research in new and valuable directions.
How Can AI Personalize Learning?
AI has the potential to create personalized learning journeys, adapting to each student’s understanding, pace, and needs. But to achieve this, the field must evolve beyond generic applications and embrace solutions specifically designed for education, developing models informed by how students learn, how educators teach, and the unique dynamics of the classroom.
To bridge the gap between AI’s promise and its current impact in education, a more focused and technically ambitious approach is needed. Several strategic directions point the way forward.
One of the most important is incorporating modeling for how and what students learn over time, a task commonly known as knowledge tracing. This functionality, paired with the existing abilities of LLMs, is the foundation for creating smart learning tools that can effectively adjust to each student’s needs, like a tutor that knows when a student is struggling or ready to move ahead. Researchers are now exploring how large language models (LLMs) can use their internal knowledge to make a more informed prediction about how well a student understands a concept. For example, if the model finds a student’s answer surprising, it might assume the student doesn’t fully understand the topic. These ideas are exciting, but are still in the early stages. Models along these lines are harder to train, require significant computational power, and need more work to make them reliable and accurate. Moving forward will require deeper research and smarter designs to better connect what these AI models “know” with the learning process of real students.
But this isn’t just about AI falling short in education because its current capabilities aren’t fully suited to the field’s unique challenges. It’s also about how education itself can drive AI innovation. While education presents real challenges for AI, such as addressing bias and safeguarding student privacy, it also offers a rich source of inspiration to push AI research in new and valuable directions.
How To Assess AI-Assisted Learning
Another key direction is figuring out how to tell the AI what “doing a good job” actually means. In subjects like math, AI has a clear goal, success is easy to measure because there are clear right or wrong answers. But in education overall, there isn’t one simple way to assess whether meaningful learning is taking place. There are some helpful clues, like whether a student is asking thoughtful questions or following instructions well, but there’s no standard system that tells the AI, “This behavior means the student is learning effectively.” What’s needed is a way to teach AI what successful learning looks like, grounded in solid educational research, so it can optimize for real student progress. Creating these kinds of learning “signals” would be a major breakthrough and could better align AI training, producing more useful models for education as a result.
Infrastructure is also a critical direction. Embedding AI into classrooms requires purpose-built systems: datasets that reflect real educational use cases (e.g., group work sessions), frameworks that support interactions like drawing on whiteboards or plotting equations, and models that account for both academic content and student learning differences. By looking at how students respond to both human and AI teaching methods, we can figure out what works best and focus on the most helpful improvements.
What’s needed is a way to teach AI what successful learning looks like, grounded in solid educational research, so it can optimize for real student progress. Creating these kinds of learning “signals” would be a major breakthrough and could better align AI training, producing more useful models for education as a result. Infrastructure is also a critical direction.
The rapid evolution of AI has proven what’s possible, but possibility alone isn’t enough. For education to benefit meaningfully from these advances, the field needs to invest in the hard work of tailoring models, building dedicated tools, and defining success metrics and signals in educational terms.
By focusing on real classroom situations, understanding how students learn, and giving clear feedback, we can make AI tools that don’t just give information but actually help students learn better.
