For the past couple of years, the conversation around AI in education has focused on what these tools can do. They can answer questions, write essays, explain concepts and hold a conversation about almost anything. But that is a party trick, not a teacher.
In 2026, the reliability bar is finally being crossed. Reasoning models can now work through a math problem step by step and be right consistently, not just sound confident. At the same time, the cost of that intelligence has dropped significantly. Put those two things together, and for the first time we have the opportunity to give every student a personal teacher that is both accurate and affordable at scale.
The question is no longer, “Will kids cheat with AI?” It is, “Can AI actually teach?”
The First Generation of AI Got Tutoring Wrong
The first generation of AI in education took a general-purpose chatbot, pointed it at homework and called it a tutor. However, those chatbots are designed to be directly helpful through providing immediate answers. And giving a student the answer is not the same as teaching them how to arrive at that answer.
Teaching requires understanding where the student is struggling and helping them work through it. That is where productive struggle becomes important as it gives the student enough help that they feel the solution is approachable, but they still need to do the thinking for themselves. Maybe that means a leading question, a hint or, if they’re truly stuck, it eventually means giving them more direct guidance. There is always a balance. If you make the struggle too difficult, the student can simply close the tool, but if you make it too easy, they do not learn anything.
That is why the second generation of AI-powered education has to be purpose-built around learning. These tools need to understand the material and the student, as well as known when to guide, challenge or step in to help the student.
Unlike other general-purpose AI tools, the next generation of educational AI cannot use time spent on the app as its metric of success. In addition to being built specifically to enhance learning, these tools should be aligned to what students are actually expected to know based on their grade and location, capable of understanding different levels of difficulty and ultimately judged by whether a student develops mastery. It should not matter how much time the student spends interacting with it.
Engagement is easy to measure. We can count sessions, messages, streaks and minutes spent on a platform. But a student spending twice as long with an educational product does not equate to a student learning twice as much. Learning is harder to measure than engagement, which is precisely why we should be investing more in measuring it.
Personalization Has to Mean More Than Changing the Response
The idea of “adaptive learning,” namely customizing learning experiences to address the unique needs of students, has become a phrase used so often in EdTech that it risks losing its meaning. How so? Because you cannot properly adapt instruction to a student unless you first understand the material itself.
Consider a seemingly simple request: generate five questions about linear equations. Linear equations can be taught across multiple grade levels. Within the same grade, a question can test basic recall, require several computational steps or demand substantially more complex reasoning.
Before an AI system can decide what a particular student should see next, it has to understand those differences. Was the question the student missed easy or difficult? Was it a simple recall error or a misconception in a multistep process? What has that student demonstrated previously on similar concepts? Only then can the response meaningfully adapt.
This is where learning science matters. The industry has invested enormous energy in making AI sound natural. We now need comparable energy devoted to making it pedagogically sound.
AI Will Not Eliminate Teachers
Whenever AI and education are discussed together, the conversation eventually turns to whether technology will replace teachers. Personally, I think that frames the problem incorrectly.
Education has historically been designed around group instruction, largely due to human resource constraints. Since one-on-one instruction is impossible in a class of thirty, classrooms have been built around a model in which one person teaches to a group at roughly the same pace. But not every student learns at the same pace. Individualized instruction is imperative for student success, but it is hard to provide at scale.
AI potentially changes that constraint. This shift could give teachers more room to focus on the parts of education that cannot be automated, such as building relationships with students, keeping them motivated, bringing lessons to life and creating opportunities for students to learn from and work with one another.
In some ways, we currently have the model backwards. Students receive group instruction from their teacher and then turn to AI when they are alone and stuck. A better model may increasingly use AI to provide individualized instruction and practice so the teacher can focus on motivation, mentorship and connection.
Buyers Need to Raise the Bar
As the market matures, parents, educators and district leaders should expect more from products carrying an “AI tutor” of “AI teacher” label.
Accuracy should be verifiable, particularly in STEM subjects where a plausible-sounding wrong answer can reinforce misunderstandings.
Adults should also have visibility into what students are doing. If a parent or educator cannot determine where a student is struggling, how the system responded or whether the student is progressing, the technology becomes a locked box.
And perhaps most importantly, buyers should examine the behavior the product encourages. Does it immediately hand over final answers? Is success measured primarily through time spent with the tool? Does the system help students reason through difficulty or remove the difficulty?
From the Average Student to the Actual Student
By 2030, I expect the biggest challenge will be that the idea of teaching toward an “average student” begins to feel increasingly outdated.
Imagine a student walking into school with a learning system that already understands which concepts they have mastered, which misconceptions keep recurring and where they need additional practice. The student no longer has to remain quietly lost for several weeks until a test reveals that they fell behind.
Classroom time can then become more human, not less: discussion, projects, collaboration, mentorship, motivation and application, paired with individualized instruction and practice that would previously have required a private tutor for every child.
That could also have profound implications for educational access. Scalable, high-quality individualized support has the potential to make what happens after school less dependent on a family’s resources.
For generations, we have grouped students largely by age, moved them through material at roughly the same pace and accepted that some will be bored while others struggle to keep up. That system was built partly around the limitations we had. AI gives us an opportunity to reconsider those limitations.
The defining question for EdTech is therefore no longer whether we can put AI into education. We already have. It is whether we are willing to build educational AI around the way students learn rather than simply around what the technology can generate. That will determine whether this next generation of EdTech becomes another collection of impressive tools or actually changes education.
