AI and the Future of Personalized Education at Scale – Part 1

Edan Shahar wrote this first part of a three-part series in 2017, arguing that artificial intelligence could help teachers personalize learning for large numbers of students while teachers stay central. A new section after the text looks at what a randomized trial and a national teacher survey have shown since.

How do you make one classroom conducive to the learning needs of every student?

Editor’s Note – This is Part One of a Three-part series

One of the greatest challenges faced by teachers is making lessons accessible, engaging, and impactful for every student in their classroom. Whether it’s a math or an English class, a science or a history class, students all bring different backgrounds, perspectives, and abilities to the subject at hand.

In many ways, this is an essential part of a good education: that it brings people with different backgrounds and perspectives into the same room and allows them to learn from each other. However, it also poses a significant problem for teachers, who are tasked with helping every student learn, leading to the great challenge of teaching – how to make one classroom conducive to the learning needs of every student.

Schools and teachers have employed various strategies to address this problem. A few of these include:

• Students progressing to more advanced classes once they demonstrate mastery of the previous levels.
• Smaller class sizes allowing teachers to more readily address the needs of individual students.
• Different levels of coursework at each grade level to allow differentiated learning.

Each of these solutions is aimed at the same thing: personalized learning for students. True personalized learning, as per the definition crafted by the Bill and Melinda Gates Foundation and others, necessitates continual assessment, adapting environments, and personal learning paths and profiles.

Of course, achieving a customized learning plan for one student is not difficult for a teacher equipped with the right tools. But so far no one has found a reliable way to help one teacher customize learning paths at scale, meaning for dozens or even hundreds of students. That is where technology has the greatest potential to make an impact: scale.

Technology has immediate and clear advantages in this endeavor. Artificial intelligence (AI) has the ability to quickly and easily identify student strengths and weaknesses and constantly adapt the student learning profile and plan based on progress. While substantial funds have been devoted to researching solutions, breakthroughs have been limited. Today, only a handful of education technologies have demonstrated a meaningful impact on students using personalized learning paths.

But the promise of AI and deep learning in this space is significant. As the next two installments in this series on AI in education will emphasize, it is a matter of when (not if) we will find workable applications of this technology across the spectrum of education.

The reason for this is simple. As humans, we are gifted in judging the subjective and the anecdotal. That means we can see curiosity, ambition, and intellect in individuals. We cannot, however, easily see data patterns and quickly analyze large data sets.

As an example, a teacher can review the results of a problem set and, given enough time, note that a certain student struggled with understanding variables. AI can not only instantaneously identify these content weaknesses, it is further able to correlate that student with a group of hundreds of other similar students also having difficulty with variables. Using the historical data on all of these students, AI can recommend exactly which steps lead to the greatest progress. These conclusions are possible because AI can see beyond the subjective and identify and process information that we simply cannot.

But let’s be clear: AI does not have the ability to replace teachers. Instead, it is best viewed as an indispensable ally in education. Teachers are the ones who inspire, motivate, and encourage learning. As long as the student is a person, the best teacher will be as well.

We owe it to teachers and students alike to explore advances in AI that can make education more effective for everyone.

Personalized learning at scale with AI: what the evidence shows so far

Edan Shahar’s 2017 article asks how one teacher can adapt learning for dozens or hundreds of students, and answers that AI could do the analysis that people cannot do quickly. Personalized learning at scale was a promise in 2017, and several years of use and research now allow a more careful assessment. The most useful evidence comes from one randomized trial of a tutoring tool and from a national survey of how teachers actually use AI.

A randomized trial of AI that supports tutors

A preregistered randomized controlled trial of a system called Tutor CoPilot, first posted on arXiv in October 2024 by Rose Wang, Ana Ribeiro, Carly Robinson, Susanna Loeb and Dora Demszky, tested a design that matches Shahar’s description of AI as an ally for people. The system gives tutors real-time guidance during live sessions. The study involved 900 tutors and 1,800 K-12 students from historically under-served communities. According to the abstract of the current version (a revision posted in October 2026), students whose tutors had access to the tool were 4 percentage points more likely to master topics, and students of lower-rated tutors gained 9 percentage points. The abstract puts the cost at about $20 per tutor per year, and reports that tutors with access were more likely to ask guiding questions and less likely to give away answers.

The result has limits that readers should keep in mind. The tool supported human tutors working in sessions and did not replace a teacher, which is consistent with Shahar’s view. The study concerns one setting and one tool, and the paper reports that tutor interviews raised problems, including suggestions that were not appropriate for the grade level. A district should not read it as proof that AI personalizes learning in a regular classroom. It does show that a tool built to support a person can produce measurable gains when the people using it are given real help at the moment they need it, and that the cost of such help can be small.

What teachers report doing

A RAND report covered by WJLA in February 2025 found that 25 percent of teachers said they had used AI, about the same as in RAND’s fall 2023 surveys. Nearly two-thirds of those teachers said they used it only for lesson planning, about 10 percent only with students, and about a quarter for both. Examples included adjusting reading levels and summarizing long articles. RAND researcher Julia Kaufman said AI may make teachers more efficient, but that researchers have no idea how much it improves the quality of their instruction. RAND called for more research and coordination to develop uses supported by evidence.

The coverage suggests that the personalized, student-facing use Shahar described is still uncommon. Teachers are mostly using AI to prepare materials, which is helpful but different from individualized instruction at scale. Principals used AI more, at nearly 60 percent, mostly for communications such as parent emails. That difference matters for planning, because the people who decide what a district buys may use AI in a different way from the teachers who would use it with students.

What school leaders can decide

• Define what personalization means in the district. Shahar cited a definition requiring continual assessment, adapting environments and personal learning paths. A product should be judged against whichever definition the district adopts.
• Ask vendors for outcome evidence from independent studies, and for the setting and age group in which it was gathered.
• Keep a teacher involved in any tool that sets a student’s path, as in the Tutor CoPilot design.
• Pilot with a small group, compare with a similar group not using the tool, and look at student results as well as satisfaction.
• Give teachers guidance. The survey coverage says many teachers use AI on their own without administrator direction.

Shahar’s central claim, that people judge the subjective well and machines can find patterns in large data sets, is a reasonable description of a division of labor. The evidence so far supports AI as help for a person more than AI as a replacement for one. Readers can continue with the series in Future of AI in Education and the articles on empowering teachers with AI and virtual tutors and personalized learning.

For a wider look at AI personalized learning, including the research and policy questions, see the edCircuit overview.

Further Reading

1. Digital Trends – That’s ‘Professor Bot’ to you! How AI is changing education
2. Forbes – Artificial Intelligence In Education: Don’t Ignore It, Harness It!
3. Education Dive – AI, analytics transforming guidance counselors’ roles

Related reading

• Future of AI in Education
• The Future of Education: Empowering Teachers with AI
• AI in Education: From Virtual Tutors to Personalized Learning

More on Artificial Intelligence