Bots Give Teachers Superpowers Says Ramesh Balan

In this 2022 column, Ramesh Balan argues that lesson bots, which follow a fixed curriculum, can give teachers extra tutoring capacity and are different from artificial intelligence. The section added below looks at what controlled studies of AI tutoring have found since then, and at how Balan’s distinction applies to tools teachers can buy today.

We have all had a full year of in-person instruction now. It’s pretty clear that the 2020-21 school year of remote learning put a good portion of our students behind instructionally. But, I’m a glass-full kind of person, so while I’m concerned about that year of learning loss, I also recognize that there were benefits to being forced into remote education. During the 2020-21 school year, we learned more about ourselves, each other and our education system than we could have in a lifetime. The most significant thing I learned is that we need more teachers and we need to lighten their load.

Even if we overlook teachers’ responsibilities as counselors, tech coaches, disciplinarians, managers, entertainers, and subject matter experts – asking them to bring students back up to speed instructionally without added class time or more resources isn’t fair, or even reasonable. We either need to hire more teachers, or we need to pull something off their plate.

For the most part, our education system is on the right track by funding tutoring programs and giving extra support to struggling students. I’m also very encouraged by the focus on the mental health of students and teachers. What is concerning, though, is that funding for tutoring is a stop-gap solution at risk of being defunded and leaving teachers in the lurch.

Limitations of Live and Chat-Based Tutoring

There are also a few shortcomings related to live and chat-based tutoring to keep in mind, meaning that this is not the entire solution we need. 

First, not all of these tutoring programs are aligned closely with existing curriculum, creating a disconnect between how much the tutor is actually able to help. Being able to open an app or call for a tutor from the school’s website or LMS is great, but unless that tutor knows the curriculum, there are limitations. For example, the tutor might teach the progression of long division differently than the school district has specified in the progression of math. Or, a tutor might not know the subtleties in language used for social studies. I’m not against live and chat-based tutoring but we do need to recognize its limitations and compensate.

It also isn’t always possible to get the same tutor, which when you are a young, disenfranchised learner, not seeing a familiar face or name could be the reason why you don’t reach out in the first place. It’s not a natural inclination for young children or even teens to turn to the school’s learning resources when they feel uncertain. They’re far more likely to Google for the answer or ask a classmate as they are to find the tutoring app and ask for help. 

Bot-Assisted Tutoring Aligned to the Curriculum

This is why my company has been refining the process of bot-assisted tutoring for years. We understand human nature and how students tend to look (or not look) for help. We’ve been working on a bot-assisted tutoring model to address this since well before the pandemic. Our lessonbots give instant feedback, and are a mechanism for guided 1:1 tutoring. We’ve figured out how to use our bots to work their capabilities on tutoring a student on any curriculum, including anything the district or individual teacher chooses to teach. In addition to guided tutoring, the bots also assess student learning progress, provide instant feedback, capture step-by-step of student learning, and overall provide a multi-modal interactive learning experience for students. 

Bots are not artificial intelligence as is often assumed. Rather, bots are programmed to take on tasks to make them speedier and able to be replicated with fidelity. That’s what makes them perfect topic-specific tutors. Bots can be programmed to take a district’s curriculum, and provide guided, specific remediation at the exact moment a student is struggling with the material. Bots can even recognize when a student is struggling which solves that very big and very real problem about students being reluctant to ask for help, not to mention that children often don’t recognize that they are struggling.

Bots are perfect tutors because they are trained with a finite set of data to do only one thing. Conversely, AI is trained on a large pool of data so that it learns to make decisions of an unknown outcome. AI isn’t consistent in how it responds to a query because it is always learning, which is what you want it to do except when it is supposed to teach something one specific way. This is, incidentally, one of the problems with wide-scale online tutoring providers; there’s no guarantee of consistency. Bots are trained on a specific task with specific data to do a specific thing, time and time again. 

In the context of teaching, bots are super helpers. They can be trained to react a certain way to an input or, for example, to answer a math problem. Bots can help a student who missed a step in long division by walking them through the correct steps all the while following the exact progression in the district’s curriculum.

Helping Teachers Reach Every Student

After all, this is how a teacher would work. If a teacher were at the student’s side – every student’s side – she would explain every step as many times as was needed for each individual. She’d watch each student’s eyes for telltale signs of understanding or loss and if she needed to explain it again, she would. She might ask the student to work out the problem in front of her several times until she was sure the process was understood. Then when that student had it, she’d move on. 

But bots can do this. Bots give teachers superpowers. 

Granted, these are very specific limited actions but when applied in the core curriculum, bots can give targeted intervention to each individual student where they have made a mistake. It works right at the point of learning, without delay, fatigue or error. It’s personalized tutoring perfectly aligned with the curriculum because it is trained on the curriculum and it is 100 percent faithful to its intended purpose – to help teachers do more.

What’s most exciting is that while our Knomadix bots can assess how a student answered a question and give contextualized feedback to get them the right answer, soon we’ll be able to do more. In the future, Knomadix bots will recognize facial cues and proactively provide assistance before a student even realizes they are struggling. Bots will notice an issue on one step of the learning progression before it escalates into failing to learn the entire concept. After all, that’s what a teacher does. She watches her students and recognizes the little hiccups before they progress into outright failure. 

We’ll never replace teachers but giving them back some precious time can put a little lift back in their capes.

AI tutoring bots: what studies show about tutors that follow a curriculum

AI tutoring bots are now a common topic in school offices, and Ramesh Balan’s 2022 column above anticipated one of the central questions: how can a tutoring tool stay faithful to what a district teaches while still helping each student at the moment of difficulty? Balan’s answer was a lesson bot trained on a finite body of material, which Balan separated from artificial intelligence because AI answers can vary from one response to the next. Since then, researchers have run controlled studies of tutors built on large language models, and their designs show how much the question of consistency matters.

A randomized trial of an AI tutor

The most cited study is a randomized trial at Harvard, reported in Scientific Reports in 2025. In an introductory physics course, 194 eligible students each studied one topic with an AI tutor at home and another in a class session taught through active learning, in a crossover design. The authors report that median learning gains with the AI tutor were more than double those of the in-class lesson, with an effect size of 0.63 by linear regression and 0.73 to 1.3 standard deviations by another method, and that students spent a median of 49 minutes with the tutor against an assumed 60 minutes in class. Students also reported higher engagement and motivation.

What matters for Balan’s argument is how the tutor was built. According to the authors, it ran on a system prompt and detailed step-by-step solutions, and it walked students through each part of each problem in the order the instructor used in class. Its answers relied on pre-written solutions, and its feedback was reactive, responding only to what a student entered. In this section’s reading, not the authors’, that design is close to the consistency Balan wanted from a bot: the language model stayed within the instructor’s sequence and solutions.

The authors list limits. The study covered one course at one university with a mixed student group, used GPT-4 and expert-written prompts for each question, and tested content at the levels of understanding, applying and analyzing. The authors do not claim AI will always beat active learning, and they note it may fall short on complex synthesis and higher-order thinking. A K-12 leader should not read it as showing the same gain for a fourth grader.

AI that supports human tutors

A second line of evidence concerns tools that support the tutor, not the student directly, which is closer to the “superpowers for teachers” in Balan’s title. K-12 Dive reported in October 2024 on a Stanford study called Tutor CoPilot, which involved 900 tutors and 1,800 elementary and secondary students in a large school district in the South. Students whose tutors used the tool were four percentage points more likely to pass math session assessments, and students of lower-rated tutors gained up to nine percentage points. The cost was estimated at about $20 per tutor per year based on usage in the study. The article reports no long-term outcomes and results from only one district.

Where Balan’s column needs updating

Balan stated that bots are not artificial intelligence, and in 2022 that described scripted, topic-specific tutors. The Harvard tutor above ran on GPT-4 with prompts and pre-written solutions, which suggests the practical question has moved from whether a tool counts as AI to how tightly it is controlled. Balan also predicted that bots would recognize students’ facial cues and offer help before a student realizes the difficulty. That would involve collecting biometric information from children, which raises the privacy issues covered in the edCircuit article on personally identifiable data, and the studies above did not test it.

Questions for a district considering a tutoring bot

1. Is it aligned to our curriculum, and can a teacher see which steps and vocabulary it will use?
2. How does the vendor keep answers consistent, and what happens when a student asks something outside the lesson?
3. Is there a controlled study of the product itself, and what grades, subjects and group sizes did it cover?
4. What student data does it collect, and where does it go?
5. Does it give teachers a record of what each student attempted, so the tool adds to a teacher’s time and does not use it up?

The edCircuit AI in education guide covers the wider decisions around these tools. Balan closes by saying bots will never replace teachers, and the studies above, in which the AI worked within a structure that human instructors designed, are consistent with that.

More on Tutoring and Homework Help