Now that August has arrived (man, that was fast!), the back-to-school season vibes are stirring and kids have to prep themselves for another year of in-person studies.
With the continued evolution of AI tools, teachers and students alike have been using them to assist them in their everyday learning. While it’s disappointing to hear stories about students producing their work entirely with AI, it’s equally satisfying to see AI tools such as Gemini Notebooks, Perplexity, and ChatGPT offer educational-themed features that don’t encourage cheating.
Speaking of AI tools built to educate students, one of the most prominent ones is StarSpark, an AI-powered math and science platform that’s designed to teach rather than simply give out answers. It primarily covers K-12 and AP coursework while adapting to a student’s specific grade level, learning style and the knowledge they’ve already acquired. And for parents, StarSpark is just as beneficial for them—the AI educational tool provides them with accessible, on-demand support that helps them grasp the way math and science are now taught.
Ashish Bansal, a father and the developer behind StarSpark, spoke with Tom’s Guide to chat about his AI educational platform, the biggest mistakes students make with general chatbots and more.
StarSpark’s impact on educating students, how it helps parents and more
Tom’s Guide: The AI tutoring space is becoming crowded. What did you believe was missing from existing AI education tools that convinced you to build StarSpark?
Ashish Bansal: It's personal to me. I'm a dad of two, and I spent 15 years building AI at Google and Amazon. When I went looking for something to help my own kids with math, everything I found was really the same thing underneath. A general chatbot that hands over the answer, or an education-branded tool that was a thin wrapper over the same model. There was no system actually purpose-built for learning.
That matters more than it sounds. These general tools were built for professionals to get work done, and they're optimized for engagement. Optimizing for engagement pushes a model toward telling you what you want to hear, and sycophancy is a very different objective than teaching. A good teacher tells a kid when they're wrong. And the “this is AI, it can make mistakes” disclaimer that comes with these tools is fine for an adult professional, but it's cold comfort for a learner, because a student is the one person not in a position to catch the mistake. If they knew enough to spot the error, they wouldn't need the tutor.
The deeper issue is that solving is not teaching. It's the player-coach fallacy. Being able to solve a problem instantly, which is what these models are great at, is not the same skill as being able to teach a child to solve it. Some of the best players make poor coaches. Handing over a correct answer and building understanding are two different jobs.
So, I built the tutor I wished my kids had. Purpose-built for learning, grade-aligned, Socratic, and designed to guide a student step by step instead of doing the work for them.
Tom’s Guide: Have you seen surprising moments where students developed confidence or problem-solving skills they weren't showing before?
Bansal: All the time, and it's the best part of this job. The pattern that surprises me most is when a kid stops asking for the answer and starts wanting to find it themselves.
A few real sessions stick with me. One young student was working through addition with regrouping, got stuck twice, and pushed through it. She ended the session by typing, “I think I'll understand math now.” That one sentence is the whole reason we built this.
Another was a high schooler on a tough number theory problem. He tried six different wrong combinations. StarSpark never handed him the answer; it just kept nudging him at the exact step where he was off. He arrived at the right answer on his own, and by the end he actually understood divisor formulas, not just that one problem.
My favorite might be the student who kept typing “easier, easier, easier,” clearly wanting to give up. Instead of caving, the tutor held the line, and a few minutes later she made up her own problem and solved it. Confidence usually shows up as persistence first.
Tom’s Guide: Everyone is worried AI is making students dependent on it. How did you design StarSpark to teach students how to think instead of giving them the answers?
Bansal: This is the exact failure mode we designed against, so it's built into the core of the product, not bolted on.
The research is pretty clear. Students who use AI to get answers lose problem-solving ability over time. Students who use it for hints keep it. So the whole design is built around productive struggle, which is the idea that real learning happens when a kid is working at the edge of what they can do, not when the work is done for them.
In practice, that means StarSpark gives hints slowly. It doesn't dump a full solution. It offers the smallest nudge first, then a slightly bigger one, and only goes further if the student still needs it, so the kid is always doing as much of the thinking as they can. There's no “just give me the answer” mode to toggle into when they get frustrated.
The other piece is proactive intervention. The tutor watches how a student is working, and when it detects that they're genuinely stuck on a question, it steps in on its own and teaches the underlying concept right then, at the exact step where the misconception is. If they get something wrong, we don't just mark it wrong. We do a two- to three-minute reteach and have them try again, as many times as they need.
The goal is to protect the struggle, not remove it. That's the difference between a tutor and a crutch.
Tom’s Guide: What are the biggest mistakes you see students making with general chatbots like ChatGPT? How does StarSpark solve them?
Bansal: The biggest one is obvious. A kid types in the problem, copies the answer, turns in the homework, and learns nothing. It feels productive and it isn't. It shows up on the next test.
A subtler mistake is trusting the explanation. General models can produce steps that look right but are subtly wrong, and in math a wrong step is a real problem. Kids don't have the expertise to catch it.
And a third is grade mismatch. Ask a general chatbot a fifth-grade problem and it might give you a technically correct college-level explanation that goes right over the kid's head.
StarSpark is built to solve all three. It won't hand over the answer, so the student stays engaged. It's pedagogically aligned through grade alignment, so it teaches a concept the way it's actually taught at that grade. Part of that is language. We fine-tune how our tutors talk to match the student's level, so a fifth grader gets shorter sentences and simpler vocabulary, while a ninth grader gets a more advanced explanation. Same math, taught in the words that kids can actually absorb. And we do step-level checking to guide the math rather than free-generate it, which reduces the chance of a confidently wrong step. It's the difference between a tool built to answer anything and a system built to teach one thing well.
Tom’s Guide: Parents say they don't recognize the way math is taught now. What problems have you heard, and how does StarSpark help?
Bansal: I hear this constantly, and honestly I felt it myself. A parent sits down to help, sees area models and number lines and three different strategies for one problem, and thinks, “this isn't how I learned it.” Sometimes the kid even gets marked down for reaching the right answer with the “wrong” method. It leaves parents feeling locked out of helping their own child.
Two things I hear most: “I don't understand these new methods,” and “I don't want to teach it wrong and confuse them more.”
StarSpark takes that pressure off in two ways. First, we make the syllabus transparent to both the parent and the student, so everyone can see what the child is supposed to be learning and where they are in it. There's no mystery about what's being covered. Second, we offload the pedagogy itself. Our tutors carry the knowledge of how a topic is actually taught at each grade. To give one example, factoring a quadratic is introduced and expected differently in grade 7 than in grade 8 or grade 9, and the tutor knows that difference, so it teaches the method the way that kid's class expects, with no conflict with the teacher.
And because it shows the steps and where the kid went wrong, not just a checkmark, a parent doesn't need to be a math expert to stay involved. My advice to parents is simple: you don't need to know the new method. Just sit nearby and ask your kid to walk you through their thinking. That alone does a lot.
Tom’s Guide: How do you see AI improving education over the next five to 10 years? Will every student eventually have a personal AI tutor?
Bansal: I do think we're heading toward every kid having an always-available tutor, and I think that's one of the most hopeful things happening in education. An AI tutor for every student in the world, in every language, is our mission. That's the future we're building toward.
There's a well-known idea from the education researcher Benjamin Bloom back in 1984. One-on-one tutoring moves the average student about two standard deviations above a normal classroom. It works incredibly well. The catch has always been that you can't afford a human tutor for every child. That's the problem the field has been stuck on for 40 years, and AI is the first technology that gives us a real shot at solving it at scale.
The other shift I'm excited about is assessment. We tend to treat assessment as judging a student, a grade on a report card. I think its real purpose is uncovering the missing concept that's holding a kid back. Because a good tutor can see the whole process - the steps, the retries, the work a kid talks or writes out - we get a window into how a student actually thinks, not just their final answer. That's a richer and more honest picture than a test.
And because the AI can give that feedback in real time, it can run assessment and remediation in a closed loop. It spots the gap, teaches to it right then, checks again, and keeps going. That's what real personalized learning looks like, and it finally lets a student move at their own pace instead of the pace of the class.
That points to the hard question underneath all of this, which I raised on the EdTech Insiders podcast recently. Are we actually ready for personalized education? Can we run a classroom where every kid is at a different point in the syllabus? That's the core problem we have to solve, and it's as much about how schools are structured as it is about the technology.
I'll be candid, though. This augments great teachers; it doesn't replace them. Motivation is a good example. A big part of why a kid pushes through hard math is the relationship with a teacher they don't want to let down, and that's deeply human. An AI can be an endlessly patient practice partner, but the teacher is still the one who inspires the kid to care in the first place. The right division of labor keeps the human relationship central. And the long-term efficacy research is still being done, so I'd rather be honest about what we've proven and what we're still proving.
Tom’s Guide: If you could give every parent one piece of advice about using AI to support their child this school year, what would it be?
Bansal: Pick a tool that teaches, not one that tells. That's the whole thing. The good ones ask questions and give hints. The bad ones just hand over the answer. The quickest test is to watch your kid use it for two minutes. If they're working and explaining and doing the thinking, it's helping them. If they're copying an answer off the screen, it's hurting them, no matter how polished it looks.
If you want a few concrete things to check before you commit to a tool:
Is it purpose-built for learning, or is it a productivity tool with a “learning mode” bolted on? Those are not the same thing.
Is it aligned to your child's grade and curriculum, so it teaches the way their class teaches?
How good is it at giving feedback on your kid's actual work, not just marking it right or wrong?
Is it multimodal? Math lives in images, illustrations, and handwriting. A kid should be able to draw and write out a problem, not just type into a chat box.
Used right, AI is a patient tutor that's there whenever your kid needs it. Used wrong, it's a shortcut around learning. Those checks are how you tell the difference.
Follow Tom's Guide on Google News and add us as a preferred source to get our up-to-date news, analysis, and reviews in your feeds. Subscribe to Tom's Guide on YouTube and follow us on TikTok.
More from Tom’s Guide
- Millions of books are reportedly being destroyed to train AI — here’s what Anthropic told Tom's Guide
- I showed my retired parents how to use Claude AI — here are 5 prompts that changed their daily routine
- Everyone is using ChatGPT Memory wrong — 5 tweaks to make it a true assistant
