Two Conversations That Changed How I Think About AI in Education and Learning

A message at 3:30 in the morning

It started with a thumbs-up.

My friend Jin Oh Bae, someone I hadn’t talked to in a long time, was in bed in Hawaii at 3:30 in the morning, struggling to finish a LinkedIn article. He noticed I had liked one of his older posts and thought, I wonder how he’s doing. So he sent me a quick message and went to sleep. I wrote back almost right away.

The timing was funny. On our side, Pooja Rupaneeja, our content writer, had been wondering how we could create original, honest content in a world that is filling up with AI-written text. Jin’s message showed up right in the middle of that question. A few emails later, we had two calls on the calendar, Pooja as host, and a list of questions.

I expected a warm catch-up. What I got were two of the most useful conversations I’ve had in years. This is the story of what we talked about, and what I can’t stop thinking about when it comes to AI in education and the future of learning. 

I should say where I stood going in. I believed that more realistic practice was the cure for test anxiety, and that AI is a tool that should never replace a teacher. Jin challenged one of those beliefs and sharpened the other.
‍

Call one: the student who knew everything and still lost her nerve

Jin moved to Seoul in 2010 to connect with his Korean roots. He planned to stay two years. “The two turned into 14.” For 14 years he taught SAT, ACT and AP test preparation in Apgujeong and Daechi-dong, Seoul’s most competitive tutoring districts, and he fell in love with his students. Today he is doing a PhD in self-leadership. Then, with a laugh, he gave us the line of the day: “I went there to learn about my culture, but I ended up learning about myself.”

My side is the software side. I started in AI research, co-founded TeachEdison in 2015, and launched EdisonOS in 2020. Today about 150 tutoring companies, community organizations and school districts use it to prepare students for the digital SAT, ACT, AP and other tests.

Pooja opened with Jin’s recent article on perfectionism and the SAT. His central idea: perfectionism isn’t a flaw inside students. It’s something their environment teaches them, often contributing to the pressure and SAT test anxiety student experience.

He described a student who scored in the 720s on practice tests and then got a 660 on the real thing. She hadn’t forgotten anything. She walked in needing every question right, and the pressure did the rest. In Korea, he explained, rankings can be public, competition is intense, and students will quietly hide which academy they attend. Inside the test, his students would “focus on the words they didn’t know,” instead of using everything else on the page.

His answer was a goal that fit a human being: fewer than five wrong, on five practice tests in a row. A student who started at 20 wrong didn’t aim for five. They aimed for fewer than 13, three tests in a row, and then lower. A 99th percentile score doesn’t need perfection, and some questions on the test don’t even count. Sometimes, he said, “you got to let someone go.” And then the line I’ve repeated to my whole team: “mastery isn’t perfection.”

I had to admit something to Jin. For years, my instinct as a product builder was more simulation. Make practice look like the real test. But Jin’s students did great on daily practice tests and still dipped on test day. Practice does help. In a survey of about 1,500 students, 72 percent said weekly retrieval practice made them less nervous. Still, rehearsing the test isn’t the same as making peace with it.

It’s the same trap in my world. Teams over-invest in one feature and polish a detail no one asked for. I remind my team that none of us takes the SAT or teaches it. We’re not the users, so we listen, and we say no to a lot. Jin summed it up: mastery is “so correlated with self-knowledge.”

It has also changed what I think software should do. Jin’s target moved with the student, from 13 wrong to fewer than five. A dashboard can show a number, but only a tutor who knows the kid can decide which number to aim for. Our job is to make that conversation easier, not to replace it.

Then we turned to tests themselves. I think test changes are healthy. Universities have limited seats, tests are a pragmatic filter, and a pattern that never changes just becomes something tutors learn to game. The digital SAT now runs about two hours and fourteen minutes and adapts to your answers. Jin understood the case for it, since students are busier than ever, but worried it might cost them stamina.

Then came the news that had caught his eye. More than 3,000 University of California faculty, including five Nobel laureates, signed letters asking UC to bring back SAT and ACT requirements. Their worry is foundations. And the national numbers explain why they’re uneasy: only about a third of high school seniors were academically prepared for college-level math. Jin said the SAT and ACT test fundamentals, and that even the hard questions are “more advanced concepts of basic foundations.” I told him what I believe: making the test easier is the wrong fix. We should help kids build the foundations.

We ran out of time. We both knew we had only started.
‍

Call two: the confession

A few days later, we were back and the first question of this call was a challenge. AI can make endless practice questions and explanations on demand, so why would a game or a simulation be worth a student’s time?

Jin’s answer was one word: systematic. After 14 years, he built a sequence of learning for students whose first language isn’t English. He teaches foundational knowledge first, such as an independent clause or a modifier, building the strong academic foundations students need before moving on to more complex concepts. When students can name the parts of a sentence, they get “that feeling of control.” AI can produce questions that look like SAT questions, he said, but when it comes to AI in test preparation, the real challenge is creating questions with the depth and subject expertise that meaningful learning requires. 

I told him we’d lived this. For six or seven months, we tried AI for building questions at EdisonOS. We learned it couldn’t do what a subject expert does, and we set a policy: no AI-generated questions. Meanwhile, some competitors announced libraries of 20,000 to 50,000 questions while we sat at 4,000 to 5,000. We felt the FOMO. Then a year later, a couple of them changed course and returned to expert-written content. In our experience, writing a good SAT question takes years of understanding why the test exists.

But I don’t want to make this an anti-AI story, because the research isn’t. In one randomized trial, a well-designed AI tutor beat an active-learning class. In another, a standard chatbot left students worse off on a later unaided exam. In a big Tennessee study, Khan Academy helped students but its AI tutor, Khanmigo, added nothing on top, because most students rarely used it. And in a Stanford trial, an AI that suggests what a tutor might say next raised student mastery by 4 percentage points, and by up to 9 when the tutor was less experienced. The pattern is clear. AI-powered tutoring works best when technology helps human educators, rather than trying to replace the expertise, guidance, and connection they bring to learning. . A tutor who reaches 100 kids can maybe reach 1,000.

Then came the question of how a bite-sized game can build the stamina a long test needs. We landed on layers. First accuracy, then speed, then the full test. Games belong to the first layer, where recall lives. Jin’s app, Daily Whisk, is built for ten or fifteen minutes a day to keep terms fresh between tutoring sessions. “I don’t think that enjoyment means fun all the time,” Jin said. And Jin did something I admire: he changed his mind about shorter tests on the spot. “I think I need you as my hype man,” he laughed.

Finally, we were asked how colleges and tutors can trust what a student does alone when AI can help with every essay.

Jin told us about his first year of the PhD. Fifteen papers a week, three presentations, no time. So he fed papers to AI and let it do his thinking. A couple of months later, his own thinking had become harder. He’d sit down to write and his mind would reach for “let’s just consult AI.” He had to write the first draft of his thoughts by hand, away from the screen. “I got sucked into that trap,” he said, and he’s a 43-year-old man with pretty good self-discipline. He’s still “unwiring” the habit.

What he tells his students isn’t a warning. It’s an appeal. “It’s more important now than it ever was that we become independent thinkers.” Developing critical thinking skills in students is becoming essential as AI makes information and answers increasingly accessible. Otherwise they’ll do “the least amount of cognitive work possible,” like everyone else. He called it protecting “sovereignty of mind.”

I told him it reminded me of a calculator. You learn arithmetic on paper first. Give a first-grader a calculator in the exam and the exam loses its purpose. Same with coding: from C and Pascal to high-level frameworks to vibe coding, we keep removing the step where you have to understand what’s underneath. And patience has been shrinking for decades: from books, to movies, to YouTube, to one-minute reels. “Patience has been my word in 2026,” Jin said. It could be mine too.

The data points the same direction. The OECD found teen reading at its lowest level this century, linked with more screen time, less reading for pleasure and more hasty reading. It also found students who use AI to draft essays or summarize texts scored about 20 points lower in science, although that’s a correlation, and struggling students may simply use chatbots more. I’d rather not overclaim. But I don’t need a study to see what Jin lived.
‍

What These Conversations Reveal About AI in Education

When I replayed both conversations, three threads kept showing up.

Foundations. Whether it was Jin’s grammar terms, the UC professors, or my team’s refusal to let AI write test questions, the answer was the same: master the basics, and the rest becomes possible.

Patience. Mastery without perfection is patient: incremental targets, not quantum leaps. Stamina is patient too. And independent thinking is the first thing we lose when we’re in a hurry.

People. A tutor brings empathy, accountability and trust. A teacher who admits his own struggles with AI is more persuasive than any policy. Even Messi and Ronaldo have coaches.
‍

Three Lessons for Students, Tutors, and Parents

• If you’re a student: Aim for “fewer than five wrong,” not perfect. And do your first thinking by hand before you ask AI.
• If you’re a tutor: Let AI take the admin and the first draft of ideas, and keep the relationship for yourself. That’s what parents are really paying for.
• If you’re a parent or teacher: Have the honest conversation. Jin believes a heart-to-heart about what we lose when we hand over our thinking beats any rule we can write. We also need to rethink assessment. At EdisonOS, we interview content writers with a pen, paper and a few books, no laptops, and we see real thinking.
  ‍

The Future of AI in Education and Human-Led Tutoring

At EdisonOS, we’re using these conversations to sharpen our own direction: more tests, a high-quality question library written by experts, and AI-powered tools that make tutors more productive instead of replacing them. It's the kind of approach we believe will shape the future of education. We’ll keep talking with Jin, and we’d love to hear from you. How do you keep your own thinking sharp? What would you tell a student?

Near the end of the second call, Jin said he hadn’t realized how much fun this would be. I felt exactly the same.
‍

Sources

• Jin Oh Bae, Perfectionism and the SAT: Building mastery without demanding perfection, LinkedIn
• Retrieval Practice, Practical tips to reduce student test anxiety
• Uniqprep, Digital SAT guide for parents
• Academic Approach, UC faculty call for reinstating ACT/SAT requirements
• NBC News, US high school students lose ground in math and reading
• Educational Psychology Review, Augmentation or Erosion? Regulating cognitive offloading in generative AI-supported STEM learning
• Hechinger Report, Students didn’t get answers from Khanmigo. They didn’t want its questions, either
• Education Week, What happens when an AI assistant helps the tutor, instead of the student
• US News/Reuters, Teen reading slumps to worst this century due to surge in screen time
• Technology.org, PISA 2025: reading, screen time and AI
  ‍