Sunday Pick: Can AI really teach your kids? with Leah Belsky | from WorkLife with Molly Graham

Molly Graham introduces Leah Belsky as someone who has spent nearly 20 years at the intersection of education and technology, including leading work at Coursera and now serving as OpenAI’s head of education. Leah says she did not intentionally enter ed tech. At Brown, she was an activist and helped conceive an online investment club during the corporate reform movement. An older student encouraged her and her friends to create a company instead, leading them to find an engineer and creatives in the computer science lab and pitch major venture capitalists along Route 95 near Boston. That experience convinced Leah that young people could create things and approach powerful institutions. Her enduring motivation, she says, is not technology for its own sake but unlocking people’s confidence, connections, skills, and capacity to learn. After college she worked at the World Bank, then joined Coursera, where she saw technology as a way to give people anywhere access to learning. Coursera made university courses available to people who otherwise would have needed admission to an expensive college. A high-school student in India or a community-college student in the United States could, for example, learn Python from professors at Michigan or another leading university.

Leah says Coursera’s ambition was to help people gain workforce skills without first paying for an expensive degree, thereby opening access to jobs and opportunity. Some of that happened, but the model ran into two major problems. First, very few learners completed courses: many explored the platform enthusiastically but did not continue very far. Second, employers still used formal degrees as a major filter, so completing online coursework did not remove the institutional barriers Coursera hoped to bypass. What Leah found valuable was a different kind of outcome. Surveys showed that even learners who completed only part of a course often gained confidence that they could learn and improve their lives. Someone might come away believing, “I can actually learn,” and imagining that acquiring IT skills could lead to a job at Google or elsewhere. Leah tells Molly that Coursera demonstrated the power of allowing students to chart their own learning journeys while also trying to change universities’ business models. It also taught her how difficult formal educational change is. Universities are expected to produce research, generate and disseminate knowledge, and teach students, all while responding to numerous competing incentives. Leah compares this with a startup: a founder would be told to choose one primary goal and pursue product-market fit, whereas universities are structurally designed to pursue several goals at once.

Molly contrasts AI with earlier promises involving online courses, MOOCs, smartphones, video, and the forced shift to remote learning during COVID. Leah agrees that AI is substantially different. Earlier educational technologies generally entered institutions from the top down: schools or universities evaluated a technology, created policies, trained teachers, and deployed it gradually, sometimes over years. AI, by contrast, spread from the grassroots. Students were among the earliest adopters, while education systems are still deciding how and whether to respond. Leah says the combination of student-led adoption and unusually rapid growth is what makes this moment distinctive. At Coursera, she saw ChatGPT become part of executive work after CEO Jeff required executives to use ChatGPT before messaging him and asked that their messages be smarter than what ChatGPT could produce. ChatGPT even appeared on a computer during executive-team meetings as a kind of thought partner. After Leah joined OpenAI, she realized the scale was much larger than Coursera’s. At the time she joined, roughly 40% of users were under 24 and about 20% of messages concerned learning or education. She says ChatGPT had quickly become one of the world’s largest learning platforms, apart from services such as YouTube or Google Search. Molly notes that the numbers made the youth-driven learning use case far bigger than many people realize.

Leah’s clearest early example came from her daughter, who is dyslexic. Leah had watched her daughter’s brother read the newspaper and wondered how her daughter would access the world’s knowledge if reading remained difficult. While the family was traveling in Iceland, shortly after ChatGPT’s voice mode launched, Leah gave her daughter the system and suggested that she use it to learn about the world. Her daughter politely asked ChatGPT to help her learn about current events, then said she was interested in astronauts and space. ChatGPT asked what she wanted to learn about and began explaining the Mars rover and the history of the space program. Leah describes the moment as a complete unlock: information could be delivered in a way tailored to her daughter’s interests and needs, rather than requiring her to access it through the same reading pathway as everyone else. Molly observes that personalized learning and AI tutors have been goals of education technology for years, without delivering as much as hoped. Leah agrees that AI has the potential to become a personalized tutor, but says the genuine version is farther away than she expected when she joined OpenAI. Benjamin Bloom’s research showed a major learning advantage for students with one-on-one tutors compared with classroom instruction, but making that experience available to everyone remains difficult.

Leah distinguishes between giving personalized information and actually tutoring. Current AI models are good at answering specific questions and presenting information in an individualized way. A real tutor, however, understands a learner’s goals, knows what the learner already understands, identifies what they need next, and interacts in a way that creates productive friction. The tutor asks the student to think, reflect, remember, apply an idea in another context, and work through difficulty rather than simply receiving an answer. Leah says AI has not yet figured out how to improve cognition, memory, reflection, motivation, and accountability in the way an effective human tutor does. This is also why she believes schools will remain important even if information becomes universally available. Schools provide a structure in which students are motivated, held accountable, and brought into contact with other people. ChatGPT can provide information, but it is not yet good at motivating a student, understanding their emotional context, or creating a learning community. She expects AI tutoring to become much more effective, but does not believe it will replace human teachers or tutors. Human teachers bring students together, interpret their circumstances, and inspire them to learn. Molly connects this view to a broader pattern in AI discussions: the technology may change work and education substantially, but people working closest to it are not describing a future in which humans simply disappear.

Molly raises the fear that AI makes people dumber by removing the effort that develops understanding, comparing it to calculators reducing people’s ability to do arithmetic mentally. Leah says ChatGPT is a tool that can be used constructively or in ways that diminish learning. She compares it to her daughter learning long division. If Leah had handed over a calculator while her daughter was still trying to learn the procedure, the child would probably not have learned long division. Once the skill is established, however, a calculator can support higher-order mathematics. The same principle applies to AI. If a student photographs homework, asks ChatGPT for help, and copies the response, the experience may feel like learning even though it has removed the struggle, reflection, and effort needed to learn. Leah argues that schools should not simply avoid AI. Because the technology will remain available, educators should teach students what learning is and how to use AI without bypassing the cognitive work. She says students themselves are among the people most worried about cheating, “brain rot,” and the effect of AI on cognition because they want to learn and succeed. OpenAI has developed Study Mode with hundreds of teachers and learning scientists. Rather than immediately providing answers, the mode prompts students to explain what they know, reflect, and go deeper. Leah also says schools need explicit rules: her son’s school tells students when technology is allowed and when it is not, particularly for writing. Leah reinforces that when work feels hard, the struggle is often where learning is happening.

Leah says her own parenting changed after two years of experimenting with AI. When she joined OpenAI, she focused on helping her son use ChatGPT as a coach or tutor for homework. More recently, she has become more interested in helping him use AI to create things. Her children wanted to start a business selling 3D-printed fidgets at farmers’ markets and online, so Leah helped her son use AI coding tools to build a website. The experience gave him a strong sense of agency: he began diving deeper because he felt able to create and build something himself. She also sent him to an AI camp. Leah believes this kind of agency should become a central educational use of AI. Students could identify a problem, build a prototype, and repeat the process, rather than treating entrepreneurship, coding, and making as extracurricular activities available only to students with special resources. She gives examples from students already doing this. One is using language models to examine existing drugs and identify medicines not currently used for mental-health treatment that might be relevant to bipolar disorder. Another built an accessible version of Wordle for a blind friend and then started a company making educational games for students with disabilities. A Penn student and a group of nontechnical peers used AI to design a system for redirecting unused cafeteria food to food banks. These examples challenge the idea that students mainly use AI to cheat. Leah says students who learn to use it productively are identifying problems and building solutions without waiting for permission, engineers, venture capitalists, or established networks.

Leah says AI should help teachers as well as students. Teachers spend substantial time grading, compiling assessments, preparing materials, and handling administration. When teachers use AI actively, Leah says they can save roughly six or seven hours per week and spend more time with students. For students, she sees a clear role for AI in learning how to build and create, but not as an invitation to keep ChatGPT open all day. Any classroom deployment should be intentional and designed for a specific purpose. Molly asks what young people should focus on if they are anxious about future jobs. Leah reframes the question: in a world where intelligence is broadly accessible, what should humans be able to do? She emphasizes emotional intelligence, leadership, judgment, perception, relationship-building, and the ability to understand other people. She also wants children to have the confidence to identify a problem, move from problem to prototype, and either improve an existing organization or build something of their own. Neither Leah nor Molly claims to know what future jobs will look like. Leah’s final position is that AI is a powerful tool and accelerant, not the future of learning replacing schools. The unresolved challenge is that society is simultaneously figuring out how students should use AI and what schools should accomplish when information and intelligence are widely accessible. Molly ends by saying that access to information is not itself learning: real learning often includes the difficulty of struggling until an idea becomes embedded in the mind.