OpenAILeah Belsky on how AI is transforming education — the OpenAI Podcast Ep. 4
CHAPTERS
- 0:22 – 1:40
Leah Belsky’s journey to OpenAI and the “education moonshot”
Leah explains how her career in global education (World Bank, Coursera) led her to OpenAI. She describes a mission-driven mandate: build AI tutoring that improves human potential and make it accessible worldwide.
- •15 years in education focused on widening access
- •OpenAI role framed as a “moonshot” to realize AI’s best educational potential
- •North Star: an effective tutor/companion available to everyone
- •Education vision tied to OpenAI’s broader mission and product goals
- 1:40 – 3:50
ChatGPT as the world’s largest learning platform—and governments leaning in
Leah argues that ChatGPT has become the biggest learning destination, driven by massive usage beyond formal schooling. She describes how ministries of education are approaching AI as national infrastructure, both for education improvement and economic competitiveness.
- •Learning is one of ChatGPT’s top use cases at massive scale
- •“OpenAI for Countries” and outreach from ministries worldwide
- •Estonia as an early example of national-level interest and readiness
- •Countries see AI as core infrastructure for schools and workforce transition
- •Goal: students graduate with AI experience, not just take AI courses
- 3:50 – 5:12
Universities: equal access, campus-wide deployment, and the trust problem
The conversation shifts to higher education adoption: campuses want to provide AI equitably and share best practices. But student hesitation emerges when school-provided tools feel surveilled, a sensitivity heightened by the COVID-era experience with monitored edtech.
- •Universities deploy AI to equalize access across socioeconomic lines
- •Institutions want collaboration on top faculty use cases
- •Students may avoid university AI unless privacy is explicit
- •COVID-era monitoring/remote learning shaped student distrust
- •Trust-building is framed as essential for adoption
- 5:12 – 6:50
From AI detectors to better assessment, policy, and classroom practice
Andrew and Leah critique early institutional reactions—especially unreliable AI detection—that damaged student-teacher trust. Leah argues the field is moving from “policing” toward clearer policies and redesigned assessments that better reflect AI’s reality.
- •AI detectors described as inaccurate and harmful to trust
- •Early response often: bans and enforcement rather than pedagogy
- •Need explicit guidance: when AI is allowed vs. not allowed
- •Push to redesign homework and assessment for an AI-enabled world
- •Signs of institutions reversing bans and adapting
- 6:50 – 9:51
Study Mode: turning ChatGPT from answer machine into a tutor
Leah introduces Study Mode as a new learning-focused experience designed to guide students to answers rather than provide them outright. It uses Socratic questioning, personalization, quizzes, and deeper follow-ups to promote understanding.
- •Study Mode emphasizes guidance, not just outputs
- •Socratic approach with tailored difficulty and context
- •Encourages follow-up questions, quizzing, and deeper exploration
- •Positioned as a first step toward ChatGPT-as-tutor
- •Designed so students don’t need expert prompting to learn well
- 9:51 – 11:35
How Study Mode was built: learning science, global experts, and “golden examples”
Leah shares the origin story of Study Mode, sparked by observations in India about tutoring costs and demand for after-school support. The team built a pedagogy-informed response schema and trained toward high-quality tutoring interactions using curated examples.
- •India trip highlighted heavy household spending on tutors
- •Product development guided by learning science and pedagogy experts
- •Creation of a tutoring “schema” for how the model should respond
- •Use of “golden examples” to tune tone, curiosity, scaffolding
- •Study Mode framed as an early iteration, not the endpoint
- 11:35 – 14:15
Where tutoring impact shows up first: confidence and out-of-class support
Leah argues AI’s earliest education impact is outside the classroom—providing the kind of adult support many learners lack. She highlights confidence-building and persistence, including stories from a student user group (ChatGPT Lab).
- •AI acts as accessible “adult support” where tutors/parents/teachers aren’t available
- •Feedback and encouragement can reduce stuckness and discouragement
- •Student story: CS learning improved after using ChatGPT as a tutor
- •Confidence and motivation framed as core learning accelerators
- •ChatGPT Lab formed to learn directly from student users
- 14:15 – 18:00
Workforce readiness: AI fluency and the return of coding as core literacy
Leah and Andrew discuss labor market shifts: AI-skilled workers are significantly more productive and highly valued by employers. They argue graduates need practical AI usage skills, and that understanding and debugging code becomes even more important as coding gets easier.
- •Workers using AI show sizable productivity gains
- •Employers increasingly prefer AI skills over years of experience
- •AI as campus infrastructure partly driven by employability goals
- •“Vibe coding” and tools lower barriers, making coding broadly useful
- •Coding literacy reframed: understand, create, and debug—even with AI help
- 18:00 – 19:30
The “brain rot” debate: when AI helps learning vs. replaces struggle
They address fears that AI weakens thinking by making learning too easy. Leah frames AI as a tool whose impact depends on use: learning requires productive struggle, but AI can also enable higher-level work when fundamentals are in place.
- •Risk: using AI as an answer machine undermines learning
- •Learning requires struggle, processing, and feedback loops
- •Analogies: calculator vs. long division; scooter vs. marathon training
- •Study Mode positioned as a design response to encourage thinking
- •Goal: expand creativity and critical thinking rather than shortcut them
- 19:30 – 21:30
A personal accessibility story: voice mode and a dyslexic learner
Leah shares how ChatGPT’s voice capabilities changed her perspective on accessibility for her dyslexic daughter. A simple conversation about current events demonstrated how AI can unlock information and independence for learners with different needs.
- •Advanced Voice Mode as an accessibility breakthrough
- •Dyslexia challenges traditional reading-based information access
- •ChatGPT conversation tailored to interests (space, robots, current events)
- •AI as a path to independence and broader world access
- •Accessibility framed as a major reason Leah joined OpenAI
- 21:30 – 25:25
Meet the students: backgrounds and first ‘aha’ moments with ChatGPT
The episode introduces two student users—Yabsera (USC, communication to business analytics) and Alaap (Berkeley, EECS). They describe early encounters with ChatGPT ranging from essay generation to playful creative prompts, and how usage matured over time.
- •Yabsera: interdisciplinary path from communication to analytics
- •Alaap: Bay Area tinkering background leading to EECS
- •Alaap’s early ‘wow’: ChatGPT wrote a full essay (which he didn’t use)
- •Yabsera’s early use: fan fiction and everyday experimentation
- •Shift from novelty to academic/coding-focused applications
- 25:25 – 29:28
How professors are adapting: harder projects, reflections, and two-track policies
Both students describe evolving classroom approaches: less emphasis on rote definition, more on application and meaning. In CS, some professors explicitly allow AI with tougher requirements and reflective write-ups, aiming to preserve learning while embracing tools.
- •Assignments shifting from definitions to applied understanding
- •More open-format evaluation to reduce rote regurgitation
- •Some courses offer AI vs. non-AI project tracks
- •AI track can be harder and requires reflection on what AI contributed
- •Students still choose non-AI paths when fundamentals feel shaky
- 29:28 – 33:20
Trying Study Mode in practice: narrowing goals, active recall, and rigor
Alaap and Yabsera compare Study Mode with regular chat, emphasizing how Study Mode interrogates goals and knowledge level before proceeding. They highlight built-in checks for retention and a more interactive, rigorous learning flow than passive long-form answers.
- •Study Mode asks clarifying questions before teaching
- •Breaks down topics step-by-step and revisits concepts to reinforce memory
- •Encourages active recall rather than passive reading
- •Yabsera: Study Mode reduces need for heavy pre-supplied context by structuring dialogue
- •Use case examples: learning AI fine-tuning; researching niche cultural topics
- 33:20 – 41:43
ChatGPT vs. social media: attention, intentionality, and deep research
The students describe pulling back from social media—especially short-form feeds—due to passive consumption and time loss. They contrast this with using ChatGPT intentionally for targeted learning, including deep research workflows and source quality control.
- •TikTok/feeds encourage convenience but can breed complacency
- •Students prefer ChatGPT for specific questions and deliberate exploration
- •Deep research seen as improving content quality vs. general web searching
- •Tactics: constrain to academic sources; cite results appropriately
- •Reframing social media as leisure and ChatGPT as purposeful tool
- 41:43 – 49:24
Cheating, misconceptions, and advice: accountability, trust, and avoiding over-reliance
They unpack how “cheating” is being redefined and why blanket assumptions miss nuance. Advice focuses on using AI to deepen understanding and productivity without outsourcing fundamentals, while schools adapt policies and assessment to emphasize learning.
- •Misconception: all students will use AI only to cheat
- •Tension: AI output may surpass human work, changing what ‘original’ means
- •Advice to students: don’t use AI as a crutch—use it to understand and iterate
- •Risk: arriving at tests/interviews without true concept mastery
- •Need education systems that reward growth, reflection, and real understanding
- 49:24 – 59:38
The future of learning with AI: hybrid education, mentorship, and risks of centralizing ‘truth’
The discussion looks forward to AI-delivered instruction paired with human mentorship, ethics, and social learning. They also surface concerns: over-skipping traditional learning, and the danger of centralized knowledge creating echo chambers—mirroring social media dynamics.
- •Prediction: AI will deliver more personalized, multimodal instruction
- •Humans remain essential for mentorship, context, and ethical grounding
- •Future classrooms may emphasize application, judgment, and impact on people
- •Risks: bypassing fundamentals; dependence that’s hard to unwind
- •Risks: AI-driven echo chambers and centralized ‘truth’—need critical thinking practices
