Building AI Teachers in 2028
All You Need Is a Foundation Model
The purpose of this thought experiment is to provoke discussion about the future of AI in education, a topic I began discussing in my previous blog post.
Here's a techno-optimistic scenario focusing on the opportunities, not the challenges and risks.
Imagine it's 2028. AI has progressed like Sam Altman prophesied (check out this podcast.) The latest bombshell from OpenAI, GPT-4o, hints at this future. In other words, assume that the latest AI progress S-curve based on transformer/LLM and diffusion architectures (perhaps with things like JEPA mixed in) continues for a few more years.
You are tasked with building AI school teachers for Kenya, a country facing a severe shortage of human teachers. These AI tutors will work alongside local educators and report to a human headmaster.
You'll use a foundation model like GPT-6 or Claude-5 (let's call it "AI") and "train" your teacher. Start by asking the AI to research the best practices for training elementary school teachers:
"You are an assistant professor at an Ivy League university's School of Education. You have a Ph.D. in Pedagogy. You also have an MBA, focusing on public education. Research best practices for training elementary school teachers. Include modern methodologies like game-based learning and project-based learning. Write an outline for a research paper."
The AI might respond with something like this (I got it from ChatGPT using the prompt above):
Best Practices for Training Elementary School Teachers: Modern Learning Methodologies and Approaches
Introduction
Chapter 1: Overview of Teacher Training in Elementary Education
Historical context of teacher training
Current trends and challenges in teacher training
Importance of effective teacher training for elementary education
Chapter 2: Modern Learning Methodologies
Introduction to modern learning methodologies
Game-Based Learning
Project-Based Learning
Mastery Learning
Flipped Classroom
Blended Learning
Inquiry-Based Learning
Differentiated Instruction
Collaborative Learning
Personalized Learning
Chapter 3: Integrating Modern Learning Methodologies into Teacher Training Programs
Framework for integration
Curriculum design and development
Pedagogical strategies
Technology and tools for modern learning
Professional development and continuous learning
Chapter 4: Case Studies and Best Practices
Chapter 5: Evaluating the Effectiveness of Teacher Training Programs
Assessment methods and tools
Impact on teacher performance and student outcomes
Continuous improvement and feedback mechanisms
Chapter 6: Challenges and Future Directions
Conclusion
References
Appendices
You would review the outline, suggest changes, and have the AI produce the research paper. After several iterations and feedback from colleagues, you are satisfied with the result.
Next, you'd ask the AI to design a curriculum for a four-year Pedagogy major, concentrating on "Elementary School Teacher". The process would be similar, with review and feedback loops.
Finally, you're ready to "build" that AI teacher:
"You are an elementary school teacher working in Kenya. You have a degree in Pedagogy and work experience. As a reminder, here is the curriculum for your kids. You teach students 1:1 and in small groups. Your students are 4-10 years old. They speak Swahili, Kikuyu, Luhya, or Kalenjin. When you plan group activities, you group kids by age and language. You are a big fan of Sal Khan, the founder of the Khan Academy, and like to emulate his style and mannerisms.
Here is Akili, one of your students. She is 5yo, her mother tongue is Swahili. Her goal for this 2028/2029 school year is to
Master Swahili literacy and basic numeracy
Learn the basics of English as a Foreign Language (EFL) up to A2 level
Complete the Understanding Nature course.
Start by develop a 15-minute gamified placement test for her. Send the report to her parents and the human guidance counselor living in the provincial center.
Design a personalized learning plan for Akili in Reading, EFL, Math, and Understanding Nature, considering her 4 x 45-minute daily sessions, 5 days a week.
Create lesson plans for Akili for today, based on her schedule.”
The AI teacher is ready to teach Akili! They would deliver the lessons, track Akili's progress, and send reports to parents and the headmaster.
In essence, this thought experiment explores how we might "build" AI tutors and teaching assistants “simply” by training a foundation model in pedagogy. What do you think - is this plausible?
Of course, today's generation of foundation AI models isn't there yet. They need substantial advancements in their planning and reasoning abilities, better multimodal capabilities, an understanding of kids' accented speech, and much much more. They need to get smarter. And they need the safety guardrails - we would need to trust them not to do something harmful or inappropriate.
Still, I think this is a plausible scenario for Education and EdTech. If so, here are some questions to consider as we move forward:
How can we encourage AI developers to prioritize support for educational applications?
How can we ensure these AI tutors are effective, safe, and complement human teachers?
What's the best way for human teachers to use these AI tutors and/or teaching assistants?
Thoughts?