How do the schools of today fall short?
The schools of today crush creativity and curiosity. We all know what they are like because their basic structure hasn’t changed since the industrial revolution. In middle school, high school, and college classrooms there is typically 1 instructor and N students. The majority of the time, the instructor lectures and the students are expected to sit still, listen attentively, and take notes. At the end of the lecture, the instructor reminds the students about their mind-numbing homework assignment without giving any explanation for why the homework is relevant to the real world. After a series of lectures, one test is given to all of the students and the students are assigned grades based on their performance on this test and a few homework assignments. Sadly, this is still how the majority of classrooms operate today. This learning format is a one-size-fits-all approach and completely fails to recognize the individuality of the students—not everyone learns at the same speed or in the same way.
That said, we shouldn’t blame the presidents and principals of these schools too much for the terrible state we find ourselves in. This is because improving learning outcomes actually seems very difficult; in the last 50 years there’s basically only one educational intervention at scale that improves outcomes1, and it is one-to-one tutoring. However, high quality tutors are too expensive for most students to afford.
How will AI improve this situation?
AI assistants will be the perfect tutors
Anyone who has spent a bit of time using ChatGPT will know that AI assistants can already be useful in many cases. But ChatGPT is just the beginning; in the future, AI assistants will be the perfect tutors.
At the moment, ChatGPT hallucinates fairly often and recalls information poorly. However, in the near future large language models (LLMs) like ChatGPT will be hooked up to the web and be able to recall and synthesize information much more accurately. DeepMind’s RETRO shows that improving LLM’s retrieval abilities is already possible2.
The LLM based AI assistants of today are pretty good at interpreting what people say. However, they must be queried using text. In the future, some AI assistants may start to use the cameras and microphones built into the phones/tablets/laptops/VR headsets that house them in order to better interpret the states of mind of their students. This could be particularly helpful when a student is not totally sure how to articulate themselves using words.
In addition to being able to query the web and better interpret the user’s inputs, future AI tutors will be able to adapt their output to fit the needs of their students. They will do this by asking clarifying questions (ChatGPT already does this sometimes) and using context clues (such as the writing/speaking level of the student and the types of assumptions the student makes) to determine where students are at in terms of their knowledge of the subject being taught; they will then give the students the content that is best suited for them. They will even be able to become the person the student wants to talk to most. Already, apps such as Historical Figures3 allow you to have a conversation with a virtual version of a number of historical figures such as Benjamin Franklin and Plato, to name a couple.
How will AI tutors change the classroom?
As AI tutors make their way into classrooms, the default teaching style will improve dramatically in ways that many of the greatest educational visionaries and thinkers could only dream about. Here are some changes that will happen: the default teaching style will change from lecturing to the Socratic method; the interactivity and self directed learning that Maria Montesorri wished for will be realized; creativity and curiosity will be incentivized because of the question/answer learning format, just as Ken Robinson desired; students will be able to use the Feynman Technique of learning when interacting with their AI tutors; high quality teaching will become accessible to all young children regardless of their backgrounds; and teachers/professors will change from dictators to facilitators.
What will be worth learning in this world filled with highly intelligent AI?
Discernment and wisdom will become very important. Yann Lecun, one of the founding fathers of modern deep learning and an ACM Turing Award Laureate, says that in a world where AI assistants are smarter than humans in many ways, the ability to know who to listen to will become highly valuable. Just like CEOs have to decide which of their (often intellectually superior) advisers to listen to, students will have to decide which AI assistants to listen to. This sounds like discernment or wisdom. Because of this, I predict that philosophy and the humanities in general will make a comeback in terms of usefulness and popularity. (As a side note: if wisdom is going to become more important in this future, a natural follow up question might be: Can one become wiser? Interestingly, there is evidence that wisdom can be taught4.)
In addition to wisdom, idea generation and integrative thinking will become more valuable. In a world where everyone has AI assistants that can carry out the mundane legwork needed to bring something new into existence, idea generation will become more important. Engineers and other roles that specialize in building things won’t disappear, but they will not be needed as often. In general, the time from idea to market will be cut down dramatically.
Finally, the ability to conduct oneself in “the real world” will still be valuable. Robots are still not dextrous/multimodal/cheap/general enough to replace humans in the majority of cases. Therefore, soft skills and trade jobs will still be worth learning.
Conclusion
On the whole, I am very optimistic about the impact that AI will have on education. It will allow for much more individualized learning, and creativity and curiosity to flourish in the classroom.
Interestingly, smaller class sizes do not lead to better learning outcomes than larger class sizes. I didn’t know this before writing this article. Here’s the review: https://onlinelibrary.wiley.com/doi/full/10.4073/csr.2018.10.
Another surprising finding is that having students tutor each other is actually just as effective as having experts tutor students. If this is true, why don’t we just have students tutor each other more? https://link.springer.com/article/10.1007/BF01383938.
https://www.deepmind.com/publications/improving-language-models-by-retrieving-from-trillions-of-tokens
https://apps.apple.com/us/app/historical-figures-chat/id6444197650