Why There Are No AI Shortcuts to a Good Education
Will AI be the technology that finally revolutionizes education?
With the rise of ChatGPT and other powerful generative AI systems, it’s worth re-asking the question, Will this new generation of digital technology improve education? I place my bets on “no” – not much, and not in a way that really matters.
I first need to explain what I mean by improving education “in a way that really matters.” Put simply, it means getting closer to the goal of high-quality education for everyone. And, that in turn means a higher-quality education for the children of lower-income and/or less-educated families of the world. The children of richer, well-educated families are not who we’re worried about – with or without technology, they’re getting a good education already and will continue to do so.[i]
Next, so as to be able to discuss concretely what current AI might do for learning, I refer to Khanmigo, an AI-powered tutor that Khan Academy has built on ChatGPT, an overview of which appears in a TED talk that founder Salman Khan gave in April 2023.[ii] It serves as a good example for several reasons. First, Khan Academy is very well-funded,[iii] so it has the ability to build state-of-the-art educational technology quickly. The talk occurred just a few months after the public release of ChatGPT, and Khan shows off a working Khanmigo demo live. Second, Khan Academy has so far appeared to be sincere about its mission to make “education free and accessible for all.” For the most part, its offerings appear to have been sincerely developed and made available free. We can probably take its idealistic educational mission at face value.[iv] Third, Khan Academy is well-known and arguably among the world’s most successful educational technology efforts. It has the potential to make big educational impact. In short, Khanmigo – and Khan’s description of it – represent what is among the best that AI-based educational technology might offer.
So, what exactly does Khanmigo do? In essence, it’s a versatile, well-informed tutoring chatbot. In his TED talk, Khan walks the audience through Khanmigo doing the following through text-based interaction:
Helping a (hypothetical) student struggling with arithmetic with a response tailored to their likely cognitive misunderstanding;
Helping someone with coding for a simple animation, again with tailored responses;
Providing career advice in the manner of a high-school guidance counselor;
Having a conversation, pretending to be the fictional Jay Gatsby;
Debating with a student about whether student debt should be canceled;
Playing a collaborative story-writing game.
Khan also mentioned that they were working on a prototype in which Khanmigo would ask reading comprehension questions based on a piece of text and have an ongoing conversation about it. For teachers, Khanmigo can provide assistance for developing lesson plans and providing feedback to students.
As far as digital tutors go, Khanmigo is impressive, both for what it reveals about ChatGPT’s strengths, and also for what the Khan Academy has done on top to customize it for education. It’s what generations of AI scientists and educational technologists have wanted to build, and had largely failed at until 2023.
But unfortunately, Khanmigo still retains the central flaw of most educational technology to-date, and as a result, I’d wager it won’t move the needle on education. (Nor will any of its virtual rivals-to-be, hundreds of which are certain to flood the market in the coming years.) The flaw is to treat the availability of knowledge, or even the highly personalized presentation of it, as the key challenge. Khan himself so thoroughly represents this misguided view throughout the talk, that I doubt he’s even aware of it. He opens the talk with a study from the 1980s that shows that students learn much better when they receive 1-to-1 tutoring rather than group instruction (i.e., regular classes). Then, he suggests that what Khanmigo does is to provide that kind of tutoring, and for the rest of the talk, he emphasizes how tailored its interactions are to the intellectual need of the student user.
But, as I’ve written elsewhere,[v] the real crux of a good education is not informational in nature. It’s motivational. For students, what makes learning difficult is that it is the mental equivalent of push-ups, sit-ups, and long-distance runs of increasing intensity, day after day for years. That takes effort, and the effort is painful. For teachers, what makes teaching difficult is generating the required motivation. Good teachers do it through an ever-evolving bag of tricks – inspiring, cajoling, approving, withholding, joking, and scolding as needed. Even mediocre teachers with the right intention can push most students along if they just sit with them. So, the primary value of 1-to-1 tutoring isn’t the fact of personalized intellectual instruction – it’s the watchful eye of a vigilant adult, under which a student cannot escape doing those push-ups. Of course, knowledge and tailored instruction are important – but that’s just the baseline, readily available in a decent textbook. Putting the textbook online (the Internet), having it presented by star teachers in videos (MOOCs), making it interactive (apps), gamifying it (more apps),[vi] or even making it hyper-customized and intelligent (AI!) doesn’t fundamentally increase student motivation, at least not for the long term. The motivation isn’t in the knowledge or in the technology.
To put it another way, the essence of real education is motivation, and motivation comes from social forces. Even with today’s AI, those social forces aren’t yet present. One day, we may have humanoid robots that students engage with as human beings – and when that day comes, some of these caveats will stop applying. (Though, even in that future world, guess who will have access to the most motivational robots? Not poor families.) And, I’ll concede that there are some unknowns with today’s unembodied AI. It’s possible, for example, that some subset of students will engage with well-crafted online bots as if they were human, in an educational twist of the movie, Her. That might be enough “social” to generate motivation.
But, I tend to doubt it, at least for the vast majority. Among other things, I’m not sure too many of us will be comfortable as parents having our children engage with an AI as if it were human. More than anything, though, even young children are quick to recognize when they’re being fed digital broccoli that’s supposedly good for them; they’re not fooled by it.[vii] Just how long will an otherwise unmotivated student sit chatting with Khanmigo so that they can craft creative stories together? Not much longer than they will play with software called “Math Fun with Elsa.” In the end, a fantastic AI tutor is still a computer – one that any child can walk away from, so as to go play the latest Mario Brothers video game.
Notes
[i] To be sure, the children of richer, well-educated families could also benefit from improvements in education, but I’m not emphasizing that here for two reasons. First and again, “improving education” at a societal level really is about improving it for the people who don’t otherwise get a good one; helping the educated rich get educationally richer only exacerbates inequality and other downstream problems. Second, given the high level of education that the well-off and well-educated already ensure for their children, the marginal benefit of AI isn’t likely to be great, even for them. There’s a limit to how quickly human beings can learn, short of Matrix-esque kung-fu training machines. Also, I should qualify that when I say “rich and well-educated families give their children a good education,” I mean it on average. There are plenty of rich, well-educated parents whose kids don’t learn; and plenty of kids of poor, less-educated parents who learn a lot – but they’re in the minority. At national or global scale, parental wealth and education is strongly correlated with children’s education.
[ii] Here’s the talk on YouTube:
[iii] In 2021, Khan Academy tax filings show net assets of over $100 million and revenue of $59 million.
[iv] There’s no shortage of bad educational technology accompanied by extravagant claims and developed by hucksters wanting to turn a quick buck, but Khan Academy doesn’t seem to be that. Salman Khan receives a cushy salary as CEO of Khan Academy (close to $900,000 in 2021), but so far, he seems to be sincere in his educational mission, and has resisted any temptation to become mega-rich through Khan Academy.
[v] For example, here: https://edutechdebate.org/ict-in-schools/there-are-no-technology-shortcuts-to-good-education/ . Incidentally, the arguments I make there from back in 2011 still apply in their entirety to AI-based edutech today. Chapter 1 of my book, Geek Heresy, presents another take on the failures of educational technology.
[vi] Gamifying is an attempt to motivate, and it works to some extent. But, it’s limited both in scope and who it helps – in terms of scope, I haven’t seen gamification motivate anything too complex, e.g., critical thinking. As for who it helps, it’s those who already have some self-motivation that benefit by being nudged to learn a bit more. Exactly the people who don’t otherwise get/have a good education don’t benefit much from gamification.
[vii] …unless they have some existing predisposition to wanting to learn that particular material, but again, those people don’t need the digital bells and whistles.