Watching ChatGPT effortlessly spin out answers to any question I throw at it is a magical experience. But I’ve been to enough magic shows to mind the gap between appearance and reality. Salman Khan, apparently, hasn’t.
As the founder of the widely acclaimed Khan Academy, which originated as an online tutoring service, Khan has been an early adopter of digital tools in his efforts to improve access to education. When OpenAI offered him an early version of ChatGPT-4, he found the technology “absolutely mind-blowing” and agreed with OpenAI President Greg Brockman that large language models will provide “the biggest benefit to education we’ve had in history.” Within months, his team rolled out an AI assistant dubbed Khanmigo—a play on the Spanish word conmigo, meaning “with me”—to provide tutoring and coaching for students in all subjects.
Not only does he apparently believe all the AI hype, Khan also breathlessly amplifies it in his book Brave New Words: How AI Will Revolutionize Education (and Why That’s a Good Thing). That might not be much of a problem if Khan and the many influencers endorsing his book and ideas—Bill Gates, Sam Altman, Laurene Powell Jobs, Satya Nadella, Arne Duncan, and others—didn’t have so much social and financial capital, which they’ve invested in the success of AI. Since AI has been unleashed on all of us, we need to grapple with how to use it well, a monumental project that begins with asking the right questions about how to educate students.
Khan’s problems begin with his simplistic view of AI as an inevitable technology that demands a binary choice from us. Since it is “futile” to ask teachers or students to avoid ChatGPT, his reasoning goes, they should instead “lean into” these tools. But should they? After all, very few informed people think that even if social media of some kind is an inescapable part of our world, we must therefore embrace it in all contexts. Rather, we wage prudential, telos-oriented, data-supported debates about the appropriate use of social media within our personal lives, our families, and our schools. Just because I can’t, as an individual, choose to live in a world without OpenAI, it does not follow that I must uncritically adopt it in all aspects of my life.
When Khan does try to assess various technologies with more nuance, his claims collapse into contradictory nonsense. Early on, he affirms the reductive cliché that technology is not “good or bad, it is how you use it that matters.” And despite acknowledging the real problems caused by social media, corporate or government surveillance, and deepfakes, Khan simply asserts that the solution to technological problems is always more technology: “The countermeasure for every risk is not slowing down; it is ensuring that those favoring liberty and empowering humanity have better AI than those on the side of chaos and despotism.” How Khan would “ensure” that only the good guys have powerful AI is never made clear.
Nor does the turgid, somewhat robotic prose do much to win readers to his argument for AI-powered education. Consider this paragraph on how AI will transform history education:
Generative AI, with its ability to mix media and content, has the potential to bring history and civics lessons to life. By offering an interactive and immersive learning experience, it empowers students to delve into historical events, engage in meaningful discussions, and develop a deeper comprehension of civic principles. Its personalized explanations, responsive question prompts, and diverse perspectives stimulate critical thinking and encourage students to form their own well-informed opinions. With these types of tools, history and civics lessons transcend conventional boundaries, empowering students to connect with the past and understand the present.
Where to begin? Do human teachers not have the ability to mix media and content? What are the conventional boundaries that AI will transcend? Is there any evidence that a chatbot helps students develop deep comprehension of civic principles or historical events? Copyleaks, a service used by many teachers to identify plagiarism in students’ writing, identifies this paragraph as 100 percent AI generated, though I needed no computer to confirm the obvious. If Kahn’s writing is evidence of the kind of prose composed by “centaurs”—those who have become, in the words of business professor Ethan Mollick, “half human and half large language model”—it is hard to believe that students who rely on AI will become, in Khan’s empty phrasing, “far more skilled and efficient writers.”
Kahn confidently positions AI as the next step in an ongoing computer revolution that has improved educational access. While it is true that computers may make information accessible to more people, we have clear evidence that Internet-connected devices have thus far not improved educational outcomes, despite the massive amounts of time and money that institutions have poured into them. A 2015 report from the Organization for Economic Co-operation and Development (OECD) on “Students, Computers, and Learning,” for instance, found that the use of computers in the classroom had “no appreciable improvements in student achievement in reading, mathematics, or science” in those countries that had invested in information and communications technologies for education. The report also found that “technology is of little help in bridging the skills divide between advantaged and disadvantaged students.”
The report’s authors offer two tentative interpretations of these discouraging findings. The first “is that building deep, conceptual understanding and higher-order thinking requires intensive teacher-student interactions, and technology sometimes distracts from this valuable human engagement.” The other conjecture is that “we have not yet become good enough at the kind of pedagogies that make the most of technology.” I am inclined toward the former explanation, and clearly Kahn would favor the latter, but his book offers few details on what good AI pedagogy might entail.
Moreover, what data we do have on the benefits of Khan Academy resources and the early trials of AI teaching tools suggest these technologies have the same limitations as older digital tools. Citing the results of one evaluation of Khan Academy’s achievements, Khan boasts that “in a study of more than three hundred thousand students using standardized test scores to inform personalized practice on our platform, ‘students who engaged…during the 2021–22 school year at the recommended dosage of 30+ minutes per week exceeded growth projections by 26 percent to 38 percent, depending on grade.’” What Khan fails to mention is that of the 329,957 students in the study, only 18,009, or 5.4 percent, used the program 30+ minutes per week. Perhaps these self-motivated students would have seen similar gains from doing other homework or extra work, but the study was not designed to measure such comparisons. Any teacher or curriculum can help the top 5 percent of a class; the real challenge is engaging and helping the 95 percent of “checked out” students.
And despite Khan’s promises, AI does not seem to help those students. Khan touts early reports from schools that have rolled out Khanmigo indicating that student users report higher self-confidence. This is confirmed in one recent study of math students from the Wharton School at the University of Pennsylvania, which found that AI contributes to overconfidence: Students with access to ChatGPT while working on their homework did better on the homework and felt better about their understanding of the concepts, but, without the aid of ChatGPT, they scored an average of 17 percent lower on the math tests than students using a version of AI designed not to give them answers and students who just did the homework without any assistance. In other words, unfettered AI hurts student learning, and even customized AI tutors have a null effect.
Why do these fancy technologies fail to improve student learning? Because the real barrier to educating students is not a lack of access to information but a lack of wise, caring mentors who can motivate them to do the hard work of thinking. As computer scientist Kentaro Toyama writes in his thoughtful book Geek Heresy: Rescuing Social Change from the Cult of Technology, education “requires directed motivation. It doesn’t matter what flashy interactive graphics exist to teach this material unless a child does the hard internal work to digest it. To persevere, children need guidance and encouragement for all the hours of a school day, at least nine months of the year, sustained over twelve years…. The essence of quality children’s education continues to be caring, knowledgeable, adult attention.” Time and again in his work with Microsoft, in India and elsewhere, Toyama found that “new technology never made up for a lack of good teachers or good principals…. If anything, [existing] problems were exacerbated by the technology, which brought its own burdens.” And unlike Kahn, Toyama supports his claims with robust research confirming that schools around the world that receive laptops report no improvement in student learning.
Khan, to his credit, recognizes the need not just to make information accessible but also to engage students. He learned this firsthand when he tutored his cousins in math. If one of them didn’t show up to a lesson, Khan could call his aunt for support. But students and their parents know that Khanmigo is a computer, not a cousin or a nephew, and it is the human relationship that enables a teacher to motivate students. Nevertheless, Khan’s vision of an AI-powered classroom entails chatbots communicating with individual students and leading student breakout groups, and then reporting back to the teacher so that the teacher can be “in the loop as to what is going on” and “understand how the students are interacting with each other.” This seems like piling on a lot of technology to replicate what any good teacher is already doing without an AI “facilitator.”
Unfortunately, many leaders in higher education have bought into the promise of AI as a cure-all for educational and institutional challenges. Arizona State University is perhaps the most-notable example. Lev Gonick, ASU’s chief information officer, justifies ASU’s “embrace” of AI by claiming that these “tools are leveling the playing field.” But again, the evidence is lacking. Other institutions are taking a more cautious approach, exemplified by Boston University’s plan to “critically embrace” AI. Even those who are more skeptical, including me, can recognize that LLMs are certainly helpful for some tasks: If I give ChatGPT last semester’s course schedule, it can update all the dates for the new semester. While useful, this tool is not going to revolutionize the way I teach students to read and think and write—or help any of us with the challenge of passing on wisdom from one generation to the next.