I spend a fair amount of my working week deciding which AI model does a given task and which one just makes it look done. So when Richard Clark writes that media are “mere vehicles that deliver instruction but do not influence student achievement any more than the truck that delivers our groceries causes changes in our nutrition” (Clark, 1983, p. 445, as cited in Clark, 1994, p. 22), I recognize the argument from the inside. Neither, Clark says, does the medium.

Exhibit A. Khan Academy Pakistan’s launch materials for Khanmigo have co-founder Osman Rashid saying the organization aims “to use Khanmigo to transform the educational landscape of Pakistan,” with Sal Khan framing it as bringing “the power of AI to tackle the unique challenges faced by students and teachers” (Khan Academy Pakistan, n.d.). Rashid’s own verb is “aim.”

Exhibit B. A Research and Markets industry report, distributed through GlobeNewswire, claims generative AI is “enhancing global edtech with . . . automated grading, and intelligent content creation, offering personalized and efficient educational pathways” (Research and Markets, 2026). This one doesn’t mention a teaching practice or a curriculum decision. Teachers and students appear only as the market the technology is arriving for.

Clark’s response to Exhibit A starts with his replaceability test: could the same learning gains come from a cheaper, non-AI intervention, a trained volunteer tutor, a well-sequenced workbook? If so, the medium isn’t the cause. But Clark’s own framework has an escape hatch here, and it’s worth granting it in full: he explicitly allows that media can change the cost and reach of instruction without changing what’s learned. A free AI tutor reaching students who’d otherwise have no tutor at all is exactly the case Clark’s theory lets you praise. That’s a real point in Khanmigo’s favour. Rashid’s claim was about learning, though, and Clark’s framework only rescues claims about reach.

Exhibit B splits cleanly on Clark’s own terms. Automated grading is pure delivery efficiency, the truck getting faster. Intelligent content creation is a different animal, and it’s where Clark’s theory starts arguing with itself.

That’s Kozma’s opening. His 1994 reply reframed the question entirely: instead of asking whether a medium’s effect can be replaced by another medium, ask whether a learner is constructing knowledge through a medium’s specific symbolic and processing capabilities, the ways it represents information and lets a learner manipulate it. On that view, media and method were never cleanly separable to begin with. Generative content creation might genuinely be a Kozma-shaped capability, not because an AI wrote it, but because it can construct a representation tailored to what a specific learner just asked, in real time, in a way a printed workbook structurally cannot. Whether that capability is actually causing anything depends entirely on whether the pedagogy uses it, or just announces it and moves on.

Here’s an observation neither press release cites, because neither vendor would choose to. Kristen DiCerbo, Khan Academy’s own Chief Learning Officer, told Chalkbeat’s Matt Barnum that after watching Khanmigo in a US high school classroom, “so far I am not seeing the revolution in education” (Barnum, 2026). Her explanation: “students aren’t great at asking questions well.” A Clark reading of that sentence is available, and it’s not weak: if the missing ingredient is instruction in how to ask a good question, the active ingredient was always the method, and the truck argument holds again. But I’d read it the other way. Kozma’s capability, the AI responding to what a student actually says, only produces anything if the student says something worth responding to. The capability existed. Nobody had designed instruction to use it. That’s not a software problem. It’s a Kozma-shaped gap in a Clark-shaped world: the field built the capability and skipped the part where you design for it.

Thirty-two years on, the field is replaying the Clark and Kozma argument with better hardware, and Khan Academy’s own Chief Learning Officer just told a reporter which half of it her organization forgot to build for.

References

Barnum, M. (2026, April 9). Why Sal Khan is rethinking how AI will change schools. Chalkbeat. https://www.chalkbeat.org/2026/04/09/sal-khan-reflects-on-ai-in-schools-and-khanmigo/

Clark, R. E. (1994). Media will never influence learning. Educational Technology Research and Development, 42(2), 21–29. https://doi.org/10.1007/BF02299088

Khan Academy Pakistan. (n.d.). Khan Academy Pakistan launched, brings AI-powered tutor ‘Khanmigo’ to country. https://khanacademypakistan.org/_us_news/khan-academy-pakistan-launched-brings-ai-powered-tutor-khanmigo-to-country/

Kozma, R. B. (1994). Will media influence learning? Reframing the debate. Educational Technology Research and Development, 42(2), 7–19. https://doi.org/10.1007/BF02299087

Research and Markets. (2026, July 8). How generative AI is shaping the future of digital learning platforms in the edtech industry [Press release]. GlobeNewswire. https://www.globenewswire.com/news-release/2026/07/08/3324182/28124/en/how-generative-ai-is-shaping-the-future-of-digital-learning-platforms-in-the-edtech-industry.html