Punya Mishra's essay “Why Sal Khan't: On Learning by Making but Teaching by Telling” lands one sharp question: Sal Khan learns by researching, drawing, questioning specialists, making connections, and producing a video—so why should the learner on the other side merely receive the finished explanation?
The question deserved the lively Hacker News debate it received. The best replies also caught the essay's weakest move: Khan Academy has never been only a pile of videos. It includes exercises, quizzes, tests, mastery progression, teacher dashboards, and a flipped-classroom idea intended to free class time for practice and feedback.
Both points can be true. Explanation is valuable scaffolding. It is not, by itself, durable learning. The design problem for any AI tutor—including Melo, explainx.ai's learning copilot—is how to turn a helpful answer into a learner-owned act of inquiry, construction, expression, and revision.
TL;DR
| Question | Direct answer |
|---|---|
| Is Mishra's central criticism fair? | Yes: receiving someone else's polished understanding is not the same as constructing your own. |
| Is “Khan Academy = watch videos” fair? | No: the platform also has substantial practice, mastery, assessment, and teacher-support systems. |
| What did Khanmigo's adoption reveal? | A tutor waiting in a sidebar cannot manufacture curiosity, metacognition, or a reason to persist. |
| What should an AI tutor optimize for? | Learner output: questions, attempts, artifacts, explanations, decisions, feedback, and revision. |
| Where does Melo fit? | It can scaffold, quiz, create practice, and ask for an explanation back; the learner still needs a meaningful destination. |
| What is the practical rule? | Use AI to shorten the distance to a good attempt, not to remove the attempt. |
The original critique is useful—but too uncharitable
Mishra contrasts Khan's own learning process with the learner's experience. Khan first builds a scaffold, follows questions textbooks leave unresolved, draws representations, tests analogies against specialists, and finally makes a public artifact. The student receives the residue of that work: a lucid explanation.
That asymmetry is real. A worked explanation gives the learner the destination without reproducing the cognitive journey that made it meaningful to its creator.
But “watch a video” is still an incomplete description of Khan Academy. The HN commenters are right to steelman it:
- A clear, free explanation can establish the vocabulary and mental model needed before practice begins.
- Self-paced video can be paused, rewound, accelerated, or revisited without the social pressure of holding up a classroom.
- Exercises, quizzes, tests, and mastery levels ask learners to do more than watch.
- The flipped-classroom model was explicitly meant to move lecture out of shared class time so teachers could use that time for problems, projects, and human feedback.
- A consistent explanation sets a quality floor for learners whose local instruction is unavailable or incomprehensible.
Those are not small wins. Several HN commenters described Khan's videos as the first explanation that made calculus or another difficult subject feel beautiful and possible. Inspiration and scaffolding matter because a learner cannot practice a concept they cannot yet see.
The more precise criticism is this: a platform can provide every ingredient for active learning and still leave the learner with no reason to combine them. Sal Khan described that engagement problem to Chalkbeat: for many students, Khanmigo was a “non-event” because they did not use it much. A capable tutor sitting in the back of the room is still waiting for the student to raise a hand.
Explanation is a scaffold, not the finish line
An explanation can do three valuable jobs:
- Compress prior work. The learner does not need to rediscover calculus from first principles.
- Provide a representation. A timeline, analogy, diagram, or worked example makes an abstract structure inspectable.
- Reduce the cost of beginning. A learner who was stuck can make the first informed attempt.
It cannot prove that the learner can retrieve, transfer, or use the idea. That requires output from the learner.
This is the same distinction visible in recent education research. Our review of homework gains and exam-score losses with unrestricted AI found that producing better work with AI is not the same as being able to reproduce the reasoning without it. By contrast, the Dartmouth Phosphor study centered constructed responses and rubric-based feedback rather than open-ended chat.
The durable loop looks less like “ask → receive” and more like this:
| Stage | Learner action | Tutor's job |
|---|---|---|
| Purpose | Choose a destination worth reaching | Help scope it without inventing the stakes |
| Inquiry | Ask, compare, notice a gap | Supply sources, examples, and better questions |
| Attempt | Solve, build, classify, predict, or decide | Withhold the finished answer long enough for effort |
| Feedback | Expose the work to a rubric or another mind | Identify the smallest consequential gap |
| Revision | Change the artifact or explanation | Check whether the change fixed the gap |
| Expression | Defend or share the result | Challenge assumptions and ask for evidence |
If the learner never reaches the attempt, the AI has delivered content but has not yet designed learning.
Dewey's four impulses as an AI-tutor design test
Mishra uses four impulses associated with John Dewey: inquire, construct, express, and communicate. They are more useful as a product checklist than as abstract philosophy.
| Impulse | What it looks like from the learner | Failure mode in a chatbot | Better design requirement |
|---|---|---|---|
| Inquire | Notices a gap, asks why, compares explanations | Learner types a vague request and receives a complete answer | Help the learner sharpen the question and inspect sources |
| Construct | Makes a model, solution, diagram, plan, or artifact | AI constructs the whole answer while the learner watches | Require a learner attempt before showing a full solution |
| Express | Explains the idea in their own words | Learner recognizes fluent prose and mistakes recognition for recall | Ask for explanation back, prediction, or defense |
| Communicate | Tests the work against another mind or audience | Conversation has no real stakes outside the chat | Create a shareable output and invite human feedback |
Notice that “communicate” is not identical to “the chatbot replied.” A responsive model can simulate dialogue, but it cannot automatically create a classmate who depends on your explanation, a colleague who will implement your plan, or a user who will reject a confusing result. Audience creates stakes; stakes create reasons to revise.
That is why social learning for AI needs both dialogic tools and human contexts. An AI can make solitary practice more responsive. It cannot make every solitary exercise socially meaningful.
How this maps to explainx.ai Learning and Melo
explainx.ai's vision says learners do not want to be lectured at; they want to build things. Turning that sentence into product behavior means giving each surface a different job instead of pretending one chat box can do everything.
| Surface | Role in the learning loop | What it does not guarantee |
|---|---|---|
/pathways | Provides sequence and a destination across related concepts | That the learner will attempt a project or persist through difficulty |
/dashboard/learn | Melo can scaffold with Teach or ELI5, test retrieval with Quiz, and prompt self-explanation with Explain Back | That a generated explanation or quiz is factually perfect, or that the learner cares |
/practice | Provides no-login, hands-on tools for manipulating AI concepts instead of only reading about them | Transfer to an unfamiliar production problem |
| Melo Practice mode | Generates an applied scenario and can evaluate the response against a rubric | Real-world consequences, users, teammates, or domain expertise |
| Melo's interactive visuals | Lets a learner watch, explore, or recall a concept inside a lesson | A fully open-ended artifact built by the learner |
| Workshops or a real peer | Adds accountability, disagreement, and an audience | Unlimited individual pacing or always-on help |
The mapping is intentionally not “Dewey proved Melo is good.” A product feature only creates an opportunity for inquiry or construction. Whether the learner takes it depends on the prompt, the task, the surrounding teacher or cohort, and the reason the work matters.
Melo's most important design choice is therefore not that it can explain. Any frontier chatbot can explain. It is that explanation is only one mode among activities that require output: Quiz, Practice, Explain Back, fill-in-the-blank checks, matching, flashcards, interviews, and interactive visual recall. The generative UI system behind those visuals uses structured components so an explanation can become something the learner manipulates and is tested on, not only another paragraph.
The hardest layer remains outside Melo: purposeful making. A learner should leave the loop with something that exists because they understood—a working agent, an evaluation rubric, a prompt experiment, a decision memo, or an explanation another person can use.
A 20-minute learning-by-making loop you can run now
Pick one concept inside an explainx.ai pathway and give the session a destination: “I need to design a retrieval test for my project,” not “teach me RAG.” Then run this loop:
/teach Give me only the minimum scaffold I need to design a RAG evaluation.
Stop before giving me the finished design. Ask me one question at a time.
/practice Give me a small RAG failure scenario and a visible rubric.
Wait for my attempt before giving feedback.
/explain-back I will explain why my evaluation catches the failure.
Challenge the weakest assumption in my explanation.
Then leave the chat:
- Build the smallest artifact that embodies the idea.
- Run it against one case you did not use while designing it.
- Show the output to another person, or write a short note for the next person who must use it.
- Revise the artifact based on what they misunderstood or what the test exposed.
This preserves the legitimate value of a Khan-style explanation while refusing to stop there. The AI shortens the route to a meaningful attempt; it does not take the attempt away.
What people are asking
“Should an AI tutor refuse to answer directly?”
Not always. Direct answers are appropriate for factual lookup, accessibility, review, and moments when missing context prevents any useful attempt. The design error is defaulting to a complete answer when the learner's actual goal is skill acquisition. A good tutor should know whether this turn is reference, instruction, practice, or assessment.
“Is making automatically better than listening?”
No. Busywork is still busywork, and an artifact without feedback can preserve a misconception. The making must require the target concept, expose a decision, and produce evidence that can change the learner's next attempt. Our review of LLM simulation games for difficult concepts reaches the same conclusion: interaction is useful when it makes a mechanism visible, not merely because something moves on screen.
“Can an AI create motivation?”
It can reduce friction, personalize examples, offer encouragement, and make progress visible. Those can support motivation. It cannot reliably supply the deeper purpose that comes from identity, curiosity, responsibility to others, or a real problem the learner chose to solve. That is why Khan's later emphasis on human systems is not a retreat from technology; it is a more complete account of what the technology cannot originate.
The better question for every AI learning product
The wrong evaluation is “How good was the explanation?” The better evaluation is:
What did the learner have to notice, produce, defend, and revise because this tool existed?
If the answer is “nothing,” the product delivered information. If the answer names a learner-created artifact and the feedback that changed it, the product may have designed a learning experience.
Khan Academy's videos can be excellent scaffolds. Its exercises and mastery system are meaningful attempts to move past passive viewing. Khanmigo can be a useful support inside that system. The lesson from weak voluntary engagement is not that explanation, practice banks, or AI tutors are worthless. It is that availability is not purpose, and assistance is not agency.
That is the standard explainx.ai should be held to as well. Melo succeeds only when its fluent answer becomes the beginning of the learner's work, not the end.
Related on explainx.ai
- Introducing Melo: the AI learning copilot built into explainx.ai
- How Melo teaches with generative UI
- Social learning for AI: why solo chatbots fail
- AI homework scores rose while exam scores fell
- Dartmouth's AI tutor study: why quizzes beat optional chat
- The doer effect and AI-graded interactive textbooks
- Introducing interactive AI learning pathways
- What schools should teach in the AI era
Sources: Punya Mishra's original essay · Matt Barnum's Chalkbeat interview with Sal Khan · Hacker News discussion
This analysis reflects the cited product descriptions, reporting, explainx.ai surfaces, and public discussion available on August 24, 2026. AI learning features and education research can change; verify current product behavior and treat generated feedback as a supplement to human judgment.