Khanmigo WANTS to Love Kids but Doesn’t Know How
Khan Academy made some product changes to Khanmigo. Let's see how that went.
A couple of months ago, I noted in my usual spirit of collegiality that Khanmigo doesn’t seem to care much about student thinking. Whether you answered a Khan Academy exercise with gibberish or something quite thoughtful, something just an inch away from correctness perhaps, Khanmigo would assume you know nothing. “Do you know the first step?” it would ask even if the student had already taken several.
The decision to feed the problem statement into Khanmigo’s context window but not the student thinking is an implicit statement of value, one which I promise you is not lost on students. This will feel to many like you’re having a conversation with someone you can tell is not really listening to you.
Well imagine my surprise last week to see that Khan Academy is now feeding student thinking into Khanmigo, making me something of an unindicted co-conspirator here.
I’m not looking for credit, but some kind of notice would have been helpful here because I found out about this product change live while delivering a keynote presentation. I was showing different edtech models to the audience and describing the regard they do and don’t have for student thinking. I was doing this live in a web browser rather than using pre-recorded video (a technique which keynote presenters call “flying too close to the sun”) when Khanmigo surprised me with this product change. A pleasant surprise, yes, but also awkward.
Anyone at Khan Academy who is unhappy about my last fifteen years of unsolicited ✌️ product advice ✌️ is welcome to hold onto this story as a treat.
I was able to improvise the talk pretty easily, though. Khanmigo is managing to integrate student thinking into its model in about the most disinterested way possible. I have asked Khanmigo to clean its room and it has pushed all of its toys under the bed. I have told Khanmigo to eat its carrots and—look, Khanmigo, we can all see that you have hidden them under your plate.
The prompt above is an extremely rich text and presents a view of students and mathematics that I consider quite profane. Let’s chat about the user experience and technology first though.
Dissociative User Experience
You click the “Tutor Me!” button and Khan Academy automatically sends Khanmigo a chat message from you. It takes a moment to realize that the line starting with “Break this question down in to [sic] steps” is something that Khan Academy has sent on my behalf. Both people in the chatbot conversation are using the pronoun “I” and I am neither one of them. It feels deeply weird. This can’t be me. I did not say the thing you said I said. Am I eavesdropping? It isn’t clear to me why they would choose this UX instead of sending the first message invisibly in the background.
Homework Rather Than Help
In this tutoring session, Khanmigo needs you to “tell me the problem you’re working on,” in this case to describe the graph of the sinusoid in words. A sinusoid is a function which, in technical terms, “has a lot going on” and if you can describe it precisely, it’s likely you don’t need much help with it. Khanmigo needs you to do some homework before it’ll help you with your homework and many students will decline.
Uneven Application
Khan Academy feeds student thinking into Khanmigo only sometimes and only sometimes accurately.
On the left side of the image above, Khan Academy has not passed the equation of the line I graphed into Khanmigo. It obviously could. Khan Academy would not have to read that equation from my mind.
On the right side, my thinking has been passed into Khanmigo but in a way that, again, communicates maximum disinterest. Rather than passing my thinking into Khanmigo as the line “y = 2x + 3,” it passes in just the numbers I entered “2,3” which Khanmigo then interprets as the point (2,3) which derails the whole conversation.
A Prompt That Profanes Students and Mathematics
Every time I press “Tutor Me!” to activate Khanmigo, the prompt is the same.
Break this question down in to steps and show it to me in a numbered list.
Also I thought the answer was [my thinking] so remember to show me your best guess as to what I did wrong in the list and reference the step number.
I’m staring at the typo in the first sentence trying so hard not to read into it a total lack of interest in this product change. I’m breathing into a bag. Sal, just add me to your GitHub repo and I’ll catch this stuff before it goes to production. I swear that is all I’ll do.
Again, I cannot speak to the heart or mind of anyone working at Khan Academy, but insofar as product design communicates value and love, this prompt regards math as machine-executable code, a numbered list of steps that a machine can execute one at a time with any error bubbling up the stack and identifying the earliest step that produced it.
The prompt regards students as buggy computers whose errors should be identified and corrected as efficiently as possible. “What did the student do wrong?” the prompt asks, which is fundamentally and tellingly different from “What did the student do right?”
The lie that Khanmigo perpetuates here is that “math is about a huge number of small ideas.” That lie makes math dizzyingly complex and impossible to hold in your head. The truth is that math is actually about a small number of huge ideas. Here that huge idea is that “graphs have meaning” and a teacher can use that idea by pointing anywhere on the graph and asking, “So what do you think this point says about the moon?”
The other lie that Khanmigo perpetuates is that “students are malfunctioning” when the truth is that students are constantly and earnestly integrating new experiences with old ones, turning old ideas into new ones. Kids aren’t buggy computers. Kids don’t want to be treated like buggy computers. Kids like and learn from teachers who see them as something more than just a small malfunctioning adult.
The skilled teacher, one who loves kids and loves math, will ask, “What does this student know about this one big mathematical idea?” Khanmigo asks, “What does this student not know about these many small mathematical ideas?”
Perhaps Khan Academy will clean up this prompt a bit. It wouldn’t surprise me. (Someone please let me know so I don’t embarrass myself in another keynote.) But you can’t prompt engineer your way to classes where students learn math and love learning math. A prompt isn’t a cheat code.
In many ways, you cast your die the moment you get into this field. You define the ceiling on your work almost immediately. You’re drawn into math edtech by a love of students, teaching, math, technology, scale, market opportunity, or some combination thereof. Whatever you love in edtech, that love will determine your product roadmap.
My love runs in exactly the order listed above. Too much of education technology runs in exactly the opposite direction and it will produce something that loves students, and is beloved by students, only by accident.
2024 Jul 17. Some time after I published this criticism, Khan Academy made some changes to the way Khanmigo handles student thinking. Check out my review of those product changes.
For what's it worth Dan, roughly 100% of the people I spoke to who heard your presentation said it was fantastic, so your nimble ad-libbing worked! I've run into the same issue, people find it SO compelling to watch chatbots do their thing in real time, but from a presentation standpoint you might find the product has changed since you last played with it.
"'What did the student do wrong?' the prompt asks, which is fundamentally and tellingly different from 'What did the student do right?'"
This flip was one of the biggest wins I felt from the beginning of this school year (when I felt like I was drowning https://danmeyer.substack.com/p/can-we-get-this-new-teacher-a-quick) to the end of my internship five days ago.
When I first started working in the classroom, if a student shared an idea with the whole class that was not what I expected, I felt the need to close that idea so we could focus on the "correct" one. I wanted to be gentle, so I would say things like "not quite" or "good idea, but..." My teaching was predicated on "no, but", which makes for terrible improvisation.
Within a few months, though, I noticed a change in my questioning. Instead of "no; good try" I would say things like "I like that you're thinking about multiplying." It became easier, with practice, to extract something productive from every student contribution.
Sometimes this meant pulling on one piece of their thinking that led towards my idea. Other times, though, it meant recognizing that the student was approaching the problem in a new way that I hadn't anticipated, and following their thinking that way instead. (These were my favorite, because it allowed me to be genuinely excited about a new idea that I was genuinely unsure about. When you're walking the tight rope without a net, students know it's real and it's so much more fun.)
This was difficult at first but became natural over time. It would be great if LLMs could do the same. I'm hopeful that if Khan Academy keeps listening to this feedback and improving their prompting (or maybe their training data), they might be able to achieve a chat bot that can extract something from the student's thinking that leads toward Khanmigo's idea of what a good solution looks like. I'm much less optimistic that LLMs, as they exist today, can get excited about a new method they haven't seen before in their training data. That kind of thinking requires genuine mathematical reasoning, not just pattern matching, which LLMs are (currently) notoriously bad at.