I think you're right on about the second problem. Generative AI is good, at least in a limited sense, at producing superficially plausible simulations of generic texts: "Make me *a* [whatever]." What it's not good at is responding to specific situations in a grounded and purposeful way: "Make me *this* [whatever]." As you suggest, effective teaching requires grounded, purposeful response to specific students in specific situations, and gen AI can't really get there - it can't make the leap from plausible(ish) to purposeful.
(And this is even setting aside the embodied aspects of teaching and learning as a specifically human interaction.)
That's very well said. My friend and teacher Dylan Kane has observed there can be a tension "between helping a student solve a problem and helping a student learn something." I think many if not most AI-as-tutor enthusiasts fail to grasp that distinction.
Agreed, teaching is hard, and developing benchmarks upon which to judge an AI tutor’s use of the disparate methods and techniques at a teacher’s disposal is quixotic and a fundamental misunderstanding of both AI and Education. Teaching, as you rightly point out, occurs in novel ecosystems—unique mixes of learner needs, behaviors, social dynamics, prior knowledge, teaching style, etc.
The confounds are seemingly endless.
My question is whether we may take narrowly defined teaching methods, a Worked Example for instance, and come to understand or generate the optimal base structure of it?
That is, don’t ask an AI to tutor a student, but perhaps it can give you a great approximation of what the average version of X teaching method or learning object is given the context provided. It won’t put all the elements together and properly apply methods (hell expert teachers don’t always get this right. There is no instructional methods checklist to follow) but perhaps the narrowly defined approximation of a single method embued with context is an excellent building block for the novice teacher.
Thanks for this thoughtful comment. I am not entirely sure what you envision in terms of defining the "optimal base structure," but I am curious about using these tools as no-stakes subjects for pedagogical experiments. I hinted at this here: https://buildcognitiveresonance.substack.com/p/the-new-version-of-chatgpt-might
I think you're right on about the second problem. Generative AI is good, at least in a limited sense, at producing superficially plausible simulations of generic texts: "Make me *a* [whatever]." What it's not good at is responding to specific situations in a grounded and purposeful way: "Make me *this* [whatever]." As you suggest, effective teaching requires grounded, purposeful response to specific students in specific situations, and gen AI can't really get there - it can't make the leap from plausible(ish) to purposeful.
(And this is even setting aside the embodied aspects of teaching and learning as a specifically human interaction.)
That's very well said. My friend and teacher Dylan Kane has observed there can be a tension "between helping a student solve a problem and helping a student learn something." I think many if not most AI-as-tutor enthusiasts fail to grasp that distinction.
https://fivetwelvethirteen.substack.com/p/learning-that-doesnt-stick
Agreed, teaching is hard, and developing benchmarks upon which to judge an AI tutor’s use of the disparate methods and techniques at a teacher’s disposal is quixotic and a fundamental misunderstanding of both AI and Education. Teaching, as you rightly point out, occurs in novel ecosystems—unique mixes of learner needs, behaviors, social dynamics, prior knowledge, teaching style, etc.
The confounds are seemingly endless.
My question is whether we may take narrowly defined teaching methods, a Worked Example for instance, and come to understand or generate the optimal base structure of it?
That is, don’t ask an AI to tutor a student, but perhaps it can give you a great approximation of what the average version of X teaching method or learning object is given the context provided. It won’t put all the elements together and properly apply methods (hell expert teachers don’t always get this right. There is no instructional methods checklist to follow) but perhaps the narrowly defined approximation of a single method embued with context is an excellent building block for the novice teacher.
Thanks for this thoughtful comment. I am not entirely sure what you envision in terms of defining the "optimal base structure," but I am curious about using these tools as no-stakes subjects for pedagogical experiments. I hinted at this here: https://buildcognitiveresonance.substack.com/p/the-new-version-of-chatgpt-might
Thank you! The incoherence of that paper could warrant an entire essay unto itself.