Seems like a very broad conclusion for such a narrow study. I’d question whether it is as successful when it comes to concepts that are more fluid, such as thos

Seems like a very broad conclusion for such a narrow study. I’d question whether it is as successful when it comes to concepts that are more fluid, such as those found in the humanities. I also don’t really see the advantage of this kind of examination. One still needs to develop tests to trigger the students thinking on the concepts—it seems we merely trade one potential form of error on the part of graders (checking students answers) for another (misinterpretations of results or bugs or in the algorithm or unanticipated changes in the model). Just because certain practices are historically prior does not mean they’re necessarily worse. Certainty is a slippery beast. The results here still depend entirely on the assumptions of the developers of this ML algorithm, their notions of what it means to “understand something” etc. Our era’s obsession with throwing technology at literally everything will bite us in the bum at some point (if it hasn’t already).