Bloom's Taxonomy, Teaching, and LLMs

Recent discussions of LLMs in the classroom have me reflecting on Bloom's Taxonomy of the Cognitive Domain. Here's a nice visual summary of its revised version.

Blooms Taxonomy of the Cognitive Domain
(By Tidema - Own work, CC BY 4.0, https://commons.wikimedia.org/w/index.php?curid=152872571)

Bloom's Taxonomy, as it is called, is a standard reference model among teachers. The idea behind it is that a learner starts from the bottom and works their way upward. As far as I know, it has not been empirically validated: it's more of a thought piece than science. This is reflected in the many, many variations I've seen in the poster sessions of games conferences, where some young scholar proposes a play-based inversion that moves some piece into a different position on the trajectory. All that is to say, take it with a grain of salt. The fact remains that this model has had arguably outsized influence on the teaching profession. (Incidentally, I prefer the SOLO taxonomy.)

There's been a constant refrain the past few decades among a significant number of educators and pundits that technology has made obsolete the remember stage. Why memorize this table of values when I can look them up? Why remember how this word is spelled? Spellcheck will fix it for me. My skepticism of the concept has only increased as I have worked with more and more students who use digital technology as a crutch rather than a precision instrument.

LLM-generated code comes up in almost every conversation I have among teachers and practitioners in software development. There are ongoing studies into the short- and long-term implications of using these tools. My observations are more anecdotal, but it's no exaggeration to say that every professional developer and almost every educator has landed in the same place: LLMs can generate useful code, but knowing what to do with it requires prior knowledge. That is, the errors within the LLM-generated code are often subtle and require knowledge of both software engineering and the problem domain. 

From the perspective of Bloom's taxonomy, a developer with a code-generating LLM is evaluating its output. They come to their evaluation by building upon the richness of cognitive domain skills that undergird it. At the very fundamental level, they bring to bear a vast amount of facts about the praxis of software development that they have remembered and understood.

If Bloom is right, then among the worst things we could do in software development education is throw students at LLMs before they have the capacity for viable evaluation. Indeed, before LLMs, the discussion around the water cooler was often about how to stop students from just searching Stack Overflow for answers and submitting those. Before Stack Overflow, it was that students were searching the web for definitions rather than remembering them. My hypothesis for learning software development then is something like this:

• Google search eliminates the affordance for learning to remember.
• Stack Overflow eliminates the affordance for learning to understand.
• LLMs eliminate the affordance for learning to apply.

This hypothesis frames the quip that I share when an interlocutor discovers that I am a professor and, inevitably, asks what I think about students using ChatGPT. My answer is that I'm considering banning spellcheck.