What Lessons Can We Learn From ChatGPT About AI and Education?

Like many other educators, I've been spending a lot of time in recent weeks learning about and playing with new generative AI technologies like ChatGPT and DALL-E. As an edtech veteran and an early AI developer, I've also been asked by a lot of friends and colleagues for my thoughts about the implications of these technologies for teaching and learning.

A lot has been made since ChatGPT's public launch a few short weeks ago about the human-like responses it generates to the sorts of prompts high school and college teachers might give their students for a typical writing assignment. Many worry that chatbots and similar technologies have already made it impossible to stop students from cheating on homework, scholarly research, and other academic work.

Is this concern justified?

Among the more vivid memories from my high school years was a spirited debate I had with a favorite teacher, Andrew Glassman, about two poems describing the plight of soldiers during the First World War: "Dulce et Decorum Est" by Wilfred Owen, and "A Soldier" by Robert Frost. I was deeply moved and profoundly captivated by the raw emotion and directness of the former, while Mr. Glassman argued on behalf of the superior sophistication and subtlety of the figurative language employed by Frost in the latter.

Remembering this exchange recently, I thought I'd engage ChatGPT in the debate:

One thing that anyone who's used ChatGPT has already discovered is that it has an allergic reaction to anything resembling an opinion - even an informed or conditional one. Ask it to tell you whether Macs or PCs are better, whether you should buy a hybrid or electric car, or if Magnus Carlsen is better at chess than Bobby Fischer, and ChatGPT goes to great lengths not to weigh in on the matter. This makes sense, since ChatGPT's human handlers at OpenAI are clearly trying to avoid ChatGPT being accused of bias, misinformation, or anything beyond "just the facts, ma'am." But this design decision is itself a bias - and a weakness - of ChatGPT, which limits its utility as an artificial intelligence, and one reason it couldn't, in its current incarnation, pass the Turing Test or game the Voight-Kampff test in Blade Runner.

In fact, ChatGPT is so averse to expressing an opinion, it often won't even be coaxed into doing so when explicitly asked to take a stand and make an argument on behalf of a position, although it is amusingly meticulous about following formatting requests to the letter. You'll notice in response to this modified prompt (apologies for the typo) that ChatGPT offers the same indefinite response, but gamely finds a way to turn it into three paragraphs:

So, knowing this critical limitation of ChatGPT in its current (early) incarnation, what if we avoid asking it to state a conviction or defend a position, and focus it instead on textual analysis. Here's a prompt that any ninth-grade English teacher would expect his students to be able to answer with accuracy and aplomb:

At the most fundamental level, this prompt requires ChatGPT to understand the basic difference between metaphor and simile, and to be able to elicit relevant examples. For a student to get an A on this assignment however, one would also expect said student to be able to compare how the two poets use these forms of figurative language to enrich their writing, and deepen the emotional and intellectual engagement with the audience. Here's what ChatGPT came back with:

While ChatGPT does make an interesting assertion (that Frost's use of metaphor and simile is subtler and more understated than Owens's - surprisingly eschewing its usual aversion to opinion!) it fails to give any examples to justify the assertion. Even more troubling, however, is its apparent lack of understanding of the distinction between - or even a basic understanding of - metaphor and simile ("Bent double, like old beggars..." is a simile, not a metaphor. "His cuts and bruises heal as fast as he can fight" is neither a simile nor a metaphor.)

So what are some takeaways for educators from this experiment? For me, the first is that the evolution of AI, whether in the form of chatbots or something else, just reinforces what good educators have always known:

1. It's more important than ever to focus students on exploring and explaining the "hows" and the "whys" vs. the "whos," "whats," and "whens."
2. Ask students to take a stand, form an opinion, and demonstrate their understanding of facts and concepts by creating compelling arguments on behalf of a view. As humans, we are born with opinions, and only over time develop an understanding of the facts that ultimately support or change those opinions. Today's AI algorithms live in the world of facts and rules. They will need to evolve to develop more informed opinions.
3. Don't assume AIs like ChatGPT necessarily even have a grasp of basic facts or concepts. In the last example above, ChatGPT gave me confident yet misleading - and downright incorrect - examples of the concepts of metaphor and simile. Given my familiarity with figures of speech, I quickly identified the inaccuracies; but I wonder how many times someone has asked ChatGPT for an explanation or an example of a more complex, less familiar concept, and received similarly inaccurate results without realizing it. 
4. Which leads me to my final thought on this experiment: Rather than obsessing over student cheating, and how to use AIs to identify the likely output of other AIs (Blade Runner, anyone?) educators should be having fun fundamentally rethinking exam and homework prompts. For instance, if I were teaching the use of metaphor and simile in poetry, rather than give them the prompt I gave ChatGPT, I would start with ChatGPT's response to the prompt, and ask my students to evaluate and grade it!

AI technologies like ChatGPT are still very much in their infancy, and there is no question they will improve quickly over time. But understanding their strengths and limitations today is important as educators seek to understand the implications for their work with students. To me the answer isn't to start an arms race between AI's and anti-AI AI's - rather, it's for educators (and students) to use this moment as an opportunity to rethink the rhythms and rituals of education, and to fundamentally improve the ways that we teach students to think and express themselves in a rapidly evolving world.

I'd like to know what you think. Please comment below!