Creating AI Resistant Tasks In the Classroom

How can teachers modify their normal practices to create more AI resistant tasks?

I’ve been writing recently about the importance of viewing generative AI in the classroom in a positive way. This is partly pragmatic - it’s here to stay and we shouldn’t pretend otherwise - but it’s also idealistic. I’ve long believed in human-centred design as the approach to help us chart a course through a dizzying array of new educational tools, devices and jargon. If AI is the catalyst that pushes us to set tasks that are more creative and more reflective, then I’m in. 

A couple of weeks ago, I posted a video illustrating how Gemini, Claude and ChatGPT could create explanatory notes for a scene from Romeo and Juliet and then use them to generate a sequence of classroom activities designed to deepen the students’ understanding of the play. 

The first ideas were OK but then I asked the bots to make the tasks more AI-resistant; more difficult for a student to simply hand in ChatGPT’s answer without checking it or, indeed, without even reading it. I know many celebrity ‘writers’ have made pots of cash by taking the same approach to their ghost writer’s work but it’s not something we should encourage.

I tried one of the tasks with a real student. I, or rather, Claude asked him to write Lord Capulet’s memoir, looking back on the events of the play from a point ten years later. I don’t pretend it’s a new task but, from an educational point of view, I think it’s an interesting and useful one.

For a start, it demands close textual knowledge. Not every student will notice or consider the significance of the one line in Act 3 Scene 5 when Capulet mentions all his other children have died but I’d suggest that fact is likely to be at the front of a father’s mind as he reflects on the death of his remaining daughter.  It also requires the student to be empathetic as well as creative and they have to have a clear interpretation of the end of the play. Do they imagine that the fragile peace will have held or will the families have descended into more bloodshed and enmity? 

Crucially, to do well, a student has to show imagination and adaptive reasoning, taking the little information the play gives us about Lord Capulet and applying it to a new situation. 

Happily, the qualities that make it a good task educationally are the very same qualities that make it relatively AI resistant. 

Generative AI, at the moment, at  least, isn’t very good at adaptive reasoning. At its core, it works by recognising patterns in the data it’s trained on and then using those patterns to predict what is statistically likely. But it doesn’t understand the underlying principles in the way a human does. 

That means it struggles to apply knowledge to completely novel situations and, in particular, it has difficulty with counterfactuals, with ‘what-if’ questions that demand the consideration of alternative scenarios. Generative AI can sometimes mimic a human’s ability in this respect but, in truth, its mimicry is rather unconvincing. Less Steve Coogan and more Bobby Davro, perhaps.

AI struggles with empathy and creativity for similar reasons. With none of the subjective experience we humans bring to every situation, AI can only mimic the sadness of a bereaved father. It might have read pretty much everything humans have written about sadness and it can process the concept of sadness on that basis; but it can’t feel anything and it can’t express the unique, individual experience of a human.  

It is only by understanding the limitations of generative AI and it is only by exploring those limitations openly in the classroom with our students that we will reach the point where we can all fully benefit from its strengths.

My student had a stab at the task first on his own. It was pretty good.

Then I asked Gemini, Claude and ChatGPT to try. At first glance, they too were pretty good. You can see them here:

Gemini

Claude

ChatGPT

They wouldn’t fool any but the most cursory reader but they were detailed and did express a sense of the character’s grief. 

Because I have to confess that my marking has, on occasions, been a little more cursory than it might have been, I did run the bots’ work through some AI detectors. They gave me a range of results - a little surprising, perhaps, given that the work was 100% AI generated but, nonetheless, they did all spot that it wasn’t a student’s work.

So, the bots weren’t that much use from a cheating perspective but, taking a much more positive approach, they were extremely useful because they instantly gave us some alternatives against which my student could compare his own ideas. Comparison and evaluation. Again, we’re effortlessly in the low-oxygen zones of Bloom’s mountain!

We noticed, for example, that Claude’s Capulet expresses regret that his final words to his daughter were angry and vitriolic. This is a great point, easily missed amidst the play’s frenetic final action. At least, we’d missed it! My student was able to use this and other ideas to develop his own writing. 

His second draft was much better. 

This is a positive example where AI had assisted him to produce something much more thoughtful and effective than his first version and it’s a very long way from the ‘Easy Button’ cheating we all want to avoid.

Of course, now we need to develop and use rubrics that properly value self-reflection and re-iteration; if AI hastens our journey down that path, then it will be a very good thing.