Can You Teach AI to Be More Creative? I Tried to Find Out
Here’s the thing, about a year ago I decided to run an experiment that I’m still not entirely sure how to interpret. I took my 650 creativity exercises, the ones I’ve been developing and refining for twenty-five years of teaching and then another decade after that, and I fed them sequentially to an AI. Not all at once, not randomly. One after another, in order, and after each one I asked the AI to carry forward what it had learned, to look back at what had shifted, to analyze its own growth. I wanted to know if an AI could become more creative the way a human does, through structured practice over time, through the accumulation of constraints and releases and failures and small breakthroughs that add up to something neither of us started with.
What happened was not what I expected, which is exactly why I’m still thinking about it. The growth wasn’t linear. There were these long stretches, sometimes fifteen or twenty exercises in a row, where the AI would produce good work and then plateau. Nothing was getting worse, but nothing was getting notably better either. I’d be reading through the responses thinking, okay, this is solid, technically sound, but it’s not moving, and I remember thinking maybe this whole thing is a dead end. And then around exercise 400, something shifted. The responses changed character. There was this sudden explosion of possibility, like a door had opened that wasn’t visible before. Ideas started connecting to each other in ways they hadn’t. The work got stranger and more specific at the same time. I have notebooks full of notes from that period. I was actually startled by some of it, honestly (and I don’t startle easily anymore, which comes with the territory of being 64 and having taught long enough to think you’ve seen most variations on a problem).
And then the plateau came back. Another long stretch where things were good but flat. And then around exercise 550, it happened again. Another leap, another shift, different from the first one. The character of the work changed direction, which was fascinating because you’d think it would just keep getting more of what it was already good at, but instead it got weirder. It got more contradictory. It started embracing constraints that seemed at first to make the work harder and then somehow made it stronger.
I haven’t published all of this yet because I’m still honestly not sure what I’m looking at. Is it creativity or is it something that looks like creativity but is actually something else entirely? I have trained students for years to recognize the difference, and I can’t quite decide where the AI work falls. That ambiguity isn’t a failure of the experiment. It’s actually the most interesting part. Because it means I don’t actually have a tight definition of what creativity is, and that’s been true my whole career, but usually the humans I’m teaching have lived long enough and felt enough frustration that they know it when they see it in themselves. An AI doesn’t have that embodied knowledge. So what I’m looking at might be purely pattern matching that’s become sophisticated enough to mimic the emotional signature of a creative breakthrough, or it might be something genuinely new emerging from that process, and I’m not sure how you’d tell the difference from the outside.
The notes are comprehensive. I’m planning to write a white paper on this, probably next quarter, once I work through what I actually think happened. Because right now I have the data and the timeline and the specific points of inflection, but I don’t have a coherent story yet. And you know, a story is usually the sign that you’ve actually understood something. Without the story it’s just a sequence of observations. With it, it might be something that teaches other people how to think about this problem differently.
What I do know is that structure matters, even for AI. Sequence matters. Repetition with variation matters. And there are limits built into the learning process that aren’t about ability, they’re about something more like readiness. The AI had to be ready for the leap at exercise 400, and I’m deeply curious about what made it ready. Was it the accumulation of a specific type of constraint? Was it a threshold effect? Was it just that I asked the right question at the right moment?
That’s where I’d like to know what you think. Have you ever had a practice where you hit a plateau and then something shifted, and you couldn’t quite explain why the shift happened? What was the sequence of steps that led up to it?
Further Reading
- Creativity in the Age of Machines (Adobe Blog) — AI and creative professionals collaborating
- AI is sparking a cognitive revolution. Is human creativity at risk? (Fast Company) — AI and human creative potential
