My Students All Used the Same Font (And What That Taught Me About AI)

My Students All Used the Same Font (And What That Taught Me About AI)

Three students walked in last week with their typographic projects and they’d all chosen Humanist sans-serif, clean and modern and versatile, and when I laid them side by side the typeface was the same, the hierarchy was the same, the visual language was the same, and I realized what had happened was that all three of them had asked ChatGPT which font to use and ChatGPT had given them the statistically most-likely answer, the one that appears in ten thousand design projects, the one that represents what a machine learning model thinks design should look like when nobody’s pushing it in any particular direction. Technically correct, aesthetically indistinguishable, and it raised a question that’s been sitting in the back of my head ever since: are we training their eyes or are we training them to defer to whatever artificial intelligence thinks is statistically appropriate?

I asked them that, and I said: “Did any of you try different fonts before asking?” and most of them hadn’t, so I said: “So you’re training your eye on whatever AI thinks is correct?” and they looked at me blankly the way students do when the question lands as criticism even though I didn’t mean it that way, and I had to backtrack and say: “Here’s the thing; the best way to learn to see is to make a wrong choice and figure out why it’s wrong, and when you ask the AI and take its first answer, you skip the part where your eye develops.” That’s not a moral position I’m taking. That’s just the mechanism of how learning works, and if you’re outsourcing the choice before you’ve made it, you’re outsourcing the part where the conditioning happens.

This became part of the creativity experiment, and not because I’m some Luddite who thinks AI is destroying everything, but because I started to wonder whether structured exercises could counteract what I was seeing; whether there was a way to use AI as a tool that helps you think instead of a tool that does your thinking for you. The distinction matters and it’s not subtle. A tool that helps you think pulls you deeper into the problem. A tool that does your thinking for you lets you skip it, and the skipping feels fine in the moment because the answer is competent and it’s immediate and it doesn’t require any struggle on your part.

So the framework became: ask the AI, get the answer, acknowledge that the answer is good, and then push past it and make a different choice. Use AI, then use your judgment on the second answer, then use your judgment on the third. The AI isn’t the problem, and I want to be clear about that because there’s a lot of hand-wringing about AI that’s really just hand-wringing about intelligence, and I don’t buy that narrative. The problem is when the first answer becomes the answer, when you train yourself to recognize the statistically probable as the correct, when you let the machine’s confidence become a substitute for your own thinking.

The students who understood this and pushed past it got interesting; they’d ask the AI for five different approaches and each one would land in a different place, and they’d sit with that tension and they’d develop something that was theirs instead of something that was the machine’s polished version of what it thinks they should make. The ones who took the first answer and moved on produced work that was competent and forgettable, the kind of thing that looks designed but doesn’t look like anything in particular, and there’s a version of that that’s exactly what the job market asks for, so I’m not going to tell them they’re wrong; I’m just going to tell them what I see when I look at their work compared to the work of the people who pushed.

When you’re training an eye to see, the wrong choice teaches the right thing. You learn what you don’t like by making something you don’t like and then figuring out why it doesn’t work. You learn what hierarchy means by creating a hierarchy that fails and understanding what made it fail. You learn what your taste is by running up against its edges and pushing back against it and choosing something different and living with the consequence of the choice. All of that is invisible when you ask the machine and take the answer, and you can’t speed that up and you can’t automate it and you can’t get the benefit of it by watching someone else do it.

This is why I’ve kept building the exercises, because most of the exercises are designed to make the automatic unavailable, to remove the easy answer and force generation in a constrained space where the first thing your brain reaches for isn’t an option. You ask a machine to do the same exercise and what you get, if you’re lucky, is something that looks like a creative violation instead of something that is one. The machine can bend its rules but it can’t experience the friction of bending them, and that friction is where the learning lives.

So the font situation became a teaching moment, but not in the way they thought it was. I wasn’t saying don’t use AI. I was saying use AI and then keep thinking, and the keeping thinking part is the part where you develop from someone who can recognize good design to someone who can make good design, and that’s the transition that matters. You know what I mean; there’s a difference between knowing what should happen and being capable of making it happen, and the difference lives in the space between the first answer and the answer you commit to after you’ve pushed past it.