Chapter 22
AI Fluency: When Anyone Can Build Anything, What Is Worth Building?
Last month, I started teaching an independent study course alongside an art professor called AI Fluency. We meet in the library’s media lab rather than a conference room because we wanted students to be able to pull up their projects, put them on a screen, and show us what they had made. The premise of the class is intentionally open-ended: students can work on almost any project they want, and they are encouraged to use as much artificial intelligence as they can to bring it to life. We assumed that much of the semester would involve teaching students how to use a growing collection of AI tools and services, how to prompt effectively, how to overcome obstacles, and how to use increasingly capable systems to expand what they could create.
Right away, however, we encountered a different problem. We didn’t know how ambitious their projects should be. Some students proposed ideas that sounded like fourteen-week projects but could probably be completed with AI in an evening or a few days if they pushed themselves and their tools hard enough. Others imagined enterprise platforms or apps whose complexity they didn’t yet fully understand. The boundary between those two categories is increasingly difficult to find because the technology’s capabilities keep moving, and they don’t move evenly.
Researchers at Harvard Business School have described frontier AI’s capabilities as a “jagged technological frontier”: these systems can perform some tasks at superhuman levels while struggling with others that appear much simpler. The challenge is that the frontier does not stand still. Each new generation of models shifts the boundary of what is possible, and unless people regularly test those boundaries, they may not realize how much the frontier has moved.
The Jagged Frontier
That jagged frontier is exactly what we are encountering in the classroom. Something that would’ve required a programmer, designer, researcher, and several weeks of work only a few years ago can now sometimes be prototyped by one student working alongside an AI system like Base 44, Lovable, Claude Code, or ChatGPT Codex. At the same time, there are still projects whose technical, organizational, or social complexity cannot simply be prompted away. The instructor’s problem is therefore no longer simply deciding whether a project is difficult or ambitious enough. It’s helping students calibrate their ambition and imagination against a capability frontier that may move dramatically during every semester of their program.
As we went around the room discussing project ideas, one student described something she wanted to build for her mother, an independent hairstylist who still ran parts of her business the old-fashioned way. Her mother had loyal clients whose preferences she knew intimately: their previous hairstyles, the colors they liked, when they had last visited, and all of the small details that accumulate through years of working with the same people. The student imagined creating a digital service that could help her mother keep track of that information, manage appointments, and charge cancellation fees when clients backed out at the last minute. At first glance, it sounded like a straightforward software-development project. With the help of AI coding tools, she could probably begin prototyping some version of it almost immediately.
But as we talked, I realized that teaching her how to build the application was probably the least interesting thing I could do. I suggested that she zoom outward first. What salon-management platforms already existed? What features did they offer? Were independent hairstylists already solving these problems with existing tools? If so, why was her mother not using them? Were the platforms too expensive, too complicated, poorly designed for independent stylists, or simply unfamiliar to someone who had spent years running a business without them? I encouraged her to conduct a competitive analysis, but more importantly, to treat her mother as a user research participant rather than simply assuming that she already understood her current needs.
Asking New Questions
That meant asking a different set of questions, ones more targeted and objective. How did her mother actually run her business? What information did she keep? Where did she lose time? What happened when a client canceled? What parts of the process frustrated her? How did she currently remember a client’s color formula or previous hairstyle? The student could then talk to other independent hairstylists to determine whether those problems were unique to her mother or represented something larger. Maybe the opportunity was not a booking website at all. Maybe it was a better way to preserve client histories. Maybe existing software already solved the problem. Maybe the real barrier was not technology but usability, cost, or habit.
Suddenly, the project looked different. Was she trying to build a website for her mother with appointment booking and cancellation fees? Was she trying to digitize the informal client records her mother had accumulated over years? Or was there a broader opportunity to create a platform that helped independent hairstylists move from an old-fashioned way of running their businesses into something more intelligent and automated? The application was still a possible output, but the learning was everything required to determine whether the application should exist in the first place. She needed to start at the beginning of the design process rather than jumping straight to the part where she started coding up a potential semester project.
That distinction has started to change the way I think about AI fluency. I assumed that a class like this would spend much more time teaching students how to use an expanding collection of AI tools: how to prompt them, research with them, analyze data with them, generate media with them, code alongside them, and eventually coordinate agents capable of accomplishing increasingly complex tasks. Those skills still matter, but the more capable the technology becomes, the less convinced I am that teaching the mechanics of individual AI tools is the hardest part of the problem. Increasingly, AI can help students figure out how to build just about anything now. The more difficult question is whether they understand what is worth building, who it is for, what problem it solves, and how they would know whether it works.