There was something about exploring a menu that nobody calls learning but was.
You open Options without really knowing what you're looking for. You see a function with a strange name, hover over it, read the tooltip. You didn't need it. But now you know it exists. Three months later, at exactly the right moment, you remember it. You use it. You save someone two hours.
That has no formal name in UX. But it should. Because it was probably the most silent and effective way that ever existed to expand what a user can do with a tool.
AI killed that. And it killed something else that is even worse.
The Accident as Method
You didn't learn Excel in a course. You learned Excel by right-clicking without knowing what you were going to find.
The rich interface, with all its flaws, with its 300 buttons and cascading menus, had a property that nobody intentionally designed: it exposed you to what you didn't know. Discovery wasn't a software feature. It was a side effect of visual density.
Navigating a toolbar was like walking through a new city. You weren't going anywhere in particular. But you came out knowing more than when you went in.
Photoshop was learned that way. Word too. AutoCAD. Not from documentation or YouTube videos, but from that clumsy, goalless exploration that in another context we'd call curiosity.
What the Chat Cannot Teach You
The problem with AI chat isn't that it's bad. It's that it's too good at what you know how to ask.
You ask how to do something, it answers. You ask it to do something, it does. The result arrives without friction, without detours, without the side corridors where you used to get distracted and learn.
If you don't know a feature exists, you don't ask for it. And if you don't ask, the agent doesn't show it. Not because it can't, but because it has no reason to. The chat operates in the universe of what you already know, with the illusion that that universe is complete.
The graphical interface was imperfect but honest about its own complexity. It showed you everything, chaotically, and you decided what to explore. The agent shows you exactly what you asked for, perfectly, and the rest stays invisible.
GPS gets you to the destination. But you no longer know what the neighborhood looks like.
The Knowledge That Stayed in the Tool
Up to here, the problem is that you don't discover what you don't ask. But there's a second problem, deeper.
The game developers who built the first 3D engines in the 90s needed to know physics. For real. For a character to jump in a way that felt natural, someone had to understand mass, impulse, gravity, friction. They wrote the equations. They tuned the values. They understood why one jump arc felt satisfying and why another felt wrong.
Today, Unity and Unreal handle all of that. The developer adjusts parameters until it "looks right." It works. But the knowledge of why it works doesn't live in the person, it lives in the engine.
With AI, that process accelerates and expands to almost everything.
The designer who uses generative AI for layouts doesn't develop intuition about visual hierarchy. The programmer who uses Copilot to write code may not understand what that code does or when it fails. The analyst who asks an agent to summarize data doesn't necessarily need to know how to read the raw data.
It's not that they're less capable. It's that they never had to be. The tool absorbed the learning that used to be mandatory to reach the result.
The Problem of Not Knowing What You Don't Know to Ask
This closes the loop.
The accidental discovery of rich interfaces gave you vocabulary. It taught you that things existed that you didn't know how to name. And that vocabulary was what later allowed you to ask better questions, search better, understand more.
If you didn't explore, you have no vocabulary. If you have no vocabulary, you don't know what to ask. And if you don't know what to ask, the agent, which only answers what is asked of it, gives you back exactly the limit of what you already knew.
The result isn't a more powerful user. It's an efficient user inside a perimeter they don't know exists.
The Point
AI isn't the problem. The problem is confusing getting the result with learning the domain. They're different things, and for a long time software, with all its interface clumsiness, mixed them together in a way that turned out to be useful.
The physics you don't need to know for the character to jump correctly still exists. The difference is that now it lives in the engine, not in you. And if the engine ever fails, or the case is unusual, or the tool disappears, nobody in the room will know what to do.
That doesn't appear in any AI demo. But it's the question worth asking.
If you want to keep reading this kind of analysis, subscribe to the macareno.net newsletter. No algorithms, no feeds. It arrives when there's something worth thinking about.