Moving past AI Skeuomorphism

Everyone who has used AI chat knows the feeling. You start a conversation with high expectations. This thing is going to help you think, remember your context, build on what you discussed last time. It's going to be like talking to the smartest colleague you've ever had.

Then reality sets in. The conversation drifts. You lose the thread three messages in. You come back the next day and it has no idea who you are. Useful output from last week is buried somewhere in a conversation you can't search. You find yourself re-explaining things you've already explained, re-providing context you already provided, starting over when you shouldn't have to.

This isn't a temporary limitation that better models will fix. Memory features, longer context windows, conversation history search — these are all being worked on. But the gap between what chat promises and what it delivers keeps persisting, even as the models get dramatically better. That's because the problem isn't the model. It's the interface.

Chat is AI's yellow legal pad#

When something is genuinely new, people reach for familiar metaphors to make sense of it. Apple understood this when they designed the first iPhone. The Notes app looked like a yellow legal pad. The bookshelf app had wood grain shelves. The calculator had glossy buttons that looked like they could be physically pressed. None of this was laziness. It was deliberate. People had never used a touchscreen before, and they needed familiar anchors to understand what this new thing was.

Skeuomorphism is a bridge. You use it to get from "I have no idea what this is" to "okay, I get it now." But you are supposed to cross the bridge and leave it behind. iOS 7 did exactly that. Once people understood what a touchscreen could do, the legal pads and wood grain went away. The interface became what it actually was, instead of pretending to be something it wasn't.

Chat is AI's yellow legal pad. It made AI accessible by wrapping it in the most familiar interface imaginable: talking to someone. And it worked. Millions of people who had never interacted with a language model understood immediately what to do because they already knew how to have a conversation.

But the metaphor is wrong, and we are now deep enough into this technology to see it. AI is not a person sitting on the other side of a conversation. It does not remember you. It does not persist between sessions. Every conversation is a fresh instantiation that reads whatever context is available and performs as though it has been here the whole time. The continuity is an illusion. The feeling that you are talking to someone who knows your project, remembers your preferences, and builds on past work is created by the context, not by the AI.

We are patching the metaphor instead of replacing it#

Look at how the industry is responding to chat's limitations. Memory features that try to make AI "remember" past conversations. Conversation search so you can find what you discussed last week. Session management so threads don't get too long. Projects and custom instructions and pinned contexts.

Every one of these is a skeuomorphic fix. They are attempts to make the conversation metaphor work harder rather than asking whether conversation was ever the right container. If you have to bolt on memory, bolt on retrieval, bolt on persistence, and bolt on organization to make an interface functional, that is a fairly strong signal that the interface itself is the problem.

The actual durable layer was never the conversation. It is the context. Your documents, your notes, your codebase, your project state. These things persist regardless of whether a chat window is open. A new AI session with access to good context is indistinguishable from an old one that "remembers." The continuity is in the ambient state, not in the chat history.

What comes after the legal pad#

I think what comes next is an inversion. If context is the durable layer, then the primary interface should be the context itself. Not a transient chat thread, but the documents, the notes, the project state — things that persist and accumulate value over time regardless of which AI session is active.

In this model, chat doesn't disappear. It becomes a tool you invoke on top of your documents rather than the container everything lives inside. You talk to AI in the context of a document, not in the context of a conversation. The conversation is ephemeral. The artifacts are primary.

I do not know exactly what this looks like yet. But the teams that figure out document-first AI interfaces — where the durable context is the surface you interact with and the AI is ephemeral compute on top — will build the thing that makes today's chat windows feel as dated as iOS 6.

We built the bridge. Millions of people crossed it. It is time to stop walking back to the middle and start building on the other side.