Will UI survive AI?

LLMs are clearly and quickly changing the software industry. No, they are not perfect, but the latest coding agents are capable enough to completely reshape how many of us build software day to day. Because this shift has happened so fast while the technology is still so young, we can’t help speculating on what it’s going to be like in the next few years.

One such speculation posits that AI will replace UI. What’s the point of menus and icons and buttons… when we can simply tell AI what we need, and it happens?

Call me old-fashioned, but I don’t think UI will go away. Not because today’s LLMs are not good enough, but because UI is inevitable; UI is inevitable because fundamentally, unmediated I/O is impossible. UI isn’t a wall stood up to block our direct access to the system behind it; UI is the access, the bridge between us and the system. Yes, medium truly is the message.¹

But this is not an original insight. So maybe when people claim “AI will replace UI”, what they really mean is that the traditional UI will be replaced by a new, natural language-oriented interface. Like how Claude Code and other coding agent TUIs are gaining over the standard IDEs like Visual Studio Code, etc. Or how the ChatGPT-like chat-based interface is replacing the standard search bar and result pages. In fact, the new kind of UI may not even need visual components, if any. Just talk to the AI and have it respond to you.

While it may be true in certain domains, I don’t think it holds in general. The greatest strength of natural language as interface is its flexibility. We’re not bound to a fixed set of menu items and buttons to input our command. Natural language excels where tasks are ambiguous, open-ended, and loosely specified. But this power and freedom come at a real cost: natural language is a serial medium with low information density, where the meaning depends heavily on the context. High latency, low bandwidth, low fidelity.

As a result, natural language as an input interface may require an essay to achieve what a carefully designed interface can do with a simple click of a button. Natural language is an even poorer fit as an interface for any output that is information-dense. Imagine having to ask a car how fast it’s going and waiting for a verbal response, rather than scanning a dashboard. Or asking a watch for the time instead of a quick look at its face. Even in these contrived, simplistic examples, what would be a split-second glance is already turned into multi-second back and forth.

The flexibility of natural language isn’t always a strength, either. The role of an interface isn’t simply to provide access to a system; it defines the shape of that access with clear structure and constraints. And the structure and constraints are placed intentionally, designed for the benefit of both the system and its users. Natural language can eliminate friction by removing menus and buttons, but in doing so, it can also strip away the direction and intentionality of each interaction. This undermines affordance instead of increasing it, and the result is a blank slate and the paralysis of infinite choice, rather than a tool that naturally guides users to focus on the task at hand.

This becomes even more pronounced as we move from simple task completion to complex management. Observability is key here, and the role of the UI shifts from direct control to oversight, as we delegate more work to AI at scale in this brave new world. When multiple AI agents are deployed to perform complex tasks in parallel, natural language with its infinite stream of words is a poor substitute for a purpose-built control center, designed to let us see how the whole machine is running at a glance and make precise interventions as needed.

Will UI survive AI? No, it will thrive as it rises to meet the new demands of AI.

Footnotes

1. I explore this idea further in “UI concerns are verticals”, where I argue that UI isn’t just a pretty skin but a core ingredient to make any system useful. ↩