Why Chat is the Wrong Interface for AI

New technologies are inherently difficult to understand. When a new computing paradigm emerges, we struggle to conceptualize it, reaching instead for familiar metaphors and frameworks that are no longer relevant. We lack the vocabulary and comprehension to reason in the terms that novel developments demand. As a result, the initial interfaces tend to reflect what is most perceptible and only later do we discover the deeper potential hidden in more powerful forms of interaction.

The earliest interfaces for new computational capabilities are almost always literal. The first computers were operated with switches and punch cards because the hardware understood discrete physical states. Early mobile devices resembled personal computers and popular skeuomorphic design perpetuated the need for recognizable parallels. The initial interfaces exposed the most direct conduit to the underlying technology. The goal was to demonstrate clarity rather than to consider what entirely new forms of behavior could exist as a result of the technological breakthroughs.

The rapid and continual advancement of large language models have exacerbated these conditions. Artificial intelligence has collapsed the intermediary stages of technological adoption, leaping directly from research labs to exponential usage. Large language models reached hundreds of millions of users before we developed a coherent model for how they should be integrated into consumer tools. Model capabilities advanced faster than our ability to reorganize software around them, creating a widening gap between what these systems can do and how they are allowed to behave. This gap is most visible in the interface. Open any leading AI product and you’re met with a text field awaiting instruction. Because text is the native form for both input and output, the chat box is a natural approach to interaction, exposing the model’s internal representation directly to the user. But, the most direct reflection of a technology’s internal state is not, in itself, an ideal user interface and, with extraordinary concentrations of talent and capital in the space, this is an awfully somber outcome. The most consequential general-purpose technology in decades, mediated almost exclusively through a chat box.

This issue goes beyond a considerable lack of creativity. Conversation is a fragile medium for sustained cognition. Chat is likely the least powerful or compelling interface for LLMs, especially in the long term. It is linear, transient, and poorly suited to preserving structure. In a transcript, context exists as an accumulation of past turns rather than as an externalized representation of the problem itself. Each new message partially degrades the context that preceded it. Memory continues to improve within the models, but I still find myself hunting for certain responses, restating prompts, and continually reconstructing the conversation. Chats are a medium for exchange. Not all tasks are malleable enough to be forced through a single linear channel.

Prompt engineering is a subtle admission of failure that shifts the burden of system design to the end user. If your consumer product leads with tooltips and tutorials–let alone creates a new discipline for users to master–the product isn’t doing the work it should be. This is the inverse of what intelligent tools are meant to do and a consequence of an intelligence that cannot yet live alongside the user.

Chat-based interfaces have stuck around because they are highly compatible with systems that lack persistent memory, continuous observation, and authority to act autonomously. The intelligence is episodic and externally hosted so the interface is optimized for exchange rather than execution. As long as intelligence must be explicitly invoked, interaction must remain turn-based.

There is also a misunderstanding in the expectation of how consumers want to interact with AI. There’s a belief that making tasks like coding effortless will usher in a new era where everyone builds bespoke software on demand. People don’t want to chat with their tools and data. They want software that does the heavy lifting for them, clearly and reliably. They want products that were refined and toiled over by developers who love software.

I don’t really want to build the apps I actually use. I don’t want to have to think about how they look or the perfect prompt or the ideal user flow. I don’t even really want to think about whether or not they use AI in how they deliver me the experiences I desire. Pride in my own creation isn’t worth sub par user experience. Call me lazy but despite the incredible advances in what is possible, I still want the apps that other people build to solve my many problems…I want it to exist for me and for it to feel perfect.

I’m certain that the convergence towards a chat-based interface was not a consequence of apathy–the technology was simply advancing too fast amid fierce competition, leaving no time for discussions of this sort. Now that leading products feel comparable, the next outsized winner won’t just be technically impressive but insightful, perceptive, creative, sentimental, and deeply embedded in daily life.

When intelligence lives in the cloud, interactions have to be explicit. Latency, privacy, and cost force everything into short, discrete exchanges. A different class of interface becomes possible when intelligence moves closer to the end user both logically and physically. If intelligence can live on the device, the interface dissolves and the requirement to explicitly invoke action begins to disappear.

There are several developments that must take place for on-device compute to reach consumers at a significant scale. Primarily, there needs to exist one or many ultra-efficient mobile optimized models. They need to be sufficiently compact and energy-efficient to operate within the tight power budgets of a mobile device while managing cooling and battery drain. In addition, memory systems must maintain a long-lived state instead of a more transient context window, supporting gradual accumulation, selective discarding of information, and stable identity over time.

Most importantly, the operating system itself must evolve to anticipate user needs, generate dynamic views, and orchestrate tasks. Static, reactive systems can't deliver a truly native AI experience; without this architectural shift, AI remains siloed in apps, unable to transform the device into an intelligent proxy. Intelligence must sit above the application layer to maintain shared context and exercise agency.

An operating system designed around these assumptions would not present intelligence as a destination and have no primary surface to open and address. Rather than scheduling follow-ups and requesting information the system would infer and execute actions on its own.

Apple is often ridiculed for not being a first mover in breakthrough categories, especially as they lag behind in the race to AGI. However, historically, they catch up quickly and powerfully given their complete ownership of the consumer device stack. I think they’re the obvious long-term winner in the AI space. If they are able to release an AI-native OS for the iPhone, current phones will look like landlines and all preceding AI consumer products will feel archaic. Unfortunately, there seems to be limited appetite for a net new device, but a non-Apple player could still capture enormous value by controlling an OS-adjacent orchestration layer or positioning themselves to be acquired into Apple’s distribution network.

The next generation of consumer AI will be defined by how little users need to think about interacting with intelligence. Today’s mobile devices are fundamentally reactive, assuming that computation must wait for instruction. An inversion of that assumption means that an AI-native device can treat patterns and behaviors as inputs rather than waiting for discrete prompts. It will be difficult to imagine returning to conversation and prompting as a dominant interface. We are only at the beginning of understanding what AI interfaces can be.