Ditch the Chat
When defaulting to conversation undermines interaction design
The chatbot interface has become the default pattern in LLM-powered human-centered AI systems. It appears everywhere: from document analysis tools to creative writing assistants to data visualization platforms. But are we defaulting to conversation when we should be designing interaction?
The chatbot interface sends a clear signal: “This is just a thin layer on top of GPT.” During the initial gold rush of LLM applications, throwing a chat interface on top of an API call became the path of least resistance for researchers and designers alike.
But as HCI designers, surely we can do better?
Natural language has its place. When users face ill-defined problems or need to explore complex domains, conversation works brilliantly. A researcher trying to make sense of a literature review benefits from being able to ask “What are the main themes here?” or “How do these papers relate to embodied cognition?” The ambiguity and flexibility of natural language matches the ambiguity of the task.
I call this the natural language trap: the assumption that because LLMs excel at processing text, all user interactions should be mediated through text. This assumption marks a significant departure from foundational HCI principles, particularly Shneiderman’s direct manipulation paradigm that has shaped GUI design since the 1980s.
Consider a simple example: adjusting the brightness of a photograph. You could type (or say) “Make the colors brighter” and wait for the system to interpret your request. Or you could drag a slider. Which feels more direct, more predictable, more satisfying? The slider every time. It provides immediate feedback, precise control, and clear mapping between action and result.
The most compelling LLM-powered applications I’ve encountered recently don’t look like chatbots at all. They embed intelligence into familiar interaction patterns. A code editor that suggests refactoring as you type. A design tool that generates variations when you select an element. Textoshop, where users apply Photoshop-like direct manipulation tools to text, dragging and transforming words with visual feedback rather than describing edits in natural language.
These applications leverage LLMs’ reasoning capabilities while respecting users’ existing mental models. They feel magical not because they can chat, but because they anticipate needs and seamlessly integrate AI assistance into established workflows.
So why do designers keep defaulting to chat interfaces? I suspect it’s a confidence issue. Chat feels safe: users understand how to have conversations. After all, the whole point of the revolutionary ChatGPT interface was to chat with the LLM. Building novel interactions requires more courage. You have to make decisions about when to surface AI capabilities, how to represent uncertainty, and what level of control to expose.
But this safety comes at a cost. Chat interfaces create unnecessary friction for routine tasks and fail to leverage the full spectrum of human perceptual and motor capabilities. We have hands that can manipulate objects, eyes that can process visual relationships, and spatial reasoning that excels at understanding layouts and hierarchies. Why reduce all of this to typing?
The HCAI field has matured beyond the initial gold rush. We understand what LLMs can and cannot do. We have established patterns for handling uncertainty, managing latency, and providing appropriate feedback. Now we need the courage to move past the chatbot crutch and design interactions that leverage the full spectrum of human capabilities — our hands, eyes, and spatial reasoning — rather than reducing everything to typing.
As LLM applications mature, HCAI must move beyond proof-of-concept chatbots. We now possess both the technical infrastructure and the design vocabulary to build systems that integrate intelligence seamlessly, support multimodal workflows, and align with established HCI principles. In other words, it is time to ditch the chat and embrace interaction paradigms that leverage the full spectrum of human capabilities.
