Why chatbots are not the future of interfaces for LLMs: the best prompts are not obvious, people will get sick of typing, responses are isolated, and more
Unfortunately for the countless hapless people I've talked to in the past few months, this was inexorable.
Context & Ripple Effects
Wattenberger's argument lands on ground already broken once before: the Messenger and Kik chatbot boom of 2017 collapsed for the same structural reasons she now attributes to chat-based LLM interfaces — text boxes put all the work on the user. The 2022-23 ChatGPT wave restarted the pattern, with hallucination warnings from experts arriving almost alongside the hype, and ChatGPT confidently failing at basic arithmetic showing how little a conversational surface does to surface uncertainty.
Her critique matters because the chat box is the default product shape of the current LLM boom, and companies are already shipping it into production: marketing and social media writers are being replaced by these tools even as quality drops. The question she forces is whether that deployment is built on the wrong interface, not the wrong model.
First-order effects
- Developers building chat-wrapper products inherit her critique directly: their interface pushes prompt-writing skill onto users, and the conciseness study showing that terse answers worsen hallucinations suggests the burden gets worse the more users try to control output.
Second-order effects
- Toolmakers competing for LLM users are pushed toward structured, GUI-based surfaces that hide prompting — a response that also mitigates the freeform-chat failure modes, since constrained inputs give the model less room to confidently invent, as it did on arithmetic questions.
Third-order effects
- If the pattern holds, the LLM market splits between raw chat as a power-user surface and embedded, workflow-shaped interfaces as the mass-market default — a repeat of the 2017 chatbot correction, but this time with models capable enough that the interface, not the intelligence, is the bottleneck.
The trend: LLM products are repeating the 2017 chatbot cycle, with interface design — not model capability — emerging as the deciding factor between chat boxes and workflow-embedded AI.