UX Collective: Designing for AI means designing like it's 1999

Patrick Neeman’s essay for UX Collective draws a pointed comparison between designing AI interfaces in 2026 and building for the web in 1999. At both moments, capability outran convention: in 1999, pages were handmade in Notepad, JavaScript was unreliable, and nobody had agreed on what a navigation menu should look like. Today, designers face protocols barely a year old, models that rewrite their own capabilities every few months, and no settled vocabulary for the patterns they are inventing.

The article’s central observation is that designers now have “extraordinary capability and almost no shared conventions for handing it to people.” Research shows frontier models are doubling their task-completion abilities roughly every seven months. A design built around what a model can do today may be obsolete before the next planning cycle. The Model Context Protocol, released by Anthropic in late 2024, was adopted by ChatGPT, Cursor, Gemini, and Microsoft Copilot within a year—before Anthropic handed it to the Linux Foundation in December 2025. Speed like that makes stable design patterns hard to establish.

Neeman is not writing a crisis piece. He frames the absent conventions as an invitation: designers who worked through the early web built the foundational patterns everyone relies on now. The same opportunity is open today, and the skills involved—observing users, identifying failure modes, proposing shared language—are the same skills designers already have.

The essay also calls out a specific problem with AI-era design work: because models evolve underneath an interface, interaction patterns can degrade silently. Unlike a bug that breaks a button, a capability shift makes an interaction feel vaguely worse without an obvious cause. Designers working with AI features need to account for this drift in a way that static UI work does not require.

Who this is useful for. Designers on product teams adding AI features, design leads trying to explain to stakeholders why AI interface work is harder than it looks, and anyone building design systems that include AI-generated or AI-augmented components. The essay gives a useful historical frame for conversations about pace and uncertainty that might otherwise feel unmoored.