The screen-building job is dying. UX isn't.

OpenAI says ChatGPT now reaches 1.2 billion people a week.

To me, the defining interface is a single text box: people describe what they want, and the software works out the steps.

I think this shift will remove a big share of today's UX engineering work, specifically the part that turns a mockup into hand-built screens. It will not remove the need for people who understand what users mean. Two forces drive the first claim, and the evidence against the stronger "UX is dead" version is just as concrete.

Force one: the interface becomes a description of what you want

On June 18, 2023, Jakob Nielsen called generative AI the first new UI paradigm in 60 years: "users tell the computer what they want, not how to do it."

Most interface work is translation. A booking form, a report filter, a settings page, an admin panel and an onboarding wizard each turn something a person wants ("show me last quarter's churned customers in Europe") into clicks the software understands. Someone designed every dropdown, built every state, wrote every validation message and fixed every layout bug on a small phone screen.

When software can take the request directly, much of that translation layer is no longer needed, and neither is the work of building it.

Satya Nadella has made a version of this argument about business software: he expects business applications to "collapse" in the agent era, because their business logic is moving into agents.

Force two: users get the screen they need, when they need it

When a person does need a visual interface, they can increasingly have one generated.

• Anthropic says users have created over half a billion artifacts in Claude, from productivity tools to educational games.
• Lovable says people have created more than 60 million projects on it since its November 2024 launch.
• Vercel says more than 4 million people have used v0.
• In the first quarter of 2026, about 60% of Figma's paid customers with more than $100,000 in annual recurring revenue used Figma Make in the quarter's busiest week.
• Google is rolling out generative UI in the Gemini app: its dynamic view designs and codes a custom interactive response for each prompt.

Together, these forces make the traditional pipeline look expensive for screens that don't last: a designer mocks it up, a UX engineer builds it, a team maintains it for years. The one-off internal screen goes first: the dashboard for one meeting, the admin tool for one workflow, the form that only needed to exist for a quarter.

Typesetting is the closest precedent I know.

Between 1988 and 2000, US employment of precision compositors and typesetters fell by roughly half. The BLS table doesn't say how much of that decline software caused.

My reading of that period: page layout moved onto computers, and the task of arranging type survived by moving to different people using different tools.

Where the "UX is dead" argument breaks

Chat is bad at many tasks. Amelia Wattenberger argues that a chat box gives no hint of what the software can do, puts the burden of supplying context on users, and makes it hard work to see what changed from one answer to the next.

Studying image-generation prompts, Nielsen Norman Group named a related problem the articulation barrier: users struggle to put what they want into words, and interfaces that pair a prompt with GUI controls can reduce that load.

I don't want to edit a spreadsheet, trim a video or pick an airline seat by describing it in prose, and I doubt most people do.

Interfaces are moving into chat. OpenAI launched apps inside ChatGPT with an Apps SDK for developers; the apps blend familiar interactive elements, like maps, playlists and presentations, into the conversation.

MCP Apps then went live as the first official extension of the Model Context Protocol, building on OpenAI's Apps SDK: tools can return dashboards, forms and visualizations that render inside the conversation.

Developers still build those interfaces, now for a chat window.

Most people take the defaults. When Jared Spool's team analyzed Microsoft Word settings files sent in by volunteers, fewer than 5% had changed any of the 150-plus settings. Programmers and designers were the exception.

The study is old, so treat it as a hint rather than a measurement of today's users. Still, "everyone becomes their own UX designer" runs into the likelihood that most people don't want that job, and someone still has to make the defaults good.

The hiring data is mixed. Indeed's index of US software-development job postings stood about 22% below its February 1, 2020 baseline on October 9, 2026.

• In UXPA's 2024 salary survey, 35% of respondents reported losing staff in the previous year, and the same share reported adding staff.
• But in a July 8, 2026 report, Indeed said software-development postings had risen almost 15% since February 24, 2025, with 71% of the increase from May 2025 to May 2026 coming from senior roles.
• The BLS still projects 5% growth for web developers and digital designers from 2025 to 2035.
• Indeed also notes that postings for the occupations most exposed to AI began falling before ChatGPT was released.

What's dying, and what replaces it

Here is my prediction, sized to that evidence.

Dying: the job of hand-turning a mockup into screens. Pixel-matching specs, building yet another CRUD form, wiring up one more filter, maintaining one-off internal tools. In my estimate that is a big share of many UX engineers' weeks, and AI already does much of it well enough.

Growing: the work one level up.

Designing the parts agents assemble from

In Vercel's AI SDK, the model decides when to call a tool, and its result is passed to a React component a developer wrote.

MCP Apps render UI resources that the server ships as bundled HTML and JavaScript.

As I read it, the agent in those systems composes the screen from parts a person made, and the parts worth designing well are components, design tokens, constraints and accessibility rules.

Google's approach generates the interface code on the fly.

In my view, that makes the rules and evaluations around it matter more.

Designing how the agent behaves

When it should ask before acting, how it shows what it is about to do, how a user undoes it.

Andrej Karpathy calls one version of this the autonomy slider.

It is interaction design, and in my view few products do it well yet.

Judging generated UI

If agents produce screens at scale, someone has to define what good means and build the evaluations that catch bad ones.

High-craft canvases

Design tools, spreadsheets, editors, games, cockpits.

Even in Nadella's agent-first picture, Word and Excel stay specialized canvases.

Google's own description of a UX engineer goes beyond building screens: functional prototypes, feasibility research, translating between design and engineering.

That in-between position becomes more valuable when one side of it is an agent.

If you're a UX engineer

It's hard to read headlines saying your job is obsolete, especially if the postings you'd apply to look thinner than they used to. In my view, the skills that made you good (understanding what users mean, knowing why a flow fails, caring about the state nobody designed for) are the ones the agent era is short of.

The old job was building the screen. The new one is building the system that builds screens, and deciding what it may show.

I'd bet on the engineers who move up to that level; competing with the generator on producing screens is a losing race.