Stripe's Katie Dill: AI is building "zombie UI," generic screens nobody checks

AI will draw you a screen in ten seconds. It will match your system, add a checkmark, maybe some confetti. What it will not do is know whether that moment deserved confetti. That gap, between a screen that looks right and one that works, is where the fight is now. Let me catch you up.

The deep cut

• AI produces output, not outcomes. Bubeck's "I don't really care" if the model understands is the shrug that trades working products for fast ones.
• A good look is the floor, not the proof. Flutter's hand-drawn iOS bar froze on the old design while Apple's Liquid Glass moved on without it.
• Autonomy needs a visible stop. Zauber flags a mismatched shipment volume for a human instead of guessing and booking it.

The ten-second screen that nobody owns

Sébastien Bubeck of OpenAI said the loud part out loud in a March 2025 debate: ask a model to build your app, come back the next day, "boom, the app is there." Did it understand what it built? "In a way, I don't really care."

That shrug is the problem. Once you stop caring whether the thing understands, you measure success by output, the screens and flows produced, not by outcome, whether a person can actually finish the task. A proof of concept shows something can exist. It does not show it works.

Katie Dill of Stripe has a name for what this produces at scale. In a talk flagged on Reddit, she compares it to the rushed buildings after World War II: "zombie buildings," generic and interchangeable, showing no care for the people inside. We are in another building boom, and the output is zombie UI.

Native moves, imitation freezes

Here is the part that costs money. In June 2025 Apple shipped Liquid Glass, and its WWDC session put it plainly: apps get the new look "as soon as you recompile." Bars, sheets, controls, all updated without a designer touching them. Apps that drew their own controls got nothing, and that lag was still visible by November 2026.

AI makes this worse, not better. A polished look now costs one prompt, so apparent quality is handed to everyone. But generated screens learn from what already filled the web, which means the old iOS. Ask a model for an iPhone app and you get last year's full-width tab bar, not the floating iOS 26 one. It is the counterfeit copying last season while the house has moved on.

A spring study of 138 generated mobile screens found the same crack: the models nailed navigation conventions but the skill was "pattern-driven rather than principle-driven." Help, error prevention, and error recovery were often missing. The picture of native is not native.

The honest tester that cannot guess

Your users hide your bad design. They try both buttons, wave the mouse to find the hover menu, log back in after a timeout, ask a coworker. Agents cannot do any of that. A designer at Deck watched an agent hit a form with two buttons labeled "Continue" and "Submit," pick wrong, and stall completely.

That makes agents the most honest usability testers you have. The places they break are the places people break, minus the quiet suffering. Ambiguous labels, state that only shows on hover, interruptions that slide over the submit button, success that never announces itself. If an overlay can stop a machine cold, ask what it costs the tired human next to it.

This is the curb-cut effect. The fixes that help an agent, real text labels, visible state, no hover-only content, are the same ones that help people on screen readers. For LLM chat itself, Eric Bailey shows the parallel work: a dedicated live region so a screen reader user hears a reply land, a loading announcement every five seconds so the thing does not feel crashed. Access is not an add-on here. It is the floor.

Where autonomy has to stop

In September 2025, researchers ran the same prompt through a model 1,000 times at temperature zero and got 80 different completions. The model is probabilistic by design. Your product cannot be, not when the output turns into a freight booking or a medical record upload.

The gap between using AI and trusting it to run things is real. The 2026 Stanford AI Index found 88% of organizations used AI in 2025, but scaled agent deployment stayed in the single digits across nearly every function. Drafting an email and sending it to 4,000 customers are different acts.

The better products decide where uncertainty is allowed and where it is not. Routine Labs uses AI to build healthcare automations but runs production on logged, deterministic workflows, no model in the live loop. Zauber lets AI read messy freight emails, then when a declared volume does not match the packing list, flags the discrepancy and waits for a human to confirm. The intelligence handles ambiguity. The stop button handles consequence.

Three questions for your team

• In our last AI-built flow, what did we actually ship, a working outcome or a proof of concept that still needs the real UX work? Name who owns that second half.
• Which of our controls that look like system components actually behave like them, including accessibility and text scaling? Pick one screen and check this week.
• Where in our product does an AI decision turn into an action with real cost, and who or what is the stop button at that exact point? If the answer is "the model decides," that is the review item.