Key Takeaways

  • Stripe Head of Design Katie Dill compares today's rapid software generation to the post-WWII construction boom, where copying modernist blueprints created thousands of soulless, generic structures.
  • Language models output the most statistically probable patterns based on past training data, which means they excel at copying yesterday's design trends rather than creating original, context-specific solutions.
  • Fast AI prototyping gives product teams a false sense of completion, confusing a working mockup with finished, user-ready product craft.
  • Because AI makes generating interfaces cheap and instant, teams treat new features as disposable and fail to plan for long-term maintenance.

The Post-War Trap

Right after World War II, a massive construction boom swept across the United States. Returning soldiers needed homes, city centers needed offices, and new techniques made building faster than ever before. Builders copied modernist architecture at scale. But as Dill pointed out at the Lenny and Friends Summit, the builders copied the exterior look while stripping away the original architectural purpose.

“The thinking got thinner and thinner, and all that was left were generic patterns ill suited to the context at hand,” Dill said. The result was a wave of bland, unlivable structures: zombie buildings.

Software development is running straight into the exact same trap. LLMs make generating screens, buttons, and user flows frictionless. Anyone can prompt a model and get a functional interface in thirty seconds. But speed removes the friction that once forced engineers and designers to think through context, user constraints, and trade-offs. “If we're not careful, our building boom could end up a little too much like the post-war building boom,” Dill warned. “The proliferation of patterns ill suited to the context at hand, essentially zombie UI.”

Three Blind Spots of Model-Generated Interfaces

To prevent your product from turning into a collection of generic templates, you need to understand the mechanics of why AI defaults to mediocrity.

Dill points to three specific traps founders face when building with models:

First, statistical regression. “First, LLMs are really good at telling you the most probable answer, which essentially means that they're able to tell you what has been or is popular now, what has worked, what was in style,” Dill explained. “They're less good at telling you what's original or specific to you, your brand, and your users context.” An LLM draws from the median of the web. It will give you a standard SaaS dashboard with a standard sidebar because that is what exists in its weights. If you accept its first output, your app looks like every other product launched that month.

Second, the illusion of finished work. A generated screen looks complete at a glance. It has buttons, tidy padding, and plausible placeholder copy. But a pretty interface is not product strategy. It does not address edge cases, handle network latency, or answer whether the user actually needs that screen in the first place.

Third, disposable software without maintenance. “And the third watch point, this work is so easy to do, so quick that it often feels disposable and worst is treated that way,” Dill said. “We sometimes let the responsibility of our decisions fall by the wayside and don't really think about the long term, like who's maintaining this.” When creating software requires zero effort, teams ship features without thinking about who owns the technical debt or who will support it two years from now.

What to Do With This

Audit the last three user-facing screens your team built with AI assistance. Open each screen and strip away the styling; look only at the user workflow and the specific problem it solves. If the screen uses a generic card or dashboard pattern that could fit into any competing product without modification, scrap the AI draft and rewrite the user requirement from scratch based on the exact context of your customer.