The year AI learned to design (but not to care): A 2025 retrospective

In August 2025, Adam Wathan, co-founder of Tailwind CSS, posted an apology that racked up to a million views:

“I’d like to formally apologize for making every button in Tailwind UI bg-indigo-500 five years ago, leading to every AI-generated UI on earth also being indigo.”

It landed because it was funny. It stuck because it was true.

For much of the year, AI-generated interfaces started to feel eerily familiar: purple gradients, Inter fonts, rounded cards, subtle shadows, neat three-column grids. You didn’t need a watermark to spot them. You could feel it in your gut. Oh. This was made by AI.

And while the sameness was easy to joke about, it hinted at a deeper issue. Because the real problem with AI-generated design isn’t the color palette.

The problem wasn’t that AI-generated interfaces looked similar. It was that they behaved similarly, too.

It’s what happens when those interfaces meet real users.

As AI-generated UX moved from novelty to production in 2025, product teams learned a hard lesson: speed doesn’t guarantee quality.

Where we started: January 2025

At the beginning of the year, AI design tools were impressive, but limited.

Vercel’s v0 could spin up React components from prompts. Figma plugins could sketch wireframes. Large language models could generate UI code quickly enough to feel magical. But most outputs still felt like drafts: useful for exploration, risky for production.

Nielsen Norman Group’s assessment in early 2025 was blunt. They found that no design-specific AI tools were in serious use by professional UX designers. Outputs were fast, but generic. Context was thin. User needs were largely ignored.

The working assumption across the industry was reasonable: AI would speed up early stages of design. Humans would still handle judgment, nuance, and accountability.

That consensus didn’t last long.

The capability explosion: Spring through fall

By mid-2025, AI design capabilities didn’t just improve. They accelerated.

As tools matured, AI-generated interfaces moved quickly from experimental demos into real production environments.

Google’s Gemini 3 introduced “Generative UI,” the ability to create complete, interactive interfaces dynamically from prompts. Not wireframes. Not static mockups. Real interfaces, rendered on demand. In Google’s own research, human-designed interfaces only outperformed AI-generated ones 56 percent of the time.

That margin caught people off guard.

OpenAI’s GPT-5 followed with stronger frontend generation, better layout decisions, and improved visual consistency. GPT-5.1 added persistent context, making it possible to hold an ongoing design conversation across an entire workday.

Claude took a slightly different approach, emphasizing polish and restraint. Its frontend guidance explicitly warned against the “generic AI aesthetic,” calling out the same purple gradients and default fonts everyone had started to recognize.

Meanwhile, Chinese models like DeepSeek and Qwen reached parity in research-oriented tasks: personas, pain points, microcopy, and synthesis. Visual output lagged, but the UX reasoning was increasingly solid.

Then Figma announced Figma Make at Config 2025. Text prompts became interactive prototypes, Designers witnessed screens connected automatically and flows assembled in minutes instead of days.

By fall, the conversation shifted.

It was no longer Can AI generate interfaces? It was Should it?

Once AI-generated interfaces started shipping at scale, cracks that were easy to ignore in demos became impossible to miss.

The quality crisis no one wants to own

Here’s where things get uncomfortable for product leaders.

While AI-generated interfaces often look polished, they frequently fail at delivering reliable, accessible UX.

Accessibility breaks first

Accessibility is where AI-generated interfaces most clearly reveal their limitations.

Microsoft’s accessibility evaluations painted a stark picture. Under typical prompting conditions (the way most teams actually use these tools) the best-performing model passed WCAG standards only about a third of the time. Most failed far more often.

That means inaccessible navigation, broken keyboard support, incorrect roles, and missing ARIA attributes. These are problems users with disabilities encounter immediately and silently leave over.

And accessibility isn’t a niche concern. It’s a legal, ethical, and business requirement.

Alt text is shallow

AI descriptions routinely miss context. “A man at a podium” instead of “Brian Teeman speaking about accessibility at a Joomla conference.” Studies confirm what screen reader users already know: AI-generated alt text is often technically present but practically useless.

SEO is inconsistent

“Vibe-coded” sites frequently ship without proper title tags, header hierarchy, structured data, or crawlable content. JavaScript-heavy pages look great in demos and disappear in search.

Edge cases vanish

Loading states, empty states, error recovery are often missing entirely. AI optimizes for the happy path. Real users rarely live there.

Responsive design is an afterthought

Unless mobile is explicitly requested, interfaces break on smaller screens. AI doesn’t infer context. It responds to prompts.

This isn’t a tooling bug. It’s a training reality.

AI models learn from the internet, and the internet is not a curated collection of accessible, resilient, conversion-optimized interfaces. It’s a patchwork of shortcuts.

The slop trap: Why 2026 could get messy

At Standard Beagle, we’ve started calling this pattern the Slop Trap.

The Slop Trap is what happens when teams treat AI-generated interfaces as finished products and assume the UX will take care of itself.

It works like this:

• AI makes interface creation fast and cheap.
• Teams under pressure ship what looks “good enough.”
• UX review gets skipped.
• Users encounter friction—especially users with accessibility needs, older devices, or imperfect conditions.
• They leave.
• The product underperforms.
• And the failure gets blamed on the market instead of the experience.

The danger is subtle. AI-generated UX looks professional. It borrows the visual language of mature products. The cracks don’t show up until real users arrive.

And the timing is risky.

A significant portion of startups (and even enterprise teams) are now shipping products that are mostly AI-generated. “Vibe coding,” a term coined to describe accepting AI output without fully understanding it, has gone mainstream.

We’ve already seen what happens next:

• Products that can’t be debugged because no one understands the code
• Subscription logic that can be bypassed
• Security holes introduced faster than teams can detect them
• Users abandoning products after a single bad AI-driven experience

When interfaces are built without validation, accessibility checks, or UX review, the failure isn’t loud. It’s quiet. Users just don’t come back.

The data product leaders should pay attention to

There’s a telling gap in how different roles view AI.

In Figma’s 2025 AI report, developers reported high satisfaction with AI tools. Designers were notably less enthusiastic. More revealing: developers were nearly twice as likely to use AI for core work.

The people trained to evaluate usability, accessibility, and flow were the most skeptical.

At the same time, many AI tools are built by teams with little UX representation. These systems are judged primarily on speed and capability, not on whether they support real human behavior.

That imbalance matters.

Where AI actually shines

This isn’t an anti-AI argument. When AI augments human expertise, the results are compelling:

• Faster experimentation
• Better insight synthesis
• Improved conversion optimization
• Increased engineering throughput
  

The pattern is consistent. AI works best when it:

• Optimizes existing designs
• Generates variations for testing
• Surfaces insights from data
• Accelerates workflows humans still control

It struggles when it’s treated as a replacement for judgment.

As Jakob Nielsen has pointed out, usability principles haven’t changed. AI just violates them faster when left unchecked.

What product leaders should do now

For product leaders heading into 2026, the biggest shift is recognizing that AI-generated interfaces require more UX oversight, not less.

If you’re building products in 2026, a few guardrails matter more than ever:

1. Treat AI-generated interfaces as drafts, not deliverables
   Speed is valuable. Accountability still matters.
   
2. Bake accessibility testing into your pipeline
   Automated checks won’t catch everything, but they catch enough to prevent obvious failures.
   
3. Prompt explicitly for edge cases
   If you don’t ask for error states, loading behavior, or mobile views, you won’t get them.
   
4. Increase UX oversight as AI usage increases
   AI productivity gains should fund more review, not less.
   
5. Watch real user behavior closely
   Demos lie. Funnels don’t.
   
6. Resist “vibe coding” as a culture
   Design you haven’t validated will eventually fail, usually quietly.

Looking ahead

2025 was the year AI learned to generate interfaces. 2026 will be the year we find out whether those interfaces hold up in the real world.

The technology is impressive. The productivity gains are real. But capability without care is a liability.

The teams that succeed won’t reject AI-generated interfaces — they’ll invest in the UX discipline required to make them work for real users.

At Standard Beagle, we believe the teams that succeed won’t be the ones who ship the fastest demos. They’ll be the ones who pair AI’s speed with human judgment, UX discipline, and respect for real users.

We help product teams use AI-generated interfaces responsibly, ensuring the UX actually works for real users.

The slop wave is coming.

The question is whether you’re building on solid ground or shipping just fast enough not to notice the cracks.