The attack of the giant AI design slop monster: a survivor’s guide

In 2026, AI has transformed UX design in unpredictable and scary ways

A lone person faces a colossal orange monster with a dripping, toothy grin in a dark, fiery landscape, representing the overwhelming AI design slop monster.
The attack of the giant AI design slop monster: a survivor’s guide. (Image credit: me + Midjourney)

In September this year, I was once again speaking at City St George’s, University of London annual conference on Human-Computer Interaction Design, entitled “Rage Against the Machine”. This story expands upon my talk at that conference.

(Some images that accompany this story are AI-generated, including the dummy product example — as stated in the image descriptions. That’s kind of the whole point. The text, however, is 100% written with my own bare hands.)

AI for UX design (a dark twist)

Re-reading my writing on the subject of AI in UX from last year, I laugh at how outdated it appears today. While every prediction I had made last year has come true, we’re well into a territory I could not have anticipated in my wildest dreams.

An oyster holds the red monster instead of a pearl. Tick marks confirm predictions about experimentation, industry trends, custom agents, AI for UX design, fully agentic AI, Jobs-to-Be-Done analysis, AI design tools and an AI research assistant. “The world is your oyster” becomes “The world is not your oyster”.
My predictions from last year all came true, but with a dark twist. (Image credit: me, Midjourney)

Research-wise, most methods I used to rely on are not needed anymore: while robust human oversight and human interpretation remain crucial, structured prompting has given way to the model’s own intent inference; hallucinated user quotes are a thing of the past; and with correct templates, creating research reports is a breeze.

So instead, today I’d like to focus on AI for UX design. Last year, I still dismissed AI for design — it was sloppy and horrible and could not compete with me, a seasoned professional. This year… well. I went on a journey that took some dark turns. Thankfully, I emerged from it changed but victorious. For now.

Let’s start with some hypothetical scenarios.

A Product Manager, a developer and a UX designer walk into a bar. The barman asks, “So which one of you vibe codes, then?”

Three portraits from The Good, the Bad and the Ugly introduce the protagonists, from left to right: the product manager, the developer and the designer, labelled “you”.
Our three protagonists. (Image credit: me + The Good, the Bad and the Ugly)

(Disclaimer: no product managers, developers or designers were harmed in the making of this episode. Any resemblance to actual persons, living or dead, is purely coincidental.)

Dubbed the word of the year by the Collins English Dictionary in 2025, vibe coding is, simply put, “the use of artificial intelligence prompted by natural language to assist with the writing of computer code.”

Scenario 1: The PM vibe codes

A cowboy portrait represents the product manager, above the words “It’s just an idea…”
The PM vibe codes. (Image credit: me + The Good, the Bad and the Ugly)

Your product manager has discovered Claude Code and, by giving it a couple of prompts, has turned every single idea from their backlog into a complete UI interface. They dump the lot on you during your weekly Teams call, prefaced with, “It’s just an idea. Don’t worry, you can draw it up properly later.”

Or worse, they don’t come to you at all but instead go straight to the senior leadership and table their “designs” — and the leadership is ecstatic. Finally, they don’t have to wait for weeks for you to draw things in Figma, or get it through your lengthy design systems approvals, or anything! Let’s go develop it tomorrow! In fact, it’s already developed, isn’t it, it’s a complete functional prototype, let’s ship it now!

And let’s say, the PM’s “design” looks like this — but no-one except you thinks it’s a problem:

Fictional, badly designed legal research interface. Its alert configuration page contains a dense three-column form, numerous required fields, practice-area checkboxes, several navigation menus and competing Save, Submit and Apply buttons.
A dummy example of what a PM might generate and believe that it looks good. (Image credit: me + Claude)

You’ve lost control. What are you supposed to do now? Tidy up the PM’s slop, argue the toss, complain to the leadership?

Scenario 2: the developer vibe codes

A stern cowboy portrait represents the developer, above the words “We are mandated to use AI-assisted coding…”
The dev vibe codes. (Image credit: me + The Good, the Bad and the Ugly)

Your dev team has been told by the leadership to use only AI-assisted code and they’ve embraced this brave new world. You might think: and so what? The developer codes, you don’t code. You don’t care what tools they use. Until they send you the cert environment link and instead of your beautiful Figma vision you see this:

Dummy legal case-alerts interface with annotations flagging wrong font sizes and colours, incorrect styling, missing focus and hover states, components disconnected from the project UI library, incorrect or missing keyboard tab order, and incorrect responsive behaviour.
A dummy example of what a developer might vibe code and believe that it looks good. (Image credit: me + Claude)

“Why did it end up looking like this?” you ask your developer. Then you discover that they wholesale fed your Figma into Claude Code via MCP without any links to your design system library or any proper componetisation. Claude hallucinated some Frankensteinian approximations of your design, but if you squint, it sort of looks OK. Just don’t try to use the keyboard to check for focus states.

The developer and the business are ecstatic because it took them 3 days to produce that instead of the usual 3 months. Once again, the only one that sees the problem, is you.

The red AI design slop monster towers over a lone person in a fiery landscape. Text reads: “Your existential crisis. Is UX even needed anymore?”
Your existential crisis monster. (Image credit: me + Midjourney)

Your existential crisis monster has landed. The monster is asking you, “Is UX even needed anymore? If you leave tomorrow, who even cares?”

Let’s look at some numbers.

The slide reports that 84% of developers use AI for coding. Of product managers’ AI tasks, 19.8% involve prototyping. Of designers’ AI tasks, 16.5% involve ideation and 13.2% involve prototyping.
Who vibe codes in 2026. (Image credit: me. Sources: https://www.hostinger.com/blog/vibe-coding-statistics/, https://www.lennysnewsletter.com/p/ai-tools-are-overdelivering-results)

This year, the tech industry has been shaken by AI like no other industry on Earth. 84% developers use AI for coding; product managers use it for absolutely everything, from brainstorming to senior leadership presentations. And PMs frequently trample the territory previously occupied by UX researchers and designers — because they can. For them, it’s now easier than ever, a mere prompt away.

UX designers, on the other hand, are behind. Perfectionism prevented us from experimenting with early AI design tools because these were so terrible, barely good enough for ideation. But UX designers today no longer have the luxury of pixel pushing — everyone else is ahead. Meanwhile, AI design tools are improving by the day.

Scenario 3: you fix what you can — and vibe code your own slop

A grinning cowboy represents the designer. Text reads: “You are the only one who cares about UX/UI.”
You vibe code. (Image credit: me + The Good, the Bad and the Ugly)

So, as the only one who cares about UX/UI, you’re going to fix it. How?

1. Fix things with the PM

Three icons illustrate the PM-related challenges: a person assembling a light-bulb puzzle for problem definition, a robot producing images for AI-generated design, and a snail beneath a speedometer for pressure to work faster.
Solutionising, AI slop and speed are all factors you may need to fix with the PM. (Image credit: me + Claude)

PROBLEM: solutionising — “It’s just an idea”.
FIX: revert back to user problem.

Make your PM state what user problems they are solving, help them determine what user evidence there is already. Fill the gaps with new user research.

PROBLEM: AI design slop.
FIX: “skeleton UI”.

Define the minimum UI rules to AI-ideate within. Everyone must follow them — whether PM or UX.

PROBLEM: you’re too slow.
FIX: make your own AI slop, fast.

You can’t win against a machine gun with a bow & arrow, so get armed with the same tools and slop away, knowing that your slop will be better than the PM’s slop.

Two dummy legal research interfaces side by side. The left shows a crowded alert configuration form, labelled “The PM can’t spot AI slop”. The right shows a case-alerts inbox grouped by matter, with consistent “Keep for matter”, “Not relevant” and “Share” actions, labelled “But you can”.
Dummy example: Even when you create AI slop, you have the designer’s eye to make it better than the PM’s. (Image credit: me + Claude)

You have the designer’s eye. You also have the researcher’s eye. While the PM may not have thought of how user research is represented, how to avoid pointless or weird UI, how to ensure visual consistency and design system rigour, you will have thought of all these things. As you vibe code like your life depends on it, knowing that your sloppy design will be a functional approximation to the perfect Platonic form you have in your head.

Feed your Claude app the body of your user research, feed it your “skeleton UI”, describe what to build in detail, link it to your design system (more on this later), give it the required accessibility standards to follow, ask it to generate realistic dummy content (which you should check with a human subject-matter expert), provide it with “done when” criteria. Press the button. Get yourself a coffee while Claude does its thing.

2. Fix things with the developers

Three icons represent deteriorating quality, quality assurance and pressure for speed: a cracked quality seal, a magnifying glass over a checklist, and a rock singer raising a fist beside a microphone stand.
Drop in quality, low frontend knowledge, and speed, could be the issues you need to fix with your developer. (Image credit: me + Claude)

PROBLEM: frontend quality deteriorates due to AI-assisted coding.
FIX: define your “minimum bar”.

Decide what your lowest acceptable UI quality looks like, spell out WCAG compliance, define componetisation rules the developer should follow.

PROBLEM: developer has low frontend knowledge and can’t spot AI mishaps.
FIX: institute QA checks of their work.

If any “minumum bar” items don’t make it to product release, insist on cleanup afterwards.

PROBLEM: “Don’t Stop Me Now” — the developer just can’t get enough of AI speed.
FIX: work fast and be prepared to compromise.

Make your QA checks rapid and responsive. You won’t win by being a perfectionist — pick your battles.

How to keep the AI design slop monster under control

Four-column AI tool comparison. Figma Make and Lovable are labelled “downstream”; ChatGPT Desktop and Claude Code with Claude Design are labelled “foundation”. Each column lists the tool’s stated strengths.
The various AI design tools out there. (Image credit: me)

So you’ve managed to stave off your existential crises with your team and you’re now happily AI-designing/vibe coding. Let’s take a look at how that can be done better.

I’ve used a handful of different AI design tools, many others are out there, but I suspect the principles of working with all of them are much the same.

Three guardrails for AI design: provide UX research findings; provide a design system and UI skeleton patterns; and use an organisation-approved, licensed tool while confining each desktop project to a single folder. Icons show a person with a magnifying glass, a drawing tool, and a shield with a padlock.
The principles of AI design. (Image credit: me + Claude)

Principle 1. Create your user research guardrails.

I cannot overstate this: always start from the user. Feed the AI tool the research findings (best as structured reports rather than raw data, as these are more machine readable). Give the AI tool the full context and test its understanding of the context before progressing.

Principle 2. Create your design guardrails.

Without these rigorous guardrails, your AI tool will ride you instead of you riding it. Fixing AI design errors can be frustrating and can quickly eat into your token allowance for the month.

Claude conversation about correcting a prototype’s buttons. The user requests “Keep”, “Dismiss” and a borderless “Share” button. Claude reports applying filled, outlined and ghost styles respectively, then checking the disabled Share state and keyboard interaction.
An example of an AI design error that took effort to fix. (Image credit: me + Claude)

This is what I meant earlier when talking about “setting up skeleton UI”. It includes components but also templates, patterns, interaction rules, accessibility — the whole UX design microcosm.

These guardrails can be set at multiple levels of sophistication, starting from simply giving the AI tool a URL link of a design system. The next step could be Figma MCP (if using Figma as a base) to get your Figma specs and your AI tool to talk to each other. Then, Code Connect can help link your Figma components with your company’s GitHub component library, preventing the AI tool from hallucinating components from scratch. You might jump between Figma and an AI-generated prototype before arriving at your perfect design solution.

The advantage of AI prototyping is that it is vastly faster than traditional design, allowing you to explore multiple options before you’ve finished that coffee of yours. Another great advantage is that it can be made to behave like a live product — brilliant for user testing.

Diagram shows Figma feeding an AI tool through Code Connect and Figma MCP, while GitHub also feeds the tool. Code Connect links design components to production code, MCP provides design structure, and GitHub provides the team’s production-ready components.
The various connectors that will allow you to cut down on AI design slop. (Image credit: me)

Principle 3. Respect safety and security.

Safety and security should not be an afterthought. AI tools, especially ones capable of accessing your computer and your company’s systems (such as Claude and Codex apps), are increasingly powerful and thus dangerous. From scandalous incidents like the agentic hacking of Hugging Face, to more mundane ones where Claude Code wiped 116GB of personal data — these stories abound.

You must at least do this: when giving the AI tool access to your files, confine it to one local folder only; keep backups of your precious data in locations that AI has no access to; monitor what permissions you are giving to the tool during each session.

This might be a no-brainer in 2026, but be aware that free and low-tier personal subscriptions allow AI tools to scrape your or your company’s data. Only share IP with the tool that you know is safe — for example, because your company has an enterprise subscription to it.

What happens next?

There are already AI-assisted UX design methods that bypass Figma altogether. If your component and pattern library is mature, once you connect it directly to Claude (for example, via a prototype environment linked to GitHub), in theory, there’s nothing stopping you from just designing in Claude. Make your prototype shareable online via Claude Artifacts (as of this month, it replaces Claude Design) and your frontend engineer might even be able to reuse your AI-generated code.

It might even be that over time, boundaries between UX design and frontend will blur — some UX designers already contribute to frontend code alongside developers, via AI tools.

Two stacked diagrams show UX and development moving from back-and-forth handovers across a solid boundary to a direct UX-to-development flow across a dashed boundary, illustrating how the separation between the roles may weaken.
The boundary between UX design and frontend may blur over time. (Image credit: me)

Spec-Driven Development is another relatively new kid on the block. Behind it is the promise that, if all the essential information (including your UX rules for research, design, accessibility and whatnot) is fed to the AI tool in a specific, machine-readable format, it can produce a complete product in one hit, with no slop to clean up. This approach shows promise — albeit so far, it is not yet the unicorn that everyone has been hoping for (but that is a subject for another story).

Spec-driven development workflow: UX, product management and development contribute to a specification. AI tools execute the specification to produce a product. A feedback loop returns to refine the specification rather than directly modifying the product.
Spec-driven development. (Image credit: me)

Epilogue: the many Ailments

In the words of a colleague, “A month is a 100 years in AI time”. Who knows what I will be writing about a year on. Today, I am certainly suffering from the “UpSkill Sisyphus” ailment, aptly described by Information is Beautiful. I confess that I have even wondered whether AI might soon take my job.

A satirical AI-ailments infographic highlights “UpSkill Sisyphus”: a person pushing a boulder marked with an upward arrow. It represents the exhausting cycle of learning new AI models, testing tools and continually trying to keep up. A grid of other invented AI ailments appears alongside it.
The many Ailments, including UpSkill Sisyphus. (Image credit: Information is Beautiful)

I remain an optimist, however. Which user problems deserve solving? Who has the overall UI vision? Who can read between the lines to establish the future UX strategy? For as long as humans need digital products, UX processes — whether AI-powered or not — will need human input.

For now, the AI monster is somewhat tamed… As for the rest, the future will show!

A lone person stands beneath an enormous orange monster in a fiery landscape. Text reads: “Forget everything I told you. Because it’s already changing.”
Forget everything I told you because it’s already changing. (Image credit: me + Midjourney)


The attack of the giant AI design slop monster: a survivor’s guide was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

Schreibe einen Kommentar