The Problem

I own two Dyson fan remotes. Both control the same types of fans (heating + cooling), but they have completely different button layouts. You know what’s even better? I can’t see either of them in the dark.

And guess when I need to use the heater? At night. In the dark. In my cold bedroom.

So there I am, fumbling around like I’m solving a Rubik’s cube blindfolded, trying to remember: “Is the long button on THIS remote above or below the power button? Which remote is this? WHERE IS THE HEAT BUTTON?!”

I’ve adapted, of course. I can now feel the shape of the buttons in the dark and build a mental map. But here’s the thing: Users shouldn’t have to develop workarounds for bad design.

This is a textbook UX failure. Inconsistent mental models across product variants. Zero consideration for low-light use cases. No tactile differentiation. As a design leader, watching users develop coping mechanisms for preventable usability issues is professionally painful.

So I thought: What if AI redesigned these remotes? Could AI solve these fundamental usability problems better than Dyson’s design team did?

Spoiler: The results reveal as much about AI’s limitations as its potential.

The Setup

I tested 9 different AI tools with basically the same prompt:

"Redesign these two Dyson fan remote controls. They control fans with heating and cooling functions.

Current problems:

  • Poor visibility in low-light/dark conditions (often used at night)
  • Inconsistent button layouts between models
  • Difficult to use without looking at the remote

Your task: Create a redesigned physical remote control that solves these problems."

Some tools got the reference image of my actual remotes. Some didn't. Some tools can explain their thinking. Some just... make pretty pictures and refuse to elaborate.

Let's see what happened.

The Contestants

  1. Adobe Firefly (with reference image)
  2. Microsoft Designer (DALL-E 3, text description only)
  3. Leonardo.ai (with reference image)
  4. Leonardo.ai (without reference image - yes, I tested twice)
  5. ChatGPT (GPT-4 with DALL-E 3, with reference)
  6. Playground.ai (nano model)
  7. Playground.ai (v3 model)
  8. Recraft (with reference)
  9. Recraft (without reference)
  10. Ideogram (4 different variations!)

Yes, that's technically more than 9 outputs. I got carried away.

Round 1: Adobe Firefly - "Let's Add EVERYTHING"

What Firefly did:

  • Added a full digital display at the top (shows "COOL 25°C")
  • Put in a rotary dial for temperature/fan control
  • Reduced overall button count to about 8
  • Premium dark aesthetic

The Good:

  • Digital display solves the "what mode am I in?" problem brilliantly
  • Rotary dial is tactile and innovative
  • Actually looks like something Dyson might make

The Bad:

  • That display will light up your bedroom like a lighthouse
  • Battery life? What battery life?
  • The rotary dial needs to be manufactured precisely or it'll feel cheap

Round 2: Microsoft Designer - "MORE BUTTONS!"

Oh Microsoft Designer. Sweet, misguided Microsoft Designer.

I told it: "These remotes have too many buttons and they're inconsistent."

Microsoft Designer heard: "YOU NEED 20+ BUTTONS AND ALSO HERE ARE FOUR DIFFERENT LAYOUTS."

What Microsoft Designer did:

  • Created FOUR different remote designs
  • Put 18-22 buttons on each one
  • Added labels like "AUTO" and "INFO" everywhere
  • Made the exact problem WORSE

The Good:

  • Um... the rendering quality is nice?
  • At least it tried?

The Bad:

  • This is a TV remote from 2005
  • Completely ignored the "inconsistent layouts" problem by making 4 inconsistent layouts
  • Those tiny text labels would be impossible to read in the dark
  • DID I MENTION 20+ BUTTONS?

This is the poster child for feature creep. When AI encounters "redesign a remote with button layout problems," it interprets this as "show me what a premium remote looks like" rather than "reduce complexity." Microsoft Designer didn't just miss the brief - it actively contradicted it.

Round 3 & 4: Leonardo.ai - The Tale of Two Approaches

With Reference Image: "Let's Play It Safe"

What happened:

  • Made a photorealistic mockup
  • 12 buttons (kept the same complexity)
  • Color-coded buttons (blue snowflake for cool, orange moon for night)
  • Basically just... cleaned up the existing design

When you show AI what already exists, it anchors to that reality. Safe. Professional. Utterly uninspired.

Without Reference Image: "LET'S GET WEIRD/CREATIVE"

What happened:

  • COMPLETELY different approach
  • Added a circular aperture at the top (mimicking Dyson's bladeless fan design!)
  • Used numbered buttons (1-6) instead of labeled functions
  • Only 9 buttons total
  • Actually creative!

The Insight: This is HUGE. The same AI tool, but removing the reference image made it 10x more creative.

When Leonardo could see the existing remotes, it thought: "I should make something similar."

When it couldn't see them, it thought: "I should make something that LOOKS like a Dyson product!"

The learning: Reference images can be creative constraints masquerading as helpful context. Sometimes the best design direction comes from constraint removal, not constraint addition.

(Though those numbered buttons remain a UX nightmare - what does "3" do? Is it temp? Fan speed? A mystery mode? Your guess is as good as mine.)

Round 5: ChatGPT - "I'm Built Different"

ChatGPT came to play. And by "play," I mean "write a complete design specification with rationale."

What ChatGPT did differently:

  1. Provided written documentation explaining every decision
  2. Created a wireframe showing the thinking process
  3. Designed a SYSTEM not just a device

Key innovations:

  • Vertical rockers instead of separate +/- buttons (one for temp, one for fan)
  • Texture differentiation - left rocker smooth, right rocker ribbed (for eyes-free operation!)
  • Hourglass shape - you can feel which way is up instantly
  • Edge-lit icons - buttons glow subtly without blinding you
  • Single universal design - same remote for all Dyson models, firmware enables/disables features

ChatGPT's Design Principles:

  • "Eyes-free operation first"
  • "One Dyson remote system"
  • "Night-optimized"

The Documentation Included:

  • Button architecture explanation
  • Tactile logic rationale
  • Accessibility considerations
  • Cross-model consistency strategy
  • Even suggested a magnetic charging dock!

This is what separates design thinking from design execution. ChatGPT didn't just solve the stated problems - it reframed them within a larger system architecture. This is the difference between a designer and a production artist.

Round 6 & 7: Playground.ai - "Text Rendering? Never Heard of Her"

Nano Model: Not Bad!

What it did:

  • Clean layout with 10 buttons
  • White illuminated indicators around each button (actually brilliant!)
  • "Mode 1", "Mode 2", "Night Mode" labels
  • Simple and effective

Playground's nano model demonstrates something important: you don't need complexity to solve complex problems. Those white indicators create a visual hierarchy in darkness without resorting to full-color displays or elaborate lighting systems.

V3 Model: "DAMOOT"

Oh boy. Oh no.

What it TRIED to do:

  • Ambitious design with 20+ buttons
  • Color-coded gradient rings (red/orange for heat controls)
  • Center elongated button for modes
  • Multiple function zones

What ACTUALLY happened: (everything went damoot!)

  • The center button says "DAMOOT"
  • Just... "DAMOOT"
  • Not "MODE" or "AUTO" or anything sensible
  • DAMOOT

The Text Rendering Hall of Shame:

  • "DAMOOT" (center button, no idea what this means)
  • "Time" (barely readable)
  • Various icon corruptions

The Lesson: AI image generators struggle with generating actual readable text on buttons and labels. When they fail, they fail spectacularly. This isn't a quirk - it's a fundamental limitation that has real implications for product design workflows. You cannot trust AI to write legible button labels in any context where precision matters.

Which, let's be honest, is pretty much every context in product design.

Round 8 & 9: Recraft - From Minimalism to Madness

With Reference Image: Beautiful Simplicity

What Recraft did:

  • Created TWO variants (left: black face, right: silver/cream)
  • Only 8-9 buttons each
  • Huge glowing buttons - amazing for dark use!
  • Color-coded: yellow power, red +/-, blue cool mode
  • D-pad on the black version

(Is that oscillation icon on the left remote... Mickey Mouse? AI's icon generation has some quirks.)

The Color Choice: Both remotes have illuminated buttons that would actually work in the dark. This is smart design.

Recraft understood something crucial: when you can't add more features, make the existing features more usable. Those giant, glowing buttons solve the visibility problem through scale and luminance, not through adding displays or complexity.

Without Reference Image: "I Can Do Annotations! Watch This!"

Recraft got ambitious. Recraft tried to create professional product documentation with labeled callouts explaining each button.

What the labels say:

  • "Powen whe miarlee cosnricals"
  • "Roert Vilyuiluaticen cownens"
  • "Nigm Medlicalgs"
  • "Ditomod Rermere Contert"
  • "Lew Deteonte a lMeale"

These are not real words. These are not words in ANY language.

But here's what's INTERESTING:

Recraft understood that professional product documentation should have annotations. It knew what a spec sheet LOOKS like. It just... has no idea what words ARE.

It's like someone who's seen a thousand product manuals but can't actually read.

What it DID do well:

  • GORGEOUS burgundy/wine red metallic finish
  • Side profile view showing 3D form
  • Professional product photography aesthetic
  • Beautiful brushed aluminum rendering

The meta-lesson here: Recraft's failure is actually more instructive than many successes. It reveals the gap between visual pattern recognition and semantic understanding. The AI knows what documentation looks like, but has no concept of what documentation is.

This is the uncanny valley of AI design tools - sophisticated enough to attempt complex outputs, not sophisticated enough to execute them reliably.

(Also, "Nigm Medlicalgs" is definitely my band name now.)

Round 10: Ideogram - "What If The Remote WAS The Fan?"

Ideogram went WILD and gave me FOUR completely different designs. Let's break them down:

Variant 1: The Pill (14 buttons, blue backlit)

  • Rounded pill shape
  • Blue glowing buttons
  • Clean grid layout
  • Nice, but nothing revolutionary

Variant 2: The Warm Glow ⭐ WINNER FOR LOW-LIGHT

  • Same pill shape
  • WARM AMBER BACKLIGHTING
  • Only 10 buttons
  • Premium aesthetic

This is IT. This is the solution. Warm amber light won't mess with your circadian rhythm the way blue/white light does. This shows actual understanding of the use case - not just "make buttons glow" but "make buttons glow in a way that doesn't wreck your sleep."

This is design thinking.

Variant 3: The Ergonomic One

  • 3D perspective showing actual ergonomics
  • Curved form that fits your hand
  • Small OLED display strip
  • Ribbed grip texture at the bottom

Finally, an AI that understands remotes are held by human hands with actual ergonomic constraints. The ribbed texture and curved form aren't just aesthetic choices - they're functional decisions about grip, orientation, and tactile feedback.

Variant 4: THE REMOTE IS A TINY DYSON FAN 🤯

WHAT.

Ideogram looked at the brief and thought: "The remote controls a bladeless fan... what if the remote LOOKED like a bladeless fan?"

Features:

  • Bladeless fan aperture at the top
  • Speaker grille at the bottom (why? because premium)
  • Gold/copper button accents (matches Dyson's luxury finishes)
  • 15 buttons arranged in a grid
  • Instantly recognizable as Dyson

This is genius because:

  • Visual metaphor that creates instant product recognition
  • Brand consistency across the ecosystem
  • Product line coherence
  • Memorable, differentiated design
  • Actually leverages Dyson's most iconic design element

This is maybe NOT genius because:

  • 15 buttons (still too many)
  • The aperture serves no function (just aesthetic)
  • Manufacturing cost would be higher
  • Form follows brand instead of function

But here's the thing: brand IS function in premium consumer products. The bladeless aperture doesn't need to work - it signals "this is Dyson" instantly. That's worth something in a crowded market.

This is the only AI that understood the assignment wasn't just "make a better remote" - it was "make a better Dyson remote."

The Big Findings

1. AI Can't Write Readable Text on Buttons 📝

Evidence:

  • Playground v3: "DAMOOT"
  • Recraft: "Nigm Medlicalgs", "Powen whe miarlee cosnricals"
  • Microsoft Designer: Tiny illegible labels

Why this matters: AI can't produce manufacturing specs, documentation, compliance labels, or anything requiring textual precision.

The irony: This limitation accidentally pushes AI toward icon-based interfaces - which is best practice for physical products anyway.

Bottom line: AI can recognize what text looks like but has no concept of what text means. It's sophisticated mimicry without comprehension.

2. Digital UI Bias 📱

AI tools default to digital interface patterns when designing physical products.

Examples:

  • Microsoft Designer: 20+ buttons (TV remote pattern)
  • Multiple tools: Added screens (mobile app pattern)
  • Widespread over-labeling (web form pattern)

The disconnect: Physical products have constraints digital interfaces don't - button size, manufacturing complexity, battery life, spatial limits, haptic feedback.

Most AI tools ignored all of these.

Why: Training data skews heavily toward digital interfaces (easier to capture, more frequently redesigned, already digitized).

Result: AI that's fluent in pixels but illiterate in atoms.

Takeaway: Don't assume AI understands physical constraints. You must specify them explicitly.

3. Brand Understanding 🎨

Most AI tools: Generic black/silver/gray remotes

Recraft: Burgundy metallic (matches Dyson's copper/gold finishes)

Ideogram: Bladeless aperture (Dyson's iconic design signature)

Only 2 out of 9 tools showed brand awareness.

Why AI struggles: Brand understanding requires cultural context, historical knowledge, strategic thinking, and subjective judgment - precisely where current AI is weakest.

The gap: AI can copy visual styles but cannot internalize brand strategy or understand why design elements carry meaning beyond function.

Implication: AI can explore variations within a brand system but cannot define or evolve it. Brand stewardship stays human.

The Winners 🏆

Best Low-Light Solution

Ideogram (Variant 2)

Warm amber backlighting is PERFECT for bedroom use. Won't disrupt sleep like blue/white light. This is the answer.

Best Brand Alignment

Ideogram (Variant 4)

Only AI to connect the remote design to Dyson's iconic bladeless fan. Visual metaphor + brand consistency = chef's kiss.

Best Simplification

Recraft (with reference)

Ruthlessly reduced to 8-9 essential buttons. Giant glowing buttons for easy dark use. Sometimes less IS more.

Most Cautionary Tale

Microsoft Designer

Made the problem WORSE by adding 18-22 buttons across 4 inconsistent layouts. This is AI's tendency toward feature explosion on full display.

Most Entertaining Failure

Recraft (without reference)

"Nigm Medlicalgs" and "Powen whe miarlee cosnricals" demonstrate the spectacular failure mode of AI text generation. I laughed. I cried. I'm still confused.

What This Means for AI in Product Design

The Good News ✅

  1. AI can explore multiple directions quickly - Ideogram gave me 4 variations, each with different strengths
  2. AI can identify solutions humans might miss - Novel approaches like warm amber backlighting
  3. AI can challenge assumptions - Recraft's minimalism, Ideogram's bladeless aperture
  4. AI can visualize concepts rapidly - What used to take days of sketching takes minutes

The Bad News ❌

  1. AI cannot produce reliable text - Every tool that tried to render text failed
  2. AI applies wrong mental models - Digital UI patterns don't work for physical products
  3. AI lacks manufacturing knowledge - No consideration for tooling, assembly, cost
  4. AI doesn't understand brand deeply - Only 2/9 tools showed brand awareness
  5. AI can make problems worse - Microsoft Designer added complexity instead of removing it

The Reality Check 🎯

AI is a tool, not a replacement.

  • Use it for exploration and inspiration
  • Don't trust it for final specifications
  • Always verify manufacturability
  • Combine multiple AI outputs
  • Keep a human in the loop

The Experiment I Want to Do Next

Remember how I said removing the reference image made AI more creative?

I want to test conversational AI (ChatGPT, Claude, Gemini, etc.) with a different approach:

No images. Just thinking.

Ask AI to describe in 3-5 bullet points HOW they would redesign the remote - without actually creating a visual design.

Why?

  • AI is marketed as "logical and analytical."
  • Can it reason about design problems without the crutch of visuals?
  • Does it understand physical constraints when forced to articulate them?
  • How do conceptual recommendations compare to visual outputs?

This tests AI's strategic thinking vs. just pattern-matching pretty pictures.

I'll update this post once I run that experiment. Stay tuned!

Key Takeaways

  • Reference images constrain creativity - Try prompting without them for bolder ideas
  • AI can't do text - Don't trust any text rendering in AI-generated designs
  • Conversational AI ≠ Image AI - They serve different phases of design
  • Physical ≠ Digital - AI needs explicit guidance about physical constraints
  • Multiple tools = Multiple perspectives - Test several to see different approaches
  • AI is inspiration, not specification - Always verify with human expertise

Would AI replace a professional industrial designer?

Not remotely. (Pun intended.) But it's a valuable tool for rapid exploration and pattern breaking.

Try It Yourself

Want to run this experiment with your own product?

The Prompt Template:

Redesign [PRODUCT NAME] to solve these problems:
- [PROBLEM 1]
- [PROBLEM 2]
- [PROBLEM 3]

This is a [PHYSICAL/DIGITAL] [PRODUCT TYPE].
Create a redesigned version that solves these problems.

Test WITH and WITHOUT reference images. The difference will surprise you.

Have you tried AI for product design? What worked? What failed spectacularly? Drop me a note at rutu.upadhyaya@gmail.com - I want to hear your "DAMOOT" moments.

And if anyone at Dyson is reading this: I'll consult on your next remote redesign. My rate is one bladeless fan with warm amber backlighting. Call me.