Working with AI means providing text. Lots of it. Prompts, instructions, context. On mobile, that means typing on a 6-inch screen with your thumbs.

Voice input isn’t new. Millions already dictate on their phone or computer. What’s new is AI-powered voice keyboards that don’t just transcribe your speech, they rewrite it. Some are making a radical bet: the QWERTY keyboard can go.

Four questions AI voice keyboards are answering

Gboard (Google Pixel’s preloaded keyboard), Wispr Flow, Typeless, and Monologue are four products answering the same design challenge differently.

1. What does the keyboard look like? How much of QWERTY stays?

2. How much should the AI change what users said? Raw transcript, or AI-polished output?

3. Should users see words as they speak? Real-time streaming, or nothing until the AI is done?

4. How do users fix what went wrong? Can users correct by voice, or do they need to do it manually?

1. What does the keyboard look like?

The keyboard is disappearing in stages.

Gboard keeps full QWERTY. The mic icon sits in the top-right corner. Users can collapse the keyboard into a floating pill toolbar. Voice is an addition to the keyboard.

Wispr Flow removes the alphabet. What remains: a full-size number and special character keyboard.

Monologue strips further. Just punctuation keys (. , ? ! ‘), delete, enter, space, and a big record button. A QWERTY toggle exists as an escape hatch.

Typeless goes furthest. A large mic icon, @, spacebar, delete, and return. To type anything, switch to a different keyboard entirely.

Verdict 👉 The less keyboard they keep, the more they bet that voice and AI can handle what used to require tapping keys.

2. How much should the AI change what users said?

When users say “I think, um, the meeting is on Tuesday, no, wait, Wednesday at 2 PM” into each product, the outputs diverge.

Gboard outputs: “I think the meeting is on Tuesday. No wait, Wednesday at 2 p.m.” Fillers are removed. Punctuation is added. But self-corrections stay. Want AI to clean it up further? Tap the icon after dictating. AI intervenes only when invited.

Wispr Flow outputs: “i think the meeting is on wednesday at 2 p.m” Fillers gone. Self-correction gone. But “I think” is preserved because they said it. Of the three ghostwriters, Wispr Flow edits the least.

Monologue, Typeless usually output: “I think the meeting is on Wednesday at 2 PM.” But sometimes they go further, dropping “I think” entirely: “The meeting is on Wednesday 2pm.” When it kicks in, it’s not cleaning up. It’s rewriting.

Verdict 👉 Same speech in, different texts out: from near-verbatim transcript to full ghostwrite.

What the AI reads beyond your words

All three ghostwriters share a baseline that goes beyond rewriting. They adapt tone to the app (casual in message, formal in email).

3. Should users see words as they speak?

What users see while they speak also differs.

Gboard streams words in real time, word by word. Users watch text form as they speak. This works because Gboard barely changes the words.

Wispr Flow, Typeless, and Monologue show nothing until the user stops. A waveform. A button sound. But no text for 1-2 seconds.

Verdict 👉 Real-time streaming and rewriting depth constrain each other. The more the AI rewrites, the less it can show in real time.

4. How do users fix what went wrong?

Every AI voice keyboard will get something wrong. How users fix it reveals what ‘voice keyboard’ actually means to each product.

Gboard offers layered voice commands: “Delete,” “Clear,” “Send.” Smart Edit adds natural language editing (“Change ‘happy’ to ‘excited’ ”) and AI edit for rephrasing. The full keyboard is always there as backup.

Typeless offers Speak to Edit. Say “change Wednesday to Thursday” or “make this more formal.” AI interprets intent and transforms the content. (Wispr Flow offers similar on desktop, but not on mobile.)

Monologue has no voice editing. If the AI gets it wrong, toggle to QWERTY or re-dictate.

Verdict 👉 Some products think voice input means text entry. Others think it means the entire workflow.

Key UX takeaways

✅ Which keys they keep defines the product’s character. Removing the keyboard is a spectrum. Special characters are often easier to tap than say. Where to draw the line between voice and tap dictates the experience.

✅ Four questions, one bet. Keyboard layout, rewriting depth, streaming, voice editing: all shaped by one fundamental bet. Coherent products are consistent across all four.

✅ Voice editing and voice commands complete the experience. If users need a full keyboard to fix text, the voice-first promise breaks. Input, correction, and action all by voice: that’s a complete experience.

Questions worth asking about your own product

1️⃣ What new frictions does AI create in your users’ lives?

AI made daily life more text-heavy: more messages, more context, more communication to manage. The bottleneck was never typing speed. It was the sheer volume of text that daily life with AI now demands.

✅ Once AI enters your users’ lives, new problems emerge from that adoption. Are you building for those, or still solving yesterday’s problems?

2️⃣ What’s your “keyboard”?

Every product has one: the familiar interface no one questions anymore. The AI voice keyboard makers asked what the keyboard was for (inputting text), then found that AI's rewriting ability could do that job without it.

You don't have to remove your current interface. Gboard didn't, and that's a valid choice. But it's worth asking where your product should lie between Gboard and Typeless.

✅ Which parts of the current interface become unnecessary, and which become more important?

3️⃣ What happens when you remove it?

That removal is your bet. And like the voice keyboards showed, when you decide how much AI should own, that decision doesn’t stay in one place. It shapes what your interface looks like, how much AI processes behind the scenes, what users see while it’s working, and how they recover when it’s wrong.

✅ Is your bet showing up consistently across your product, or contradicting it in another?

Clicked generate, waited 90 seconds, got something half right, tweaked prompt, waited again. Every attempt to generate something from AI costs something real: tokens, GPU cycles, waiting time. This is the iteration tax.

The tax is heaviest where outputs are expensive: video, image, code. A 10-second video render isn’t like a chat response. Users feel the weight before they click.

The best UX reduces this weight before the expensive commit.

Three patterns to lighten the tax

Each pattern intervenes before the expensive commit.

1️⃣ Sharpen the intent

2️⃣ Cheapen the search

3️⃣ Show the price

1. Sharpen the intent

The cheapest iteration is the one that never happens. Smart UX helps users say what they mean before the meter starts running. Plans and outlines surface structure before render.

Gamma shows an editable outline before generating any slides. Users decide structure first, then visuals.

Replit’s Plan mode breaks down complex projects into a task list. Users provide their requirements and it will generate an ordered task list, and nothing runs until you approve.

Sora lets users build or generate storyboards frame-by-frame.

👉 The pattern: show the skeleton, get agreement, then render the flesh.

2. Cheapen the search

When full generation is expensive, the best tools offer a cheaper way to check direction first. Draft modes exist for iteration.

Midjourney V7’s Draft Mode renders 10x faster at half the cost. It explicitly designed for exploration before commitment.

Luma's Draft Mode is 5x faster and cheaper, creating low-res quick video for "exploring ideas in a state of flow" before mastering to 4K.

👉 The pattern: Let users browse cheap before they buy expensive.

3. Show the Price

Users can’t spend wisely if they don’t know the cost. Transparency turns “should I?” into an informed decision.

Leonardo AI puts the token count on the generate button itself. Users see exactly what they’ll spend before clicking. The price updates live as they adjust resolution, model, and quality settings. No surprises.

ElevenLabs shows estimated credit cost before you click generate.

👉 The pattern: Make cost visible, learnable, and predictable.

Three Key UX takeaways

✅ Intent-sharpening has the highest leverage, but requires domain insight. Users often know more than they can easily express. Slide structure exists before visuals. Code architecture exists before code. The challenge is identifying what users already know that your product isn’t yet capturing.

✅ Cheap exploration only works when fidelity is a spectrum. Draft modes make sense for images and video—where “rough version” is meaningful. Code either runs or doesn’t. If your output is binary, cheap exploration means smaller scope, not lower quality.

✅ Cost transparency builds trust, but only when cost is predictable. Image and video have knowable costs: dimensions × model × steps = tokens. Code doesn’t. When cost is unpredictable, showing scope helps more than showing price: checkpoints, task breakdowns, progress.

Actionable questions for your AI products

💡What does your user already know that you’re not capturing?

💡Does your domain allow draft modes, or only smaller chunks?

💡When users regenerate immediately, what failed upstream?

💡Can users predict the price before generating?

AI’s next UX frontier: Continuity

AI tools are getting better at staying in sync with you: your past chats, your projects, your voice.

Behind that improvement is a shift from prompts to persistent context.

For product makers, this changes how we design experiences: not just responses, but relationships over time.

Why AI forgets: Context window constraint

AI models can only process a limited amount of text at a time, called the “context window”.

Today’s models can handle up to 1 million tokens (roughly 750k words or several novels).

Here’s the catch: that window resets with every new conversation. An AI can remember a whole book during one chat, but forget your product name in the next.

Real UX challenge: Context continuity

The core problem centers on what happens when the window closes. People don’t think in sessions or chats. They expect AI to remember things like a teammate would.

The question becomes: How do we sustain continuity after every reset?

Three approaches to persistent context

Each represents a distinct UX approach for maintaining continuity: by storing, structuring, or reconnecting context when needed. These often appear together because continuity is required across multiple levels of user interaction.

Three continuity archetypes

1. 🤝 The Companion (e.g. ChatGPT’s Memory): Learns implicitly and builds personal continuity

2. ⚙️ The Tool (e.g. Cursor Rules): Uses explicit settings to ensure consistent behavior

3. 🗂️ The Librarian (e.g. Gemini + Workspace): Retrieves context from existing data when needed

How AI products are tackling the challenge

(Examples illustrate a primary case. Each product also mixes the others where it fits.)

❶ 🤝 The Companion: ChatGPT’s Memory

🔷 UX Approach: Learn implicitly, remember invisibly

ChatGPT’s Memory quietly tracks user’s writing style, repeated prompts, and preferences. When it learns something new, it shows a brief “Memory updated” notice.

✅ Why It Works:

• Reduces cognitive load through quiet persistence

• Provides clear controls to view or delete memories

• Uses timely injection to create a sense of ongoing familiarity

⚖️ Trade-Off: Convenience vs. transparency (users may not know what triggered a memory)

❷ ⚙️ The Tool: Cursor’s Rules

🔷 UX Approach: Nothing implicit, everything explicit

Cursor lets users create “Rules” files that define how the AI behaves. They act like shared guidelines that persist within the defined boundaries.

✅ Why It Works:

• Keeps user’s mental model clean: “These rules belong to this project”

• Prevents accidental crossover between workspaces

• Reuses patterns without repeated setup, saving user effort

⚖️ Trade-Off: Control vs. setup effort (requires users to configure everything)

❸ 🗂️ The Librarian: Gemini + Workspace

🔷 UX Approach: Don’t remember, just reconnect

When a user asks a question, Gemini searches their Workspace in real time. Existing Docs, Sheets, and Gmail serve as its external memory.

✅ Why It Works:

• Scales without expanding the model memory

• Keeps outputs grounded in source-of-truth documents

• Works with existing data organization without any new setup

⚖️ Trade-Off: Seamless access vs. data exposure (All-or-nothing based on Workspace permission)

How each pattern answers key UX questions

Note: A feature can blend patterns. “Referencing past chats” sits between Companion and Librarian: implicit signals can trigger it, and the system retrieves from your past chats.

UX insights: Designing for continuity

✅ UX can solve the continuity problem

Users experience “forgetting” as friction. Even if a model resets every session, thoughtful design can preserve a sense of ongoing familiarity. Continuity is not about longer memory, it’s more about smarter context handoffs.

✅ Design how AI remembers

Each approach represents a different mental model:

• Implicit feels personal (ChatGPT Memory)

• Explicit feels structured (Cursor Rules)

• Retrieval feels reliable (Gemini + Workspace)

✅ Keep context within reach

Users care less about whether AI “remembers” and more about whether it finds what matters. If full persistence isn’t feasible, design smooth reconnection to past data or context at the right moment.

Key takeaways

To design continuity in your AI product, ask these questions:

• 💡 What should AI remember? → Map what needs to persist

• 💡 When should it activate? → Set clear boundaries

• 💡 Who manages it? → Give users visibility and control

• 💡 How to implement? → Layer archetypes. Most products need more than one

Generative video is expanding its territory from memes to Netflix. The models are getting better with more realistic video and audio generation.

However, we see a lot of room for improvement in UX, because AI video generation is more complicated than text and image generation due to the inherent complexity of video.

Two forces making AI video generation clunky

1. Creative Complexity

Unlike generating text or an image, video demands control over dozens of variables: lighting, camera movement, pacing, audio, and more. Good UX must guide users through these choices without drowning them in controllers.

2. Resource Sensitivity

Every generation burns real money—credits and GPU minutes. Users need to balance quality against budget, while platforms need to prevent cost spirals.

Understanding the user journey

Essential UX features for AI video creation

❶ Inspire: Clarify what’s possible and spark concepts

What's needed: A gallery showcasing creative, out-of-the-box use cases

Why it matters: Users need to see what's possible beyond their imagination

Working examples: Sora and Midjourney lead with galleries on their landing pages, immediately showing the breadth of styles possible

❷ Initiate: Bridge the gap from simple prompts to creation-ready inputs

What's needed:

• Smart prompt enhancement: Turns "dog running" into "golden retriever running toward camera through autumn leaves, wide angle, shallow depth of field"

• Image-first creation for the starting frame before ‘animating’ the scene

Why it matters: Users often underestimate the context AI needs. Poor inputs cause disappointing outputs

Working examples:

• Sora's Storyboard breaks complex ideas into scenes with separate enhanced prompts

  

• Midjourney's text-to-image-then-animate workflow lets users validate the visual before committing to video generation

❸ Iterate: Enable guided exploration while managing costs

What's needed:

• Directional controls that guide AI's creative randomness rather than promising precise edits

• Transparent credit usage display

Why it matters: Users want to iterate but AI generation is inherently chaotic and expensive

Working examples:

• Sora's “Remix strength” setting allows users to control how significantly a video is altered

  

• Midjourney's "Low/High motion" setting provides control over the degree of movement

  

❹ Export: Ensure production-ready output

What's needed:

• Video spec options including resolution upscaling and frame rates

Why it matters: Professional users need high quality

Working examples:

• Runway's 4K upscaling separates draft exploration from final quality

  

Three Key UX insights for AI video generation

✅ Bridge the imagination gap early

The gap between "dog running" and cinema-quality output is huge. Successful UX helps users build creation-ready inputs through prompt enhancement and reference galleries.

✅ Make every dollar count visibly

Video generation burns real resources. Users need constant visibility into costs and smart ways to explore cheaply before committing to expensive final renders.

✅ Embrace creative unpredictability

AI video generation isn't precise editing, it's controlled randomness. Each generation is unique, which means traditional "undo/redo" mental models break down. UX should frame this as creative exploration, offering ways to guide randomness rather than fighting against AI's inherent variability.

Five UX questions when you are working on creative AI products

💡 Inspiration: Do users immediately understand the creative ceiling? Can new users quickly grasp what's possible with your tool?

💡 Input Quality: How well do you bridge from simple ideas to AI-ready prompts?

💡 Creative Expectations: How do we help users embrace AI's creative unpredictability?

💡 Cost Transparency: Can users explore ideas without fear of budget burnout?

💡 Professional Readiness: Does your output work in real-life professional settings?

The reality check

Nearly half of U.S. consumers now use AI tools during online shopping, with 69% interacting with AI-generated prompts when searching for products.1

But here's the disconnect: AI models are racing to redefine e-commerce while still struggling with basic shopping fundamentals.

We’re in a hybrid era where the question is:
How well do AI models support online shopping experience?

Four key UX factors for AI-powered shopping

• ✅ Conversation-Browsing Integration
  Can users fluidly switch between chatting and browsing, with shopping-specific UI when needed?

• ✅ Fit Explanations & Context
  Does AI explain "why this product fits for me" based on user’s specific preferences and context?

• ✅ Deal Confidence
  Can users trust the AI’s recommendation, and independently verify price, reviews, and alternatives?

• ✅ Trusted Purchase Pathways
  Do users feel assured and informed when checking out?

Let’s zoom in on 3 leading AI models worth studying

Perplexity

❶ Conversation-Browsing: 🟢 Strong
Clean product visuals are tightly integrated with chat; allowing users to browse and chat together.

❷ Fit Explanations: 🔴 Weak
Acknowledges user requirements in summaries but lacks comprehensive "why this fits you" explanation

❸ Deal Confidence: 🔴 Weak
Shows one website only, prohibiting price comparisons; oversimplifies reviews into one section

❹ Purchase Pathways: 🟡 Moderate
‘Buy with Pro’ allows one-click check-out, but missing policy-related information on return and delivery

👉 Bottom Line

Prioritizes seamlessness over confidence. Great UI, but lacks decision-making depth.

ChatGPT

❶ Conversation-Browsing: 🟡 Moderate
Easy to switch from browsing to chatting, but visuals are separated from text responses

❷ Fit Explanations: 🟢 Strong
Delivers strong, tailored insights on why a product fits, and guides what to consider when choosing

❸ Deal Confidence: 🟡 Moderate
Includes pricing information from multiple sites; reviews are summarized by source, making validation limited

❹ Purchase Pathways: 🔴 Weak
Offers basic shipping & return policies, but no built-in checkout; users have to complete purchase from external sites

👉 Bottom Line:

Strong on context and decision guidance. Weak on visual UX and checkout integration.

Google AI Mode

❶ Conversation-Browsing: 🔴 Weak
In-context switch between browse and chat is not supported

❷ Fit Explanations: 🔴 Weak
Pulls user inputs into product details without connecting them to meaningful recommendations

❸ Deal Confidence: 🟢 Strong
Surfaces trusted reviews, prices, and specs with clear links, using existing Google’s shopping modules

❹ Purchase Pathways: 🔴 Weak
Displays transparent shipping information yet breaks user flow by requiring external checkout

👉 Bottom Line:

Relies heavily on existing shopping modules rather than providing a truly AI-augmented shopping experience.

How the models stack up

No single platform has cracked AI shopping yet.

• Perplexity: Seamless experience, trust gaps

• ChatGPT: Context-rich, visually disconnected

• Google AI Mode: Data-heavy, fragmented interface

Key UX insights

🤝 Confidence beats convenience: Seamless UX isn’t enough if users don’t trust the outcome. If the system feels like it’s pushing a decision instead of supporting one, users hesitate or drop off.

🔍 Verification matters: Even with strong recommendations, users need proof. When users can verify, they engage more deeply, and trust the AI more over time.

👕 Don't neglect visual design: Shopping is inherently visual, even with AI assistance.

🧠 Augmented decision-making is the future: Users want to offload mental effort but still make the final call. AI should support this with transparency and adaptive learning over time.

Consider these questions for your AI products

• Does your AI feel like it's supporting decisions or pushing them to users?

• Can users easily verify your AI recommendations independently?

• Are you prioritizing confidence over speed, or vice-versa?

• Does your interface support fluid transitions between conversation and visual experience?

The winners will be those who solve for user confidence first, seamless experience second.

Two ends of technical expertise spectrum

'Vibe coding' spans from complete beginners to seasoned developers:

• Low-code end: Assumes no prior coding experience, functions as a "code generator"

• Pro-code end: Serves experienced developers as an "AI-assisted IDE"

Different audience, different UX

Each UX tells users what to expect: ’Try it out’ or ‘Build with it’

• Lovable: Optimized for quick wins and broad adoption

• Windsurf: Crafted for complex, market-ready products

UX overview across user journey

Let's see how they handle each UX step differently

❶ Onboarding

💜 Lovable

• Guidance: Simple interface with placeholder prompt

• Low friction: No tutorial needed

🌊 Windsurf

• Familiar UI for developers with VS Code-based layout

• Onboarding wizard: Imports settings from VS Code or Cursor, lets user choose themes

Key UX Difference from Onboarding

✅ Speed vs. Setup

• Lovable prioritizes immediate engagement with zero setup

• Windsurf invests in configuration that pays off for professional developers

❷ First Prompt

💜 Lovable

• Single prompt: Generates full codebase from single high-level prompt

🌊 Windsurf

• Multiple entry points: User can start coding directly or use AI assistance via chat

Key UX Difference from First Prompt

✅ Lead vs. Assist

• Lovable leads with AI, ideal for beginners

• Windsurf supports with AI, giving developers control

❸ Response

💜 Lovable

• Focus on visual outcome: Displays instant preview—showing what was built, not how it was built

• Black-box generation: Code created behind the scenes

🌊 Windsurf

• Code-first approach: Details code structure, styling, and functionality, focusing on codebase rather than visual preview

• Transparent process: Shows file and code creation step-by-step, explaining logic behind

Key UX Difference from Response

✅ Outcome vs. Process

• Lovable treats code generation as magic and prioritizes the working result

• Windsurf reveals the process and educates through transparency

❹ Handoff (AI → Human)

💜 Lovable

• Hand-holding: Guides beginner through the development process, suggesting next steps such as refinement or GitHub sync

• Educational nudges: Explains in simple language and prompts to connect databases, check docs, manage knowledge

🌊 Windsurf

• Technical empowerment: Trusts developer expertise, offering control through chat or direct code editing

• Minimum intervention: Only offers suggestions when needed

Key UX Difference from Handoff

✅ Guided vs. Self-directed

• Lovable acts as a mentor explaining next steps without jargon

• Windsurf operates as a companion that respects developer autonomy

❺ Iteration

💜 Lovable

• Chat-centric: Iteration is driven through chat conversation where AI navigates and implements changes

• Visual editing: Provides basic visual editor, while code is not directly editable

🌊 Windsurf

• Code-first iteration: Fully editable codebase allowing granular AI-assisted edit line by line

• Fluid collaboration: Smooth transitions between manual coding and AI assistance

Key UX Difference from Iteration

✅ Pair Programming vs. Co-editing

• Lovable creates a pair-programming dynamic where the human navigates and AI drives

• Windsurf blends human and AI coding so seamlessly that the boundaries blur

❻ Fixing errors

💜 Lovable

• One-click solutions: "Try to Fix" button for immediate error resolution

• AI as expert: Handles bugs that users might not understand, focusing on making it work over explaining why

🌊 Windsurf

• AI-augmented debugging: AI detects errors and suggests corrections

• Educational fixes: Explains the "why" behind solutions, allowing users to learn from observing AI's debugging patterns

Key UX Difference from Fixing Errors

✅ Auto Fix vs. Learn

• Lovable shields users from complexity by focusing on fixes

• Windsurf educates through the debugging process, helping developers improve over time

Key UX takeaways

• ✅ Match context to confidence: Lovable removes friction for novices; Windsurf meets developers where they are

• ✅ Abstraction vs. Transparency: Beginners benefit from abstraction, experts from transparency

• ✅ Guidance vs. Autonomy: Lovable suggests what’s next; Windsurf waits for user intent

• ✅ Progressive disclosure: Both tools reveal complexity at different rates based on user needs

Vibe coding spectrum shows that successful AI tools adapt their entire UX to match users' technical comfort levels

Consider these questions for your AI products

👉 Position your product

• Where does your product fall on the expertise spectrum?

• Is your UX aligned with that position?

👉 Choose transparency level

• How transparent should your AI features be?

• When is it better to reveal the process vs. just show the result?

📩 Email isn’t the problem, but the endless volume is

For knowledge workers, the email inbox is a mix of high-stakes communication and digital clutter.

An average professional spends around 11 hours a week sorting, replying to, and searching through emails.

How AI reinvents the core email journey

From incoming chaos to clear communication, email clients, like Notion Mail, Superhuman, and Gmail are tackling overload by embedding AI throughout the user journey.

Here's how AI can address key pain points at each step

Managing incoming messages

The pain point: Inboxes are one-size-fits-all, making it hard to prioritize what matters across work and life. Sorting emails, or even deciding what to ignore, is a time-intensive process.

Goal: Clarity over chaos

Notion Mail

✅ Key UX approach

Auto-label by context, not just sender:

• Learns how the user organize emails (e.g., ‘recruiting’) and auto-applies that logic to similar threads and incoming emails

Clarify user’s intent:

• Suggests matching emails for the user to confirm or correct before labels are applied

Communicating out

The pain point: Writing replies takes too long — especially when it feels overwhelming but still needs to sound personal

Goal: Inbox Zero

Superhuman

✅ Key UX approach

Minimize overload to initiate replies:

• AI initiates the full draft, the user simply reviews and confirms

• Transforms the user’s rough bullet points into a conversational email

Match user’s writing style:

• Suggests drafts based on the voice and tone in the emails the user have already sent

Finding & retrieving

The pain point: Finding information from an old email or long threads can feel like digging through an endless digital junk drawer

Goal: Get the right answer effortlessly

Gmail

✅ Key AI UX approach

Answer-first search: find answers, not emails

• Lets the user ask direct and conversational questions like “When is my flight?”

• Answers concisely, quoting sources below

From quick scan to full context summaries

• Provides skimmable bullet-point summaries

• By suggesting "Show me a longer summary" as a next step, they bridges quick skimming and in-depth reading

Real world limitation 🚧

• Trust gaps: For sensitive and critical emails, users still hesitate to fully trust AI automation

• Edge cases: Complex workflows in fields like law or finance can break simple AI processes

• Early days: Personalization is improving, but still has a long way to go

Key UX takeaways from AI adoption in email

• ✅ AI-first, user-final flow: AI starts the task, but users always finish — keeping control while moving faster

• ✅ Mental model alignment: Organize emails by context to match how users naturally think

• ✅ At-a-glance clarity: Use summaries and highlights to make content easier to scan and act on

• ✅ Information over retrieval: Users aren’t looking for an email, they want the answer inside it

One more layer that great products nail beyond solving tasks:

The attitude their AI takes toward users

Different AI attitudes shape UX differently

AI attitude = behavior style

It’s how your product chooses to help: push ahead, adapt, or quietly support

As AI in products becomes more advanced, matching the right AI attitude to specific context becomes key to great UX

I’ve been drawing little stories from the tech world, moments that make me pause or think. Here’s one for today. A small snack for your mind. 🍪

P.S. Drawn by human, not AI 💛

Vibe coding? Vibe cleaning!

AI is being embedded into every product and workflows

It is easy to focus on what AI feature we can bring to our products.

However, the biggest UX challenge isn’t what the AI can do. It’s how to hand off control between human and machine.

Handoff matters

Poor transitions between AI and users disrupt workflows, and can leave users stranded.

Effective handoffs preserve momentum and enhance collaboration, ensuring every interaction feels purposeful and efficient.

What makes a good AI-human handoff?

• 🦾 Clarify AI’s Role: Set clear expectations for what the AI will handle

• 👁️ Show Previews or Diffs: Avoid jumping straight to the final output

• 🗣️ Ask for Clarification: Prompt users for more detail when needed to stay aligned

• 🛑 Get Consent Before Action: Don’t execute actions without explicit confirmation

• 👣 Guide Next Steps: Offer clear, accessible paths for what to do next

• ✍️ Make Outputs Editable: Allow users to easily refine AI-generated content

  Here's how today's AI products approach the handoff challenge

  

Comparison: Great UX vs. Weak UX

Let's compare how leading tools handle these transitions across three common creative tasks.

✍️ When editing text

Critical transition:

Handoff occurs after the AI suggests edits and the user is invited to take action.

What matters at this moment:

• Can the user see what will change?

• Are the primary next steps surfaced clearly?

• Does the user feel confident about what will happen?

Great 👍 - Notion AI

✅ Why It Works:

• ‘Track changes view’ allows users to confirm edits visually before they commit

• Clearly surfaces the key next steps — Accept, Discard, Insert below, Try again

• UX writing reinforces clarity and intent (e.g., "Insert below," "Accept")

Weak 🫤 - Google Docs

🛑 Where It Misses:

• No visual cues for users to track which words are changed

• A key primary next step, “Replace,” is buried in a dropdown, while “Insert” alone feels vague and lacks context

🏞️ When generating images from inspiration

Critical transition:

Handoff occurs when a user selects how the AI should use the image reference.

What matters at this moment:

• Is it clear how the AI will use the selected content?

• Is the follow-up interaction easy to engage with?

Great 👍 - Midjourney

✅ Why It Works:

• Clearly communicates how the AI will use the reference as prompt, image, or style

• Offers a clear path to the next step — tweak the prompt or add additional references

• Injected prompts are often short and manageable to tweak

Weak 🫤 - Canva AI

🛑 Where It Misses:

• “Use” label is ambiguous

• Only offers a more demanding option: editing a raw prompt

• Inserted prompts are often long and difficult to modify

👩‍🏫 When generating presentations

Critical transition:

Handoff occurs after AI provides initial response and when user provides feedback.

What matters at this moment:

• Does it communicate and align with users before generating final outcome?

Great 👍 - Gamma AI

✅ Why It Works:

• Show an editable outline before creating the outcome

• Ask for essential inputs like purpose, style, and tone upfront — reducing back-and-forth

Weak 🫤 - Canva AI

🛑 Where It Misses

• Doesn’t clarify user’s intent. The AI auto-generates with full of guesswork

• No opportunity to preview the structure or edit key elements beyond color before creation

Key UX takeaways

For better AI-human handoff for your product:

• ✅ Map out critical transitions: Where do handoffs happen in your AI features?

• ✅ Set clear expectations: How can you show what the AI will do before it acts?

• ✅ Ask for just enough input: What’s the minimum user input needed upfront to clarify intent?

• ✅ Design for preview-before-commit: What kind of preview helps users feel confident before committing to the AI’s output?

Search behavior is shifting

Two major changes are transforming how people search.

1. AI-powered summaries (Google AI Overview & AI Mode beta)

2. LLM-based searches (Perplexity, ChatGPT Search)

Both reduce organic clicks by giving users answers directly.

Zero-click search is surging

Users no longer need to click to find answers:

• Almost 60% of searches end without a click. (SparkToro, 2024)1

Users make decisions inside AI, not on product websites.

• Click-through rates have fallen by 27% in B2B software queries since Google introduced AI-generated summaries. (Bain, 2025)2

Traditional search before AI

What shapes user expectations:

• Your landing page with your marketing copy

• Your onboarding experience that guides users with intent

• Users’ self-directed exploration

Search with AI

What shapes user expectations:

• AI-generated summaries pulled from Reddit posts, reviews, and your website

• Brief, simplified explanations that may overlook your key differentiators

• Features AI highlights based on user queries

Why AI search matters to product makers

AI shapes users' expectations even before they visit your product website, compressing the decision-making flow.

Consequently, user’s purpose for visiting your site is changing from exploration to validation.

Actionable tips for discovery UX in AI era

✅ Map the AI discovery journey

• Treat AI summaries as part of the user journey.

• Audit how your product is presented before the user even interacts with your UX.

✅ Identify the gap between AI summaries and your product story

• Ask diverse questions and analyze the range of answers AI provides about your product.

• Interview new users to understand how AI-generated responses shape their expectations of your product.

• Compare where user expectations align or diverge from their actual experience.

✅ Influence AI-readable narratives

• Provide enough structured, machine-readable content to reduce hallucination.

  • Structure with proper headings and text tables

  • Include feature lists, how-tos, and FAQs

• Seed content with context — the user, their problem, and when they’d need you — so AI knows when to surface your product.

✅ Design onboarding for AI-informed users

• Early product UX should reinforce differentiators and clarify potential misconceptions.

• Make task completion obvious from any page.

• Remember: users often land directly on internal pages via AI summaries, not your homepage.

Your product now lives in two parallel worlds:

Product success depends on your ability to bridge these worlds.

Democratizing software development

Vibe coding is reshaping how we build software by replacing coding with intuitive, AI-powered prompting.

This approach blends low/no-code interfaces, AI assistance, and rapid iteration cycles to enable non-engineers to bring ideas to life quickly.

However, these tools still face challenges in achieving production readiness, with varying levels of polish and practical limitations.

UX is the key deciding factor

Each vibe coding tool offers a distinct user experience that has the potential to significantly impact everyday workflows.

Understanding these UX differences is essential to fully harnessing the potential of vibe coding in real-world development.

How easy are vibe coding tools for non-engineers?

Each tool requires a different level of technical background. This article highlights those for users with low to moderate technical expertise.

Let’s look at how UX differs among vibe coding tools

Bolt.new

Bolt.new targets makers who want to go from zero to a running app with little effort.

It requires low technical knowledge to get started, though familiarity with code is needed when fine-tuning.

UX Approach: “One-shot” app creation⚡️

❶ Fast-track creation: Generate a prototype even with minimal user input. It fills in gaps to create logical functionality while keeping assumptions minimal.

❷ Minimal interface: Emphasize simplicity with just three core views: Chat, Code, and Preview. Users build apps via chat, and can modify them through either chat or code.

👉 Verdict: ⭐️⭐️⭐ — Quick & easy, great for validating functionalities

Lovable

Lovable is tailored for non-technical product creators focusing on UI design with attention to UX practices. It assumes essentially no coding capability.

UX Approach: AI as your front-end design-minded engineer 🧑‍🎨

❶ Maximum Autonomy: Fill in a lot of gaps beyond the prompt, applying changes without asking. It shows some product-thinking to achieve the user’s goal.

❷ Design-oriented: Apply basic UX principles and micro-interactions to pre-designed UI components by default.

👉 Verdict: ⭐️⭐⭐️⭐️ — Fully automated and delivers a good-looking UI

Replit

Replit serves everyone from beginners to seasoned developers, with an AI Agent for app building via chat and an AI Assistant for refining code in a full dev environment.

UX Approach: AI-assisted full-featured dev environment 🤹

❶ Predictable: Confirm plans and show a preview before making changes, keeping users involved and aligned with outcomes.

❷ Explainable: Every step is explained, offering technical details. The generated code includes additional comments for future reference.

❸ One-click styling: Offer simple design theme options — light/dark mode, color scheme, style, rounded edge.

👉 Verdict: ⭐️⭐️⭐⭐ — Start with AI agent, scalable with AI assistant, explainable

Tempo Labs

Tempo is designed for cross-functional product teams collaborating on development. No coding is required to start, but technical familiarity helps with customization.

UX Approach: Feels like a design tool, works like an IDE 🧑‍🎨+🧑‍💻

❶ Transparent progress: Generates PRD and visual user flow diagrams from initial prompts

❷ Visual-first editing: Allow users to add components via drag-and-drop and to edit styles similar to Figma or Webflow.

❸ Design system integration: Can import components from Storybook and reuse them to ensure consistency.

👉 Verdict: ⭐️⭐️ — Great product strategy, but overall UX is less intuitive

Different vibe coding tools, different UX approaches

• Bolt.new: the simplest workflow

• Lovable: the most autonomous experience

• Replit: involves the user the most

• Tempo: the highest flexibility for customization

Real world limitation 🚧

• Error handling and AI quirks: These tools can fix errors but they sometimes lead to repetitive loops. Non-engineers may find it challenging to troubleshoot these issues.

• Limited customization: Although most tools offer a variety of pre-built components, deeper customization just by prompting is cumbersome and limited.

• Confidence gap: Without a foundational understanding of coding principles, non-engineers might feel less confident directing the AI or interpreting its outputs.

Key UX takeaways from vibe coding tools

Consider these questions when designing AI products

• ✅ Autonomous vs Approval-driven: Does the AI act on its own to build and fix things, or explicitly involve users in the loop?

• ✅ Flexibility: Once the AI generates something, what options, and how many, are available for the user to customize it?

• ✅ Explainability: Do these tools show how or why the AI did something?

Enjoyed this post?

Leave a comment below—I’d love to hear your thoughts. Or, forward this to a fellow product maker who might find it useful.

💡 Let’s make UX better together.

Leave a comment

AI products can churn out almost unlimited creative works in a snap.

This leads some users to feel less connected, less proud, and even guilty about their creations.

This is not a new problem

In the 1950s, Betty Crocker’s instant cake mix wanted to revolutionize baking — just add water, bake, and enjoy!

But it flopped. Why? Guilt. Housewives didn’t feel like they baked it with their own hands.

The fix was to let them put in effort by adding eggs and decorating the cake. Sales soared.

The effort paradox can trigger AI guilt 😞

• The effort paradox says that people value things more when they put in more effort.

• When AI removes too much effort from creative work, creators feel disconnected from the final output.

• AI guilt happens when users feel like they’re not truly involved in their work due to AI.

AI can flatten creativity 🤖

• If everyone uses the same AI models, the results start to feel formulaic.

• It’s like how Instagram filters made photos look the same.

• Creativity is more about the process than the outcome. And when AI shortcuts the entire process, the magic disappears.

1. Give users creative ownership

Instead of AI fully automating a task, let users edit and refine the output meaningfully before finalizing it.

ChatGPT’s Canvas

• Allows users to directly edit and refine AI-generated content within the editable interface.

• Keeps users involved in the creative process with a sense of ownership.

2. Preserve effort where it matters

Remove repetitive tasks but preserve meaningful effort to help users “own” the process.

Adobe Premiere Pro's AI Search

• Enables editors to find clips using natural language, eliminating the tedious process of manually scrolling through footage.

• While AI assists with laborious tasks, editors remain fully in control of the editing process and storytelling.

3. Encourage unique, personal input

Let users train AI on their unique style by uploading reference images, past work, or defining preferences for more personal results.

Midjourney’s Personalization feature

• Allows users to create multiple personalized style groups by uploading images that reflect their desired aesthetic.

• Ensures that the AI-generated images align closely with the user’s unique vision.

Consider these questions when designing AI products

• 💡 What meaningful creative decisions should remain in users' hands?

• 💡 Which repetitive tasks can AI handle without diminishing satisfaction?

• 💡 How can you allow for personal expression and customization?

Let’s put it into practice

For example, if you’re to design a collaborative AI writing assistant,

• Let it handle research and citations (low-satisfaction tasks).

• Let users define style through their existing content.

• Keep high-level structure and key decisions with the user.

🚀 For your own product: Identify one feature in your product where users might experience "AI guilt" and apply these principles.

Enjoyed this post?

Leave a comment below—I’d love to hear your thoughts. Or, forward this to a fellow product maker who might find it useful.

💡 Let’s make UX better together.

Leave a comment

Reference:

• https://www.allrecipes.com/article/history-of-boxed-cake-mix/

• https://uxdesign.cc/the-effort-paradox-in-ai-design-996a0bc2f7f6

Grok and Claude have unveiled their latest models, each enhancing transparency in their thought processes.
With this shift, major AI players are making real-time reasoning more visible and understandable, as transparency is crucial for building user trust.

Different AI models, different UX

Interestingly, every company has taken a different approach to making AI reasoning visible. Some prioritize simplicity, while others lean into depth.

Here’s key UX factors for real-time AI reasoning

When displaying AI reasoning, each model strives to balance several competing UX principles:

• Transparency vs. Overload – How much reasoning is shown without overwhelming the user?

• Feedback Mechanisms – What indicators signal that the AI is actively thinking?

• Clarity of Presentation – How structured is the reasoning display?

Let’s look at how each AI model implements real-time AI reasoning.

1. ChatGPT o3-mini

ChatGPT concisely displays its reasoning and collapses when it's done. Users have to manually expand it for details.

UX Approach

• Flashing and changing text labels signal thinking progress.

• Keep short reasoning visible, reducing cognitive load while providing something to read while the model is thinking.

• Hide reasoning when done, allowing users to jump into answer

👉 Verdict

ChatGPT makes reasoning unobtrusive while providing some transparency—which may still feel opaque to users seeking deeper understanding.

2. Claude 3.7 Sonnet

Claude doesn't show its full reasoning process by default. Users can expand it anytime for details.

UX Approach

• An animated icon, dynamic text label, and a time counter serve as progress indicators.

• Reasoning is expandable, but it’s not immediately clear that users can expand it.

• Reasoning section is separately scrollable and structured with bullets, giving clarity.

👉 Verdict

By default, Claude offers minimal transparency, letting users focus on the answers. When necessary, users can easily navigate the reasoning process.

3. Grok 3

Grok displays scrolling snippets of its reasoning process in a distinct section, though they aren’t fully readable as it’s fast.

UX Approach

• A time counter, an animated icon, and scrolling snippets indicate active reasoning.

• Reasoning section collapses once done, but clearly guides users to expand for details.

• Reasoning is very detailed and long, but often repeats the final answer.

👉 Verdict

Grok manages user's perceived wait time with a scrolling UI. However, redundant and unstructured reasoning may overwhelm users.

4. DeepSeek R1

DeepSeek generates reasoning continuously, scrolling downward as its chain-of-thought expands.

UX Approach

• The “Thinking...” label, continuously generated texts, and a throbber serves as indicators.

• Reasoning is highly detailed but lacks a structured bullet format.

• Keep users engaged with scrolling downward display.

👉 Verdict

DeepSeek emphasizes progressive reasoning, maximizing transparency. However, the lack of structure and too much detail may overwhelm users.

5. Gemini 2.0 Flash Thinking Experimental

Gemini represents its reasoning through continuously generated texts. But it’s not automatically scrolled.

UX Approach

• A throbber and dynamic "Thinking..." label serve as indicators.

• Users can scroll down at their own pace to read through during reasoning process.

• Reasoning is structured with bullet points and numbers for clarity.

👉 Verdict

Gemini provides high transparency with detailed reasoning while giving users control over scrolling. However, it’s hard to recognize when the thought process is complete.

All AI models' reasoning at a glance

DeepSeek offers the most transparency, while Claude provides the most structured, low-overload reasoning experience.

Key UX Insights

• ✅ More transparency ≠ Better UX: AI should explain itself, but too much detail can be overwhelming. The right balance builds trust without adding cognitive load.

• ✅ The elevator mirror effect: Since AI takes time to generate responses, well-designed progress indicators can reduce perceived wait time and make interactions feel faster.

• ✅ Focus on user's goal: Users primarily seek answers, not reasoning. Overemphasizing transparency can inadvertently shift focus away from delivering answers.

AI transparency isn’t about revealing everything—it’s about showing the right reasoning at the right time to build trust without overwhelming users.

Enjoyed this post?

Leave a comment below—I’d love to hear your thoughts. Or, forward this to a fellow product maker who might find it useful.

💡 Let’s make UX better together.

Subscribe now

It sounds like a no-brainer—labeling a product as “AI-powered” should signal innovation and boost sales.

But customers aren’t buying it

Here’s what studies from WSU and Growth Unhinged found—AI-labeling:

• ❌ Lowers purchase intent on hardware. (TV: 14% ↓, Car: 21%↓)

• 🤨 Fails to increase trust, perceived value, or willingness to pay for software.

• 💔 Weakens emotional trust. Customers question AI’s reliability.

• ⚠️ Especially harmful for high-risk products. (Illness Diagnosis Tool: 35% ↓)

Why customers tune out “AI”

1. Skeptical of buzzwords—“AI” feels like empty jargon.

2. Bad experiences with clunky AI—AI hallucinates.

3. Trust gaps—privacy concerns, lack of transparency, and limited understanding fuel uncertainty.

4. Fear of malfunctions—especially in health, safety, or finance.

* Both studies tested general U.S. consumers, ensuring a broad and real-world applicable dataset rather than focusing on tech experts or early adopters.

How to market AI without saying "AI"

Focus on real value, not buzzwords

• ❌ “AI-Powered Document Creator”

• ✅ “Draft a client-ready proposal in under 2 minutes”

• ❌ “AI-Powered Sales Assistant”

• ✅ “Close deals 40% faster with automated follow-ups”

If you must use AI in messaging:

• ✅ Pair it with clear benefits.

• ✅ Still avoid using “AI” as the headline.

• ✅ Address privacy concerns & AI’s role upfront.

Good example—Notion

Notion doesn’t sell AI as the main attraction—it sells a better way to work. Instead of overhyping AI, it integrates AI as a writing and collaboration assistant.

AI isn’t the value prop— your product is. What matters most is how it solves real problems, not that it’s AI-powered.

Enjoyed this post?

Leave a comment below—I’d love to hear your thoughts. Or, forward this to a fellow product maker who might find it useful.

💡 Let’s make UX better together.

Leave a comment

My insights are backed by research from:

• Growth Unhinged, 2025

• Washington State University, 2024

They are losing paying users1:

• Bumble: -57K (2%) in Q4 2024; expecting another -100K~120K in Q1 2025

• Match Group (Tinder, Hinge, etc.): -1.3M (8%) vs. 16.5M peak in Q3 2022

Their stock price dropped (vs. peak):

• Bumble: -93%

• Match Group: -80%

Swipe fatigue is real

At its core, swipe fatigue isn’t just about too many choices—it’s about a system that prioritizes swiping over meaningful connections.

Users also feel burned out from ghosting, rejection, and a lack of real relationship success.

Where AI shines💖–
Behavioral moderation & scam detection

Major services like Hinge, Tinder, and Bumble use AI for:

• Content Moderation: AI flags potentially harmful or abusive messages and prompts users to reconsider before sending.

• Scam Detection: AI analyzes behavioral patterns to identify romance scams and suspicious profiles.

The potential risk zone🚨–
AI matchmaking assistant

AI is reshaping how users present themselves and connect.

• Profile Optimization: AI coaches users on improving their profiles and photo selections.

• AI Intermediaries: AI serves as an agent, filtering matches and guiding conversations.

• Personalized Recommendation: AI curates potential matches based on user behaviors and preferences.

Here are the recent moves by major apps in adopting AI features.

Hinge—AI as a personal coach

AI provides feedback and coaches users to improve their prompt answers. (Live)

• UX Strategy: Nudge users to provide more detailed responses based on their initial inputs, helping them showcase their personality without controlling their voice.

• Potential Drawback: AI-guided answers may become more specific, but not necessarily more authentic—90% of users admit to lying on their profiles.2

Tinder—AI matchmaking

AI curates personalized recommendations, offering users an alternative to the traditional swipe system. (Planned)

• UX Strategy: Reduce swipe fatigue by helping users find best matches more efficiently, shifting focus from quantity to quality.

• Potential Drawback: While AI-curated recommendations aim to reduce swipe fatigue, users may question how different it truly is from the existing swipe algorithm.

Bumble—AI Concierge️

The AI concierge could interact with other user's AI concierges to evaluate compatibility, narrowing down potential matches. (Announced)

• UX Strategy: Save users time and emotional energy by filtering out less compatible matches, allowing for more focused and meaningful interactions.

• Potential Drawbacks: Over-automation may raise concerns about authenticity, user privacy, and risk creating a dehumanized dating experience.

Grindr—AI Wingman

AI Wingman chatbot assists users in crafting messages and suggesting date ideas. (Testing)3

• UX Strategy: Empower users to navigate conversations more confidently by alleviating the complexities of online dating.

• Potential Drawback: For a platform serving the LGBTQ+ community, AI access to personal conversations raises serious privacy concerns, including risks of outing and data misuse.

But here’s the tension

Many users were already using AI tools to tweak photos and bios before dating apps integrated them. For some, it’s harmless “digital makeup.” For others, it blurs the line between curation and deception.

But AI dating agents push this even further—raising concerns about dehumanized connections and making dating feel more like a simulation than reality. It’s a slippery slope that edges closer to Black Mirror territory.

UX lessons from AI-powered dating apps

• Balance assistance & authenticity: AI should enhance user decisions, not replace genuine connections.

• Privacy vs. Personalization: Transparent data collection practices and user control matter, especially in intimate and private products.

• Don’t ignore the “Real Life” factor: Digital experiences shouldn’t widen the gap between online behavior and real-world outcomes.

Enjoyed this post?

Leave a comment below—I’d love to hear your thoughts. Or, forward this to a fellow product maker who might find it useful.

💡 Let’s make UX better together.

Leave a comment

Apple launched “Invites,” a digital invitation service that closely mimicked Partiful - a free social event-planning app that gained significant popularity from Gen Z and young millennials.

It was crowned Google Play's Best App of the Year and named a finalist on the App Store in 2024.

How Partiful changed the game

Before Partiful, Gen Z and young millennials used a mix of different platforms (iMessage, WhatsApp, Messenger, etc) to plan and coordinate parties. But each had its friction points that made the process messy and less seamless.

Partiful makes it easy to collect RSVPs while keeping things fun. It also allows people interact before and after the party, making it feel like a group chat with event planning built-in.

Why Partiful’s design works

1. Remove friction to adoption

• No app required, just a simple link for RSVPs.

• Instead of forcing users into a new system, Partiful fit into existing behaviors (texting, social media, and group chats).

• SMS reminders are more powerful than app notifications. Unlike apps, people actually check their messages, reducing no-shows.

2. Emotional & social-first UX

• Parties are associated with fun, not logistics

• Partiful made event planning feel lighthearted and exciting.

• Their design - soft multi-colored gradients layered with animated confetti and fun memes creates hype from the planning stage.

3. Create a growth loop

• Partiful invites created an organic word-of-mouth growth engine.

• Every party hosted on Partiful introduced the platform to new people.

• Attendees saw how easy and fun it was, so when they hosted their own event, they were more likely to use Partiful.

Deeper dive into Partiful’s UX

1. Dampening fear of uncertainty: “Should I go?” 🤔

Partiful didn’t just make event planning easier—it reduced the social anxiety that stops people from going to parties in the first place.

• People hesitate to commit to events when they’re unsure what to expect or worry about feeling left out.

• The guest lists allow invitees to keep track of who’s coming before they commit.

• Pre-event group chat lets attendees interact before the event, breaking the ice early.

2. Keeping the hype alive: after-party UX 🎉

Partiful didn’t just focus on getting people to the party—they extended the experience after the event, reinforcing positive emotions and increasing the likelihood of repeat use.

• People may want to stay in touch but without a structured way to reconnect, new friendships fade quickly.

• Group chat stays open post-event and encourages guests to share photos, thank the host, and keep the conversation going.

• Guests can DM each other (optional privacy settings) makes it easy to follow up with new friends.

Worth thinking about

• What hidden anxieties might prevent people from using your product, even if they need it?

• Are you designing for engagement before, during, and after the core experience?

What do you think

When a big player like Apple copies a rising startup, does it validate the idea or kill innovation? What do you think?

DeepSeek introduced a novel UX that showcases how its AI reasoned through problems using logic that resembles human logic.

Following DeepSeek's success, Google and OpenAI have now made similar adjustments, reinforcing this trend.

Why it matters

1. From black box to open book

For years, AI chatbots have been black boxes, offering answers with limited transparency into their reasoning process. DeepSeek’s 'thinking out loud' feature flips this model by making AI’s real-time thought process visible and understandable.

2. AI's new era of transparency

This move toward transparency is already being adopted by major AI companies. By making AI reasoning more trustworthy and better aligned with human intentions, it’s transforming the field.

Why this design works

1. Building trust through transparency

• Users are often skeptical about AI-generated results, as they know hallucinations happen.

• Transparency allows users to identify where errors occur, making it easier to correct them.

• Users feel more confident in verifying and correcting errors, which increases trust.

2. Enhancing user engagement and involvement

• Users don’t just want answers—they want to understand them.

• Watching AI’s thought process keeps users engaged, and makes them active participants in problem-solving.

• When users feel mentally engaged in an interaction, retention and satisfaction improve.

3. Leveraging feedback for improvement

• User feedback is crucial for model improvement.

• More detailed feedback (e.g., “Step 3 doesn’t make sense”) from users helps fine-tune the model rather than just “good” or “bad” ratings.

Why Now?

• Growing concerns about AI bias, hallucinations, and safety make transparency critical.

• The AI industry is racing to gain user trust, and explainable AI is a major competitive advantage.

• Governments and regulators are pushing for AI accountability, making this feature not just a UX improvement but also a compliance strategy.

Worth thinking about

When you’re dealing with AI products: how can you make your product more explainable?

Adobe introduced AI-powered search in Premiere Pro that lets users find footage using natural language descriptions like "person skating with a lens flare."

How it works

• Import videos and Media Intelligence will automatically recognize objects, locations, and camera angles

• Everything is processed locally on your device without sending content for AI training

• Enter natural language queries in the search panel

• Filter with visuals, text, and metadata to get precise results

Why it matters

Traditionally, video editors spent countless hours manually reviewing and organizing footage through meticulous logging and tagging systems to find what they needed.

By delegating media analysis to AI, editors can focus on more creative storytelling, enhancing both productivity and artistic output.

Key insights

1. Enhancing human creativity

• AI doesn’t replace the editor’s creative role.

• Keep creative control with users while taking over repetitive and mundane tasks.

• Avoid intrusive automation that could frustrate professionals.

2. Prioritizing copyright and privacy for creators

• Creators want to protect their original work from being used AI model training.

• Ensure user’s content is never used to train any AI models.

• Process all AI operations on users' local devices.

3. Making AI invisible and integrated to whole product

• The best AI is the one users don’t even notice

• Media Intelligence working in the background

• Integrate AI within existing search features to maintain a familiar user experience

Reality limits

• At this stage, this feature doesn’t recognize audio content. Users have to run AI-powered transcription feature separately to access transcript.

• The search cannot detect specific people, emotional expressions (e.g., happiness), or technical characteristics (e.g., stable camera movement or focus).

Worth thinking about:

When you’re dealing with AI features:

• Map your product's workflow. Identify which tasks are "mechanical" vs. "creative/strategic”

• Which tasks are preventing users from doing their best work?

• How can you seamlessly integrate AI into your existing system?

Bluesky introduced "Starter Packs" in June 2024. Think of it as a playlist for accounts - users can create and share themed collections of accounts to follow. New users can join these curated communities instantly with just one tap, making it easy to find their people on Bluesky.

How it works

• Visit the Starter Packs tab on your profile to create a pack with up to 150 recommended users and 3 custom feeds.

• Choose between auto-generating recommendations or selecting users manually.

• Share your pack using the provided link or QR code.

• Get notified when someone joins Bluesky through your pack.

• Users can opt-out by blocking the pack creator or report packs to moderation.

Why it matters

As users were looking for alternatives to X/Twitter, Starter packs played a crucial role in their migration to Bluesky. Here's why it mattered:

1. Solving the Cold Start Problem

• When users migrate, they start with zero connections.

• Starter packs solved this by enabling communities to move together, reducing time to first meaningful interaction by 50%.

• During the X/Twitter ban in Brazil, Portuguese-language starter packs helped Brazilian users migrate as a community.

2. Reaching critical mass

• Starter packs drove 43% of daily follows at peak, with 94% of these connections remaining active.

• This quickly built the critical mass needed for the network to become self-sustaining.

3. Preventing AI/algorithms dominance

• While centralized curation is vulnerable to spam and manipulation, decentralized curation lets users control what they see.

• Decentralized curation allows users to control their discovery rather than being fully controlled by recommendation engines.

Reality limits

Reinforcement of Inequality ("Rich Get Richer" Effect)

While already-popular accounts gained even more visibility, these packs have created a widening gap between popular and less-known accounts.

Worth thinking about:

When you're competing against dominant platforms:

• Identify the "switching costs" that keep users locked into existing platforms.

• How can you creatively lower these switching costs?

• How can you help new users find value in your platform more quickly?

Instagram is shifting its profile grid from square (1:1) to vertical (4:5) thumbnails, prioritizing short-form video content formats over legacy square layouts.

Why it matters

Market adaptation vs. User backlash

This change sparked strong opposition from creators who spent years perfecting their square grid aesthetics for their brand.

Yet Instagram prioritizes market adaptation over Instagram’s original identity.

What are they building?

• Built a new vertical grid system that displays rectangular thumbnails.

• To move Story Highlights to a separate tab to make more space for vertical thumbnails.

• To allow to re-order the entire grid.

• To add direct-to-grid posting option.

Why would they change?

1. User behavior reality

The platform is aligning with users’ evolving content consumption habits and filming preferences, matching current upload trends of vertical photos and videos.

With most users discovering content through algorithm-driven feeds, a creator's grid layout has become less important for driving engagement.

2. Gradual video-focus shift

Instagram has a Reels tab on profiles, but it's secondary to the main square multimedia grid. Instead of elevating the Reels tab to primary position (which could alienate photo-sharing users), they're adapting the main grid to vertical format.

This change optimizes the primary tab for both photos and videos, making it a more versatile showcase for all content types.

3. Strategic timing

Instagram aims to capitalize on TikTok's current market challenges to drive user acquisition.

To ease migration for TikTok creators, they are offering the familiar vertical thumbnails (4:5), along with financial incentives.

Worth thinking about:

When you’re dealing with a successful legacy product:

• Analyze if your past success formulas are against users’ evolving behaviors.

• Quantify the impact: How much do legacy features affect primary metrics?

• Consider creating migration tools or guides for affected users.

What do you think?

Instagram's Home and Explore tabs haven't yet adopted its video-focused approach.

Will Instagram create an innovative new standard, or will it follow TikTok's design patterns again?