Most AIUX is just search with extra steps?

We were given the most transformative technology in decades and we turned it into a fancier text box. I think we can do better! Here’s how…

Open ChatGPT. Open Claude. Open Gemini. Open Copilot. Open Perplexity.

A text input. A submit button. A wall of generated text. Every time.

Open any AI product. What do you see?

A text input. A submit button. A wall of generated text.

We’ve had two years to figure out how humans should interact with AI. The collective output of the design industry has been: slightly nicer search engines.

I’m not being glib. I’ve spent six months analyzing 50+ AI products, cataloging their interface patterns, trying to understand what makes some AI experiences feel transformative while others feel like typing into a void. And the uncomfortable conclusion I keep reaching is that most of us are designing AI interfaces on autopilot.

We’re not designing for AI. We’re designing for what we already know.

The search box won

Here’s what happened: generative AI emerged, and we needed interfaces for it immediately. There was no time for research, no time for experimentation, no established conventions to draw from. So we reached for the nearest familiar pattern.

The search box.

It made sense. Users type a query, the system returns a response. That mental model was already universal. Why reinvent it?

Jakob Nielsen, the usability pioneer, has called AI “the third user-interface paradigm in computing history” — a shift from command-based interaction to what he terms “intent-based outcome specification.” But here’s the irony: while AI represents a fundamentally new paradigm, we’re packaging it in the interface conventions of the old one.

But search and AI are fundamentally different interactions. When I search Google for “weather in Tokyo,” I expect a factual answer. There’s a correct response, and Google either has it or doesn’t. The interface reflects this: a clean result, maybe a widget, no ambiguity.

When I ask an AI to “write a marketing email for my product launch,” I’m not searching for an answer. I’m initiating a collaboration. The AI doesn’t retrieve something that already exists. It creates something new, shaped by countless probabilistic decisions I have no visibility into.

Same interface. Completely different interaction. And we’re pretending that’s fine.

What we lost by defaulting to search

The search paradigm carries assumptions that don’t apply to AI:

Search assumes there’s a right answer. AI generates possibilities. When you search “capital of France,” there’s one correct result. When you ask AI to draft an email, there are infinite valid outputs. The interface should help you navigate that possibility space. Instead, it gives you one output and a “regenerate” button: the equivalent of saying “try again and hope for something better.”

Search assumes the query is the hard part. With Google, articulating what you want is the main challenge. Once you’ve typed the right keywords, the system does its job. With AI, the query is just the beginning. The real work is iterating, refining, steering the AI toward what you actually need. Most AI interfaces treat that iterative process as an afterthought.

Search assumes you’ll recognize the answer when you see it. If I search for a fact, I can verify it. If AI generates a persuasive-sounding paragraph with a subtle factual error, I might not catch it. The interface should help me calibrate my trust. Instead, AI outputs are presented with the same confidence as Google search results, as if hallucinations weren’t a defining characteristic of the technology.

Search is transactional. You query, you get a result, you leave. AI interactions are often conversational, building context over time. But we’ve designed AI interfaces like transactions: each prompt treated in isolation, previous context buried in scroll, no sense of an evolving collaboration.

We took a paradigm designed for retrieval and applied it to generation. Then we wondered why users feel like they’re fighting the interface to get what they want.

The features that prove we know better

Here’s what’s frustrating: the companies building AI products clearly understand these limitations. You can see it in the features they bolt on.

ChatGPT added custom instructions because the search paradigm couldn’t accommodate persistent context. They added memory because treating every conversation as isolated was obviously broken. They added canvas because generating text in a chat bubble was the wrong container for actual work.

GitHub Copilot never fully committed to the search paradigm in the first place. Inline suggestions, appearing and disappearing as you type, accepting with a keystroke. That’s a different interaction model entirely. It’s closer to autocomplete than search.

Midjourney built an entire interface around the assumption that the first output won’t be right. Variations, upscaling, remixing: the core experience is navigating possibility space, not receiving a single answer.

These aren’t minor UX improvements. They’re acknowledgments that the search box was the wrong starting point.

But instead of rethinking the paradigm, we keep patching it. Custom instructions here, memory there, a canvas mode when chat bubbles don’t work. We’re adding rooms to a house built on the wrong foundation.

What AI interfaces could be instead

I don’t have a complete answer. Nobody does yet; this is genuinely new territory. But I can see glimpse of better paradigms in the products that are working:

Workspaces instead of chat threads. Linear’s AI features don’t live in a separate chat interface. They’re embedded in the workspace where you’re already doing your job. The AI operates on your actual work artifacts, not abstract prompts in a conversation. This approach aligns with Google’s People + AI Guidebook, which emphasizes designing AI that fits into existing workflows rather than demanding users adopt new ones.

Steering instead of prompting. Some creative tools let you adjust parameters in real-time (temperature, style weights, constraint sliders) rather than trying to articulate everything in natural language. The interface gives you direct manipulation of the generation process, not just input and output.

Confidence as a first-class element. Perplexity shows its sources inline, letting you trace claims back to evidence. Some medical AI tools display probability distributions rather than single answers. The interface acknowledges uncertainty rather than hiding it. Microsoft’s Guidelines for Human-AI Interaction explicitly recommends making clear “how well the system can do what it can do” — yet most AI interfaces ignore this guidance entirely.

Iteration as the core loop. Midjourney’s entire experience assumes you’ll need multiple rounds. The interface is designed for exploration, not for getting it right the first time. “Regenerate” isn’t a fallback — variation is the primary interaction.

Ambient assistance over explicit queries. The best Copilot interactions happen when I forget it’s there. It suggests, I accept or ignore, we move on. There’s no prompt, no submit button, no waiting for a response. The AI fits into my existing workflow rather than demanding I enter its interface.

None of these are revolutionary in isolation. But they share a common thread: they’re not search.

The cost of getting this wrong

This isn’t just an aesthetic problem. When we design AI interfaces poorly, we create real costs:

Users don’t trust outputs they should trust. I’ve watched people refuse to use AI-generated content that was perfectly good because the interface gave them no way to evaluate it. The black-box presentation creates anxiety that makes AI less useful than it could be.

Users trust outputs they shouldn’t trust. The flip side: confident presentation of uncertain outputs leads people to accept hallucinations, miss errors, and over-rely on AI judgment. Research on trust calibration shows that both “overtrust” (uncritical reliance) and “undertrust” (avoiding AI even when beneficial) lead to poor outcomes and the current search-box paradigm does nothing to help users calibrate appropriately.

We’re leaving capability on the table. Current AI can do more than most users ever discover. But the interface — that simple search box — doesn’t reveal the possibility space. Users type basic prompts because the interface suggests that’s all there is.

We’re creating learned helplessness. When users can’t get what they want and don’t understand why, they conclude AI “doesn’t work for them.” I’ve talked to designers who gave up on AI tools entirely, not because the technology failed, but because the interface failed to teach them how to collaborate with it.

The technology is advancing rapidly. The interfaces are barely keeping up. And the gap between what AI can do and what users can access through current UX is widening.

Confessions from the field

I need to be honest about something. I’ve been documenting AI UX patterns for the past eight months, building a resource at aiuxdesign.guide And for most of that time, I was making the exact mistakes I’m criticizing here.

I kept building features. A pattern simulator. A Figma prompt generator. More patterns, more documentation, more examples. When the traffic didn’t come, my instinct was to build more. Surely if the product was good enough, people would find it.

Then someone told me something that stung: “You have a discovery problem, not an engagement problem. Stop building features. Start driving traffic to what you’ve already built.”

I was doing with my own site what I’m accusing the industry of doing with AI interfaces: defaulting to what felt familiar (building) instead of doing the harder work of understanding how users actually find and use things.

There’s a phrase that haunted me from user feedback: “It’s a menu without recipes.” I had documented 28 patterns with real examples and working code. But I wasn’t showing the messy reality the delays, the failures, the edge cases that make AI implementation actually hard. I was packaging my knowledge in the same tidy format everyone else uses.

Sound familiar? Take a search box, add AI, call it innovation.

The turning point came when I admitted something uncomfortable: I didn’t have a formal methodology. I was just… looking at AI products, noticing what worked, writing it down. For months I felt like an impostor because I wasn’t doing “real research.”

Then I learned there’s a name for what I was doing: observational pattern mining. Design anthropology. Nielsen Norman Group calls it “field studies” — going where the users are, watching, asking, and listening. Observing people in context interacting with systems. It’s how most design knowledge actually accumulates not through academic papers, but through practitioners paying attention and sharing what they notice.

The same is true for AI interface design. We don’t need to wait for someone to publish the definitive research on post-search-box paradigms. We can observe what’s working, notice the patterns, and build on them.

But that requires something uncomfortable: admitting that the current approach isn’t working. Looking at the data. Accepting that “users understand the search box” isn’t the same as “the search box is right.”

I’m still learning this lesson. Building is comfortable for me as a product designer and developer. Distribution is hard. Challenging the search-box paradigm is uncomfortable. And yet.

A call for design courage

Here’s my challenge to all designers (including me) working on AI products:

Lets stop defaulting to the search box.

I know why we do it. It’s safe. It’s familiar. Users already understand it. Stakeholders won’t push back because it looks like what they expect AI to look like.

But “users understand it” isn’t the same as “it’s the right solution.” Users understood command-line interfaces too. That didn’t make GUIs unnecessary.

As Saleema Amershi and the Microsoft Research team wrote when introducing their human-AI interaction guidelines: “While classic interaction guidelines hold with AI systems, attributes of AI services, including their accuracy, failure modalities, and understandability raise new challenges and opportunities.”

We’re at the GUI moment for AI. The paradigm that will define how humans interact with AI for the next decade is being established right now, through the design decisions we make today. And we’re sleepwalking into a future where that paradigm is… a text box.

The products breaking out of the search paradigm are finding new interaction models — steering wheels instead of query boxes, workspaces instead of chat threads, confidence gradients instead of binary outputs. These aren’t just better UX. They’re better mental models for what AI actually is.

We have an opportunity to shape how a generation understands and works with artificial intelligence. We can give them interfaces that reveal AI’s nature — probabilistic, iterative, collaborative — or we can give them interfaces that hide it behind a familiar facade.

I know which choice is easier. I also know which one we’ll regret.

Two years in, and we’re still designing AI like it’s a database with better marketing. Input query, receive answer, repeat.

The technology deserves better. The users deserve better. And honestly? Designers capable of creating genuinely new interaction paradigms deserve a more interesting problem than “make the chat bubble prettier.”

The search box won the first round. It doesn’t have to win the war.

References & Further Reading

• Nielsen, Jakob (2023). “AI Is First New UI Paradigm in 60 Years.” Nielsen Norman Group.
• Amershi, Saleema et al. (2019). “Guidelines for Human-AI Interaction.” CHI Conference on Human Factors in Computing Systems / Microsoft Research.
• Google PAIR (2024). “People + AI Guidebook.” Google Research.
• Kapania, Shivani et al. (2024). “A Systematic Review on Fostering Appropriate Trust in Human-AI Interaction.” ACM Journal on Responsible Computing.
• Stanford HAI (2024). “How Do We Design and Develop Human-Centered AI?” Stanford Institute for Human-Centered Artificial Intelligence.
• Lee, John D. & See, Katrina A. (2004). “Trust in Automation: Designing for Appropriate Reliance.” Human Factors.

All diagrams and screenshots created by the author. Product screenshots used under fair use for commentary and criticism.

The author documents AI UX patterns at aiuxdesign.guide.