Essay · AI · Design Practice

Designing with AI: From Illusion to Intention

On what large language models actually are, why that matters for how we design with them, and the move from blind trust to genuine, informed collaboration.

01 / The Question

Do we actually know what this thing is doing?

A few months ago, I gave a talk to Harver's Product and Engineering teams. I didn't pitch a framework or walk through a case study. I just wanted to ask a question we don't ask enough in rooms full of people building with AI.

Not in a conspiratorial way. In a genuinely curious one.

01 / The Question

We work with AI the way people once worked with electricity.

We know it works. We've learned not to touch certain things. But the inner life of it? Still pretty mysterious.

And that mystery, I've found, is where a lot of well-intentioned design goes wrong.

01 / The Question

Apple called it an illusion of thinking.

Apple published research on reasoning-enabled Large Reasoning Models — the ones marketed as capable of "deeper thinking" — and found something uncomfortable: as tasks got more complex, these systems made more errors, not fewer.

The researchers called it an illusion of thinking. That phrase sat with me for a long time. Because when you believe AI is really thinking, you make different decisions. You defer more. You question less.

The question that started the talk

"Do we actually know what this thing is doing?"

Asked to a room of designers, PMs, and engineers building with AI every day.

AI as infrastructure placeholder

We use it. We rely on it. We rarely understand it well enough to question it.

Apple Research · Large Reasoning Models
Finding
As task complexity increased, reasoning-enabled LRMs made more errors — not fewer.
Named
"An illusion of thinking" — the appearance of deep reasoning without the underlying capability.
Design implication
If the model isn't really thinking, the responsibility shifts back to the person directing it.

02 / The Research

The placebo nobody talks about.

A CHI 2024 study found that participants consistently believed AI had improved their performance — even when they'd been explicitly warned beforehand that it might make things worse.

The belief in AI's capability improved outcomes more than the system's actual performance.

02 / The Research

The belief improved outcomes. Not the tool.

Participants brought confidence to their work because they thought they had a powerful tool. That confidence changed how they showed up. And the results reflected that — not the tool's actual capability.

This isn't a story about AI being good or bad. It's a story about how much our expectations shape what we experience.

02 / The Research

We're not just designing interfaces.
We're shaping belief.

We're shaping what people believe is possible. And that belief has real consequences — for how they work, how they trust, and how they act on outputs.

As a designer, I find this both fascinating and a little unsettling. The interface is never neutral. It is always an argument about what the system can do.
CHI 2024 Study Finding
Condition
Outcome
Warned AI might hurt performance
Still believed AI helped
Believed in AI capability
Performed better
Driver of improvement
Expectation — not the tool

The design responsibility

"We're not just designing interfaces. We're shaping what people believe is possible."

And that belief has real consequences for how people work, trust, and act.

It isn't reasoning
the way you reason.

When you type a message into an LLM, it gets broken into tokens — numerical representations of words and parts of words. The model processes those tokens through layers of mathematical calculations, finding the most statistically likely next token based on everything it was trained on.

It does this again. And again. Until it has a response.

It's finding the nearest
related vectors.

The output can look like thinking. It can feel like thinking. But understanding what's actually happening under the hood changes how you work with it.

You stop waiting for the model to figure things out. You start designing the conditions under which it performs well.

04 / Prompting as Design Work

Prompt engineering is interaction design for intelligent systems.

When you write a prompt, you're not just asking a question. You're designing a reasoning pattern. You're setting up the conditions for a particular kind of output.

The same instincts that make you a good designer make you good at this too.

04 / Prompting as Design Work

Three techniques that changed how I work.

Meta-prompting — asking the model to think about how it should approach a problem before it tries to solve it. It's like briefing a collaborator before they dive in.

Function encapsulation — breaking complex tasks into smaller, well-defined steps. Don't ask it to do everything at once.

Output templating — giving the model a structure to fill rather than a blank canvas. It performs better, and you get outputs that are actually usable.

04 / Prompting as Design Work

None of this requires code.
It requires design instincts.

Clarity about the goal. Empathy for what the system can and can't do. Iteration based on what you observe.

The skills are already there. The practice is new.

How a prompt becomes a response
Input Design with AI
Tokens 42891 411 15836 → vectors
Layers Mathematical calculations across attention heads
Output Most statistically likely next token — then the next — then the next.
Three prompting techniques
01
Meta-prompting Ask the model how it should approach the problem before it tries to solve it.
02
Function encapsulation Break complex tasks into smaller, well-defined steps. One reliable thing at a time.
03
Output templating Give the model a structure to fill, not a blank canvas. Better performance, usable output.
Design instincts placeholder

Clarity · Empathy · Iteration — the same skills that make a good designer make a good prompter

When we treat AI as math, not magic, something shifts.

When we treat AI as magic, we abdicate something.

We stop questioning. We stop iterating. We accept outputs that we should push back on. We build products that reflect the model's limitations instead of our users' needs.

And the people using those products pay the cost — in decisions made on flawed outputs, in systems they can't interrogate, in trust that gets broken without explanation.

When we treat AI as math, we engage differently.

We ask better questions. We stay in the driver's seat. We design the conditions under which the model performs well, rather than hoping it figures things out on its own.

That's a more honest relationship with the tool. And it produces better work.

06 / The Intention

Not trust that's blind.
Not skepticism that's paralyzing.
But genuine, informed, curious collaboration.

The move from illusion to intention isn't about being skeptical of AI. It's about being thoughtful with it. Learning to think with it rather than surrendering to it.

Where you know what the tool is, what it isn't, and what that means for the work you're doing together. That's what good design with AI looks like.

← Back to Thinking