Discover how AI actually works — #AIwithUX Series #02

What I learned looking under the hood — moving beyond prompt engineering, and why understanding how machine learning actually works makes us better designers

Picking up where I left off

My last post was about prompting — how I talk to AI, and how much that changes what I get back. This post is the next question I had: okay, but what’s actually happening underneath these tools?

Over the last few years, most of us watched AI go from a distant engineering topic to the main feature request on our roadmaps. When I first started designing around it, my focus was purely on the surface — how do I design a clean chat box? What does an “AI sparkle” icon even mean to a user? How do I handle the loading state while it’s thinking?

Then I realized something: designing for AI without understanding how machine learning actually works is a bit like designing a mobile app without knowing what a touchscreen is. You end up with interfaces that look fine until real-world unpredictability shows up and then they fall apart.

So I went back to basics. I worked through Google’s AI Essentials course on Coursera and put together a deck breaking down how this stuff actually works, through a design lens. This post is that, written out , for anyone else who’s used AI constantly but never really learned what’s running underneath it.

AI is already everywhere, we just don’t call it that

Before getting into how it works, it’s worth noticing how normal AI already is. Netflix and YouTube personalize your feed based on what you’ve watched. Google Maps reroutes you using real-time traffic data. Companies across industries are using it to speed up testing and streamline how they work.

None of this is futuristic anymore. It’s just infrastructure. Which is exactly why I think it’s worth understanding.

From rule-based to machine learning: the shift that actually matters

We used to design rule-based systems, even if we didn’t call them that. If a user clicks button A, show screen B. Early chatbots worked the same way — fixed, human-written logic, if this then that. The moment a user asked something outside the script, the experience broke.

Machine learning flips that. Instead of a person writing every rule by hand, you feed the system a large set of examples ,training data and it learns to recognize patterns and respond to things it hasn’t seen before on its own.

I’ve started thinking of it as the shift from deterministic UI to probabilistic UI. A rule-based system gives you the same, predictable output every time. A machine learning system gives you an output based on its best guess — which means our job isn’t just mapping a fixed user flow anymore. It’s designing for a system that’s usually right, but not always certain.

Why training data quality is a UX problem, not just a technical one

This might be the part that stuck with me most, because it’s really a data quality issue and that’s something designers already understand.

An AI model’s accuracy depends completely on the data it was trained on. There’s a simple analogy that made this click for me instantly: if a model is only ever shown pictures of red apples, it won’t recognize a green or yellow apple as an apple at all. That’s algorithmic bias in its simplest form.

Here’s the reframe that mattered to me: in technical terms, that’s a training data problem. In UX terms, that’s a broken user journey — someone trusted the system, and it quietly failed them. Bad or narrow data doesn’t just produce technical errors, it produces “hallucinations” and unreliable output that a user has no way of knowing is wrong. Which makes me think designers need to be asking, early: what was this actually trained on, and how would a user even know if it got something wrong?

Three ways machines actually learn

Once training data shapes everything, the next question is how a model learns from it. There are three main approaches.

01. Supervised learning: the model learns from labeled examples. Show it thousands of cat photos labeled “cat,” and it learns to recognize cats in new ones.

02. Unsupervised learning: no labels this time. The model looks at raw data and finds patterns or groupings on its own.

03. Reinforcement learning: it learns through trial and error, refining itself over time based on a kind of reward or penalty for each action.

Different mechanisms, same underlying idea — learning from experience instead of being told exact rules.

Generative AI, and the tools I’m actually talking to

Generative AI is a branch of machine learning that does something different from the above — instead of just recognizing or predicting, it creates new things: text, images, audio, even video, from a plain language prompt.

This is where tools like Claude, Google Gemini, and ChatGPT come in for me day to day. They take an open-ended request and give back something that actually fits the context, rather than a rigid, pre-written response. In practice, that’s shown up in a few specific parts of my design work:

• Rapid visual prototyping : quick hero images or mockup visuals, without needing a full shoot.
• UI copy and microcopy : faster drafts for onboarding flows, notifications, and emails.
• UX ideation and brainstorming : a thinking partner that helps me work through product questions or catch edge cases I might’ve missed.

None of it replaces judgment. It just shrinks the gap between having an idea and having something to react to.

What AI is genuinely good at

If I had to sum it up in one line: speed. AI can take a long thread of user feedback and boil it down to the key points in seconds , something that used to eat up an entire afternoon. By handling routine drafting and answering repetitive questions instantly, it frees me up to spend more time on the parts of the job that actually need a human — strategy, and genuinely understanding the user.

What it still can’t do and why that’s a design problem too

AI doesn’t learn on its own in real time, it needs people to keep updating and retraining it. And it will sometimes state something false with total confidence, which is a much harder failure to catch than an obvious error.

That’s the part I keep thinking about as a designer: how do you design something that’s honest about not being sure? A system that’s confidently wrong is more dangerous than one that visibly struggles, and I don’t think our field has fully figured out how to design for that yet. What I do know is that nothing AI produces should go anywhere near a real user without a person checking it first.

Why designers need to stay in the room

AI doesn’t have empathy, and it doesn’t have moral judgment. Handling anything sensitive — privacy, ethical calls, how something actually affects a person, still needs a human involved. And that’s not just an engineering responsibility. Keeping AI aligned with real human values takes ML engineers, product leads, and UI/UX designers working together, not any one of us alone.

Key Takeaway: Looking under the hood didn’t make me want to become a machine learning engineer. If anything, it made me realize how much this work still needs designers in the room — not after the model is built, but while it’s being shaped.

AI is a powerful engine, but without real UX thinking behind it, it stays confusing or hard to trust. Understanding how training data, bias, and probability actually work is what lets us turn it into something that feels less like a black box, and more like something people can actually rely on.

What’s the part of working with AI that’s been hardest for you to design around?

( This is post two of the AIwithUX series. Still learning this in public, one concept at a time.)

#AIwithUX #ArtificialIntelligence #UXDesign #UIDesign #MachineLearning #AIDesign #ProductDesign #LearningInPublic #UXwithAI

Discover how AI actually works — #AIwithUX Series #02 was originally published in Bootcamp on Medium, where people are continuing the conversation by highlighting and responding to this story.