UX Design with AI ┃ 1.2 LLMs for UX Design

Module 1. AI Fundamentals for the UX Workflow

Everyone's using LLMs. Almost nobody knows what's actually happening when they hit enter.

Let's fix that.👇

What Is a Large Language Model?

A large language model (LLM) is an AI system built to understand and generate human language. Because it has been trained on enormous volumes of text, it can produce sentences that fit a given context, answer questions, and condense long documents into their essentials.

The easiest way to picture it? Think of an LLM as an extremely smart autocomplete.

That analogy matters more than it might seem, so let's be precise about it. When we say an LLM "understands" language, we don't mean it interprets meaning the way you do. It isn't drawing on lived experience or grasping what words refer to in the world.

What it's actually doing is prediction. Based on patterns absorbed from its training text, it calculates which words are most likely to come next given the context you've provided — and generates a response from that.

This is why LLMs produce writing that reads so naturally and sounds so convincing. And it's also why that writing isn't automatically accurate, or automatically right for your users.

Fluent is not the same as correct. Persuasive is not the same as grounded.

Which leads directly to the working principle of this entire course: whatever an LLM generates, a designer still has to review it, test it against real user context, and shape it toward the design's actual goal.

What LLMs Are Actually Good At

Let's get concrete about the capabilities you'll reach for in real UX work.

• Text generation — Emails, explanations, landing page copy, interface copy. Anything made of words.
• Summarization — Compressing long documents and research materials down to what actually matters.
• Translation — Converting text between languages, and refining how something is phrased.
• Question answering — Answering questions based on material or context you provide.
• Information extraction — Pulling recurring needs, pain points, and repeated phrases out of interview transcripts and user feedback.
• Classification — Sorting user feedback and VOC data by theme.
• Structuring — Turning scattered notes into tables, checklists, and frameworks.
• Ideation — Generating and expanding ideas and concepts around a design problem.
• Code generation — Writing simple code or building quick prototypes.

You don't need to master the technical details behind any of these. What you do need is a working sense of what an LLM does well, and where your review is non-negotiable. Get that calibration right, and you'll use AI far more effectively in your design work.

The video below — Large Language Models explained briefly — is the clearest explanation of this topic I've come across. Watch it first, then continue with the lesson.

Where LLMs Break

An LLM is a genuinely powerful tool. It's also not a complete one — and in UX work specifically, four limitations demand your attention.

1. It Doesn't Actually Understand

An LLM does not understand the world the way a person does.

You interpret meaning and context through experience. An LLM generates plausible-sounding responses from statistical patterns in text. Those are fundamentally different operations.

So the output can read as smooth and authoritative while being disconnected from your actual users' context — or simply unsupported by any evidence at all.

☑️ Concept Check

• Hallucination — when a generative AI produces content that is factually wrong or entirely unsupported, and presents it in a plausible, confident tone.

The confidence is the dangerous part. A hallucination doesn't announce itself. It arrives sounding exactly as certain as the truth does.

2. It Inherits the Bias in Its Training Data

An LLM reflects the biases baked into the text it learned from. If stereotypes about a particular group, gender, occupation, or culture exist in that data, they can surface in the model's output.

For UX designers, this is not an abstract concern. It's a practical one, and it shows up in your deliverables.

When AI generates a persona, a user scenario, or a piece of copy — ask yourself: has it quietly generalized around a specific age, gender, occupation, or cultural background?

You are the one responsible for checking whether your work reflects the real diversity of your user base, and whether it's actually grounded in your research rather than the model's assumptions.

3. It Can Only Hold So Much at Once

There's a ceiling on how much information an LLM can consider in a single pass. That ceiling is called the context window.

This has an immediate practical consequence. Feed it a long research report or a full interview transcript, and important details from the beginning may not be adequately reflected in what comes back.

The fix is straightforward: break long material into sections, summarize each one separately, then bring the pieces together at the end to reconstruct the whole picture.

❇️ Tip

• With long materials like research reports or interview transcripts, don't ask for one summary of everything. Organize by topic or by section first. Then lay those summaries side by side and look across them — that's where recurring problems, real user needs, and genuine insights start to surface.

4. Ethical and Copyright Concerns

An LLM can generate content with unclear sourcing or copyright issues. It can also produce false information presented as fact.

And there's a subtler risk that matters enormously in product work: AI-generated output can mislead users — or make claims about value your product doesn't actually deliver.

So AI output has to be reviewed. Always.

This is especially true for anything that shapes a user's judgment: research summaries, user analysis, product copy, policy statements. A human being has to verify that content and take responsibility for it.

That practice has a name.

• Human in the Loop (HITL) — keeping a person in the decision path, responsible for reviewing and validating what the AI produces.

It isn't a formality. It's the thing that separates a designer using AI from a designer being used by it.

What's Next

So that's the LLM. Powerful, fast, occasionally full of it — and now you know exactly where to look.

Next, we go deep on the other side of the toolkit: image generation models. What they're really doing, how to use them in a UX workflow, and why "make it pretty" is the worst prompt you can write.

Follow along so you don't miss it. See you in the next one.

Thank you! 👍