How to Use AI as a UX Designer, Not a Guesser
How to use AI as a UX designer: fast first drafts of flows and copy, and why a claim about what users actually do still needs real testing behind it.
A UX designer’s week splits into two different kinds of work filed under one job title: producing volume — a first-draft user flow, a spread of microcopy options for a button or an error state, a quick summary of a competitor’s onboarding screens — and judging whether a design actually solves the real problem a specific user has, based on what that user has shown you, not what a design pattern usually does in general. A general AI tool is genuinely fast at the first. It has no real access to the second: it was not in the room for your last usability session, and nothing about how confidently it describes "what users expect" tells you whether that is true for the people who actually use your product.
This sits alongside how to use AI as a market researcher, which covers the same volume-versus-judgement split for the interview and survey work a lot of UX research is built from. This one is about where that line falls once research turns into an actual screen.
Where AI genuinely helps
- A first-draft user flow or wireframe copy from a feature brief, for you to check against the actual constraint set before anyone builds it.
- A spread of microcopy options for a button, an empty state or an error message, for you to pick and sharpen rather than write from nothing.
- A first-pass summary of a competitor’s onboarding or checkout flow, screen by screen, so you start from what is actually there rather than rereading every screen yourself.
- Turning raw interview notes into a first-pass list of recurring themes, which you then verify against the actual transcripts or recordings.
- Drafting a first set of usability-test tasks or interview questions from a research goal, for you to check for leading or ambiguous wording before anyone runs a session.
Where it will quietly make things worse
The costly failure here is not an obviously bad draft. It is a confident sentence about "what users expect" or "a common usability pattern" that is really a generic best practice with nothing behind it specific to your product, your audience or the task you actually tested. A BCG and Harvard Business School study of 758 consultants found real performance gains on tasks inside a model’s current capability and a measurable drop on tasks that look similar but sit just outside it. Generating a spread of copy options sits inside that frontier. Telling you how a real user will actually react to a specific screen does not — a model has never watched someone hesitate at your exact sign-up flow.
A model can describe a dozen plausible-sounding usability patterns in a minute. It cannot tell you which one your actual users will struggle with until someone watches them try it.
Hallucination in large language models is well documented: asked about user behaviour or research findings, a tool will sometimes produce a specific-sounding statistic or a named study with nothing actually behind it, stated with exactly the same confidence as a verified one. In practice this shows up as an invented "70% of users prefer" line, or a described usability finding that sounds exactly like the output of a real test but was never run. A claim like that is dangerous here precisely because it reads like research.
A worked example
Weak: "Design a sign-up flow for a fitness app."
Better: "Here is our actual sign-up flow, our target user (a beginner who has never used a fitness app), and the two usability issues we found in last week’s testing session: people did not notice the skip button, and people dropped off at the permissions screen. Draft three alternative flows that specifically address those two issues. Do not claim any version is tested or preferred by users — flag anything you are suggesting from general design convention rather than from our actual findings." The second version keeps the model working from real data instead of a generic pattern, and makes clear which parts still need an actual test before anyone trusts them.
The confidentiality line
An unreleased feature’s designs, a usability session recording, or a research participant’s personal details should not go into a general consumer tool with no agreement covering how it is retained. Deciding in advance what research material and design work may and may not see is the core idea behind the NIST AI Risk Management Framework — a line to set before a screen or a transcript goes anywhere near a tool, not a cleanup after it already has.
Checking before anything ships to users
Build one fixed checkpoint before a design ships, not after support tickets show people could not find the button a model assured you was intuitive. Task-level research mapping real AI usage onto occupational data keeps finding that these tools help most with producing and formatting material against something already decided — a brief, a research finding, a constraint list — and help least with judging whether that material actually holds up against a person who was never in the conversation. For UX specifically, that means a generated flow or a described "common pattern" is a starting hypothesis, not a finding, until it has been in front of an actual user. Checking an AI answer when you are not the expert covers that discipline in general terms, and accuracy of AI: what a major study found covers the research behind why a confident description and a tested one are not the same thing.
Getting started this week
- Pick one recurring, mechanical task first — microcopy options, a competitor screen summary — because the output is instantly checkable against what you already know.
- Feed the tool your actual research notes and constraints rather than a general brief, and ask it to flag anything it is suggesting from convention rather than from what you gave it.
- Treat any claim about what users want or do as a hypothesis to test, never as a finding, until you have watched a real person confirm or break it.
- Write down, once, which research material and in-progress designs never go into a general consumer tool at your team.
The wider pattern this fits
This sits close to how to use AI as a content writer, which covers the same drafting-plus-checking split for the words that end up inside the flows designed here, and alongside what AI is actually bad at, which covers the general version of why a fluent description and a true one are not the same property in these systems. What jobs are safe from AI covers where design and research-heavy roles generally sit on that measured scale. The same instinct — name the exact constraint, do not trust an open brief to avoid inventing a finding on its own — is standard advice from OpenAI and Anthropic in their own prompting guidance.
Use AI to clear the mechanical grind — flow drafts, copy options, a competitor summary — fast. Do not let a fluent description of "what users expect" substitute for actually watching a real person use the thing, and never let an invented statistic or a described finding that was never tested reach a deck under your name.
Coursium teaches that checking habit directly: short lessons on your phone, a quiz that checks the point actually stuck, and a practice task in using and checking general AI tools well. Stay ahead of AI rather than trusting a flow because the explanation for it sounds reasonable.