Table of contents
What is actually different in 2024
Design tools now generate. Figma's AI features and Uizard turn a written brief into a starting wireframe in under a minute. The output is rarely the final design, but it skips the blank page. For designers who have spent careers staring at empty Figma frames, that alone is worth paying for.
Copy is fast and contextual. ChatGPT and Jasper produce microcopy variants for buttons, errors, and empty states in seconds. The teams that move fastest write a one-line context, get back ten variants, and pick one. The teams that are still slow are the ones writing the copy themselves between design reviews.
Personas can be data-driven instead of imagined. Tools that ingest real user research and generate persona profiles produce something less fictional than the standard "Sarah, 35, marketing manager" template. Imagined personas described what we wished users were like. Data personas describe what users actually do.
The video below covers where AI tools still fall short. Worth watching before getting overinvested in any of them.
What AI does well today
It detects patterns in big data sets. Session recordings, support tickets, survey responses condense into themes faster than a researcher reading them by hand. Accuracy is reasonable for surface patterns. Deep insights still need a person who can spot the weird thing in the data and ask why. AI flags that something is happening. It rarely tells you why.
It writes good microcopy. The example everyone cites: "Your basket is waiting for you" becomes "Your basket is ready. Don't miss your favorite items." Both work. The second creates more urgency. Pick by context. The deeper use is generating ten variants of an error message in three seconds and picking the one that fits your brand voice. That used to be a thirty-minute exercise. Now it is a paragraph of prompt and a coffee.
It drafts personas in minutes. Feed real user research into an LLM and ask for three personas with citations. The output takes another hour to edit. That hour still beats four hours from scratch. The trick is the input: feed it real research notes, transcripts, support tickets. Feed it nothing and you get a recycled stereotype. The AI is only as useful as the data you give it.
It generates starting wireframes. Uizard and Figma's AI features take a description and produce a layout. Generic, but the blank page problem is solved. The wireframe will be wrong in important ways. That is fine. It is faster to react to a wrong wireframe than to start from nothing.
It accelerates usability test analysis. Maze and UserTesting now ship features that flag friction points across hundreds of sessions, picking up patterns a single human reviewer would miss in the volume. The ratio of insights per hour of analysis went up by a factor of five for some teams. Whether the analysis is correct is a different question, and that is where humans still need to look closely.
Where AI is still bad
Visual judgment is missing. AI cannot tell whether a layout feels right. A senior designer can spot a generated UI in seconds: alignment that is almost-but-not-quite consistent, spacing that is mathematically correct but visually wrong, color choices that hit accessibility minimums and nothing else. The output is technically correct and emotionally flat. That gap between technically correct and actually good is most of what design is. Models are not yet on the right side of that gap.
Research that surprises you. AI synthesizes what is already in the data. It cannot ask the unexpected follow-up question. It cannot read body language. The most valuable research moments are unplanned, and AI is the opposite of unplanned. The user who hesitates for ten seconds before answering tells you something the transcript will not. The user who looks confused but says "yes, I get it" is signalling a usability problem the words do not capture. AI reads the words, not the silence.
Edge cases and error states. Generated UIs default to the happy path. They forget what happens when the API is down, when the user has no data yet, when something is loading slowly, when permissions are missing. These states are often what makes a product feel polished or broken. The error state for a billing flow is more important than the success state, because the success state is where everything works and the error state is where the user is already frustrated and needs help.
A product with personality. Generated UIs are confidently average. Memorable products are slightly weird in some specific way, and weird needs a human who chose it on purpose. Linear's keyboard-first interface, Notion's blocks, Stripe's documentation tone. None of these would come out of an AI prompt. They came out of teams making opinionated choices that the average product would not. AI is the opposite of opinionated. It produces the median.
The ethics part nobody wants to talk about
Two real problems and one fake one.
The first real one is data leakage. If you feed user data into a hosted LLM, that data may train future models. Most teams do not realize this until legal asks. Use enterprise tiers with no-training agreements, or run models locally for sensitive data. The NN Group review of AI design tools covers some of the privacy questions worth asking before you adopt one. Specifically: where does the data go, who can access it, what happens if the vendor pivots or shuts down, what are your obligations under your customer contracts. None of these are hypothetical. All of them have caused real incidents in the past two years.
The second is bias amplification. AI trained on existing UIs reproduces existing patterns, including patterns that exclude users: small touch targets, low contrast, missing language switches, copy that assumes Western context, screen-reader-hostile flows. AI does not fix accessibility problems. It often deepens them. The model has seen a million inaccessible designs and zero accessible ones with the right metadata. It will produce more of what it has seen. The cure is simple and unsatisfying: review every AI output for accessibility before shipping. The tools that audit this exist. Most teams do not run them.
The fake one is the existential "AI is replacing designers" panic. AI is replacing some tasks. Production work, layout drafts, copy variants, pattern analysis. None of those tasks were what made a designer valuable. The designers who lose their jobs are the ones who never did anything else. The designers who do strategic work, build trust with cross-functional teams, and exercise taste are not replaceable by current models, and the models are not on track to replace them in any timeline I can see. The panic is loudest among designers whose careers were built on the work AI is best at automating, which is not surprising.
What to do this year
Use AI for the production tier. Wireframes, copy, persona drafts, test analysis. Skip it for the parts where you needed to be in the room: research conversations, design critique, reviewing edge cases, deciding what to ship. The teams that do this well have a clear mental model of which tasks are "AI-first" and which are "human-first" and stop arguing about it.
Set up a workflow. Mine looks like this: AI for the first draft of anything I write, anything I sketch, anything I summarize. Human review for anything that ships, anything that involves taste, anything that involves trust. The boundary moves slightly every quarter as the tools improve. The structure stays.
"The only way to win is to learn faster than anyone else." — Eric Ries, The Lean Startup
If you only learn one tool this year, learn the one your team actually adopts. The AI ecosystem changes every quarter. Tool fluency is less valuable than the judgment to use any tool well. The designer who knows three tools deeply will out-ship the one who knows ten tools shallowly, every time.
The longer game: AI changes what designers should be hired for. The interview question is no longer "show me a beautiful Figma file." It is "show me how you decide what to build." The first one will get cheaper to fake every year. The second one is harder to fake than ever, because it requires actually knowing what you are doing.
