The Most Common AI Mistakes UX Designers Make

Dovetail AI condenses interview quotes into neat themes, but this removes the stammers, contradictions, and emotional weight that signal where a user's actual problem sits. You design for the summary instead of the person.

The fix

Read at least 20 percent of raw transcripts yourself before looking at AI summaries, and flag any theme that contradicts another theme as a priority to investigate manually.

ChatGPT can build a persona from your research notes, but it flattens out the messy parts of actual behaviour. You then design for consistency instead of the real inconsistency of human choice.

The fix

Use AI personas only as a starting point, then overlay behaviour data (what people actually did, not what they said) to find the gaps.

When Maze tells you 78 percent of users clicked button A instead of button B, you feel confident and move on. You miss that the 22 percent might be the people who need your product most.

The fix

Always ask why the minority chose differently. Run a follow-up question in Maze or speak to those users directly.

ChatGPT can instantly list interaction patterns from fintech, e-commerce, and social apps. You adopt the pattern because it worked elsewhere, not because it serves your specific users' mental model.

The fix

For each pattern you consider, write down the assumption about how your user thinks. If you cannot name that assumption, do not use the pattern.

Adobe Firefly and Figma AI make it fast to move from research to wireframe, but the work of sitting with ambiguity and asking hard questions about what the data means is where your real design thinking happens. Speed removes reflection.

The fix

Set a rule: no AI tool touches your work until you have spent at least four hours with your notes on paper, asking what does not fit and why.

Figma AI proposes component arrangements based on design patterns, but those patterns assume a user goal that may not match your research. You inherit the hierarchy without asking whether it serves your user's actual task.

The fix

Before accepting any Figma AI layout, map it against the three main questions your user needs to answer in order, and reorder if those questions do not align.

Adobe Firefly can generate illustrations and backgrounds that look polished and inclusive. But visual language carries meaning about tone and intent. If you do not choose it yourself, you are not choosing the story you tell about your product.

The fix

Use Firefly as an option to compare against, not as your first choice. Create at least one alternative version manually to see what emotional meaning shifts.

ChatGPT can complete a microinteraction flow or suggest how a user should move through a screen. When you accept the suggestion, you lose the moment of choosing whether that interaction actually teaches the user something or just moves them forward.

The fix

Write out your interaction logic first without AI. Only ask AI to poke holes in your reasoning or suggest edge cases you may have missed.

A Figma AI prototype can look complete and functional, but you have not watched a user navigate it. The interface may be efficient for an AI-modelled user who knows what they want, not for a real person who is uncertain.

The fix

Test early and often with real users, especially on flows Figma AI generated. Watch for hesitation, not just success.

When ChatGPT or Figma AI suggest colour contrast ratios, heading hierarchy, or keyboard navigation, you can implement them without knowing why they matter for your specific users. You tick the compliance box but miss the real need.

The fix

For each accessibility recommendation AI offers, research one user who benefits from it and understand their actual workflow before implementing.

When Figma AI or ChatGPT proposes a solution, it feels neutral and authoritative. You skip your own judgement and adopt the suggestion because it sounds backed by pattern research. But it is optimised for average, not for your specific user.

The fix

Ask ChatGPT to argue against its own suggestion. If it cannot, you have not thought through the choice yourself.

Research synthesis through AI pulls out what users said they wanted. It cannot detect the needs they felt but could not name. This is where empathic design happens. AI efficiency removes the space for that kind of thinking.

The fix

Spend one hour per week alone with a single user's story. Write down what they wanted without saying it. What are they trying to become?

Dovetail AI or ChatGPT can summarise data without human bias, but they have their own bias built in by their training. You feel safer with the AI version and stop trusting your own pattern recognition.

The fix

When an AI summary surprises you, go back to the source. Your instinct that something is off is usually right.

AI tools encourage you to build coherent, repeatable design systems. But systems work best when you have solved a real problem for real users first. You end up with consistency that serves nobody.

The fix

Do not build a design system until you have solved the same problem three times for three different contexts and noticed what stayed the same.