AI is everywhere, but so are its limitations.
The UX Days 2026 were structured around the theme " To design is to choose consciously The first day was entirely dedicated to workshops: hands-on sessions where participants rolled up their sleeves, tested methodologies, and made mistakes in a safe environment. The second day offered a full program of conferences, with speakers from around the world providing fresh, international perspectives on the topics that matter most to us right now.
Two very different formats, but a single common thread: how to use AI in a way intentional, rather than by simple reflex ?
The tools and use cases varied from session to session. But the message was surprisingly consistent. AI can help us to work fasterto explore further and focus on what matters. It can also make us lazier, produce inaccessible code and generate results that seem correct but are not.
The difference lies entirely in how we use it.
Accessible ≠ usable: the confusion persists
One of the most significant moments of these two days was the intervention of Mathieu Froidure, The director of Urbilog is blind. He has worked in the field of digital accessibility for over 30 years.
His message was simple, but it hit the mark.
"A website" technically accessible can absolutely be completely unusable »
He demonstrated this live with audio readings of complex data tables: an inaccessible version and a technically compliant version. The second was better. But it still generated such a cognitive load that a real user would struggle to extract any useful information from it.
Accessibility isn't a checklist you tick off at the end of a project. It starts at the design stage. It's a design issue, not a development issue. And the only way to know if something truly works is to test it with the people it's intended for.
His recommendation: integrate 5 to 10 users extremes in your testing process. Not as an optional bonus, but as a standard quality indicator.
AI doesn't solve accessibility problems, it amplifies them.
Anna Cook, Platform and Systems Designer at Microsoft, took the thinking further.
She came to shatter a growing myth: the idea that AI will solve the accessibility problem for us. Her response was unambiguous.
“AI does not solve accessibility.” ELLE amplifies what is already there. If your system is poorly structured, AI makes the situation worse.
WebAIM's data for 2026 confirms this: 95,9% home pages still exhibit basic accessibility flaws, with errors per page increasing by 10,1% from one year to the next and the increasing complexity of the pages 22,5%A significant portion of this increase is attributed to AI-generated code.
This is a crucial point, because the temptation to delegate accessibility to AI tools is very real, especially under time pressure. But an AI trained on faulty and inaccessible code will reproduce... faulty and inaccessible diagrams.
The solution is not to use AI less. It is to better structure. Accessible systems are well-structured systems. And well-structured systems are exactly what AI needs to function properly. The two reinforce each other, but only if the designers do the groundwork done beforehand.
The trap of cognitive laziness with AI
Bastien HuguesThe Head of Design at Bouygues Telecom asked the question that most of us avoid, I think: Is AI quietly making us... less competent ?
He wasn't alarmist. He was precise.
"Our brain is programmed to save energyHe's the real culprit.
The risk is not that AI will take our jobs. The risk is that we will gradually stop think deeply because thinking requires effort, and AI makes it easy to bypass this step.
He proposed a simple mental model: a " NutriScore "The use of AI ranges from level E (I ask, the AI executes) to level A (The AI challenges my thinking and pushes me further). Levels D and E are perfectly suited for repetitive, low-stakes tasks. For anything requiring genuine design judgment, aim for level B or A."
The framework I keep coming back to: AI is reproductive, not generativeIt makes predictions based on what already exists. That's genuinely useful for many things. But new ideas don't come from there. That part still belongs to us.
https://ici.radio-canada.ca/rci/fr/nouvelle/2178183/chatgpt-ia-intelligence-artificielle-cerveau-cognitif-mit
The real work begins before the prompt
Two workshops on the first day supported the same viewpoint from very different angles.
Yohann Elmaleh, a UX Writer at Bouygues Telecom, explained to us how he designed a AI writing assistant for his team, entirely trained on their internal content system.
The initial problem is one that many designers will recognize. When you ask a generic AI to rewrite a sentence with your brand's tone, it has no idea what it means. It doesn't know that Bouygues Telecom uses "vous" (the formal "you") and not "tu" (the informal "you"), that discounts are always written "-5€/month" and not "5 euros discount," or that theinclusive writing This means completely rewriting the sentences rather than simply adding a midpoint. The result looks correct, but it doesn't belong to anyone.
Yohann therefore built something different. He began by treating the designers on his team as users, asking them about their real difficulties with writing. From these interviews, he structured a content system with clear and precise guidelines, based on real-world use cases: general rules, tone, conversational style, accessibility, microcopy at the component level, and a glossary of brand-specific terms.
This documentation has become the knowledge base ofa personalized AI assistant, designed at thewith the help of a pre-prompt structure This defines the assistant's identity, its objectives, and the exact steps it must follow before generating any results. The result is an assistant that doesn't just produce text. It analyzes what you give it, flags what doesn't align with the brand, and offers several alternatives along with a UX recommendation tailored to the channel.
What makes it truly powerful is that it's not a generic tool with a thin layer of branding slapped on. It's built from the ground up on real guidelines, tested by real users, and continuously refined. As the content system is updated, the assistant improves.
L'atelier Figma Make animated by Romain Dao highlighted the same principle in a very concrete way.
The workshop focused on "vibe-coding" a mobile application using AI-generated components. And the most important lesson had nothing to do with prompting.
" The preparationThat's the work. The prompt is just theexecution »
Before we delved into AI, we worked on a moodboarda file of graphical guidelines and a user journey map. All of this was then provided to the AI.
Without this preparation, the results would have been generic and inconsistent. Thanks to it, the results were consistent, brand-friendly, and genuinely useful. The first prototype was obviously not perfect, but what's great about Figma Make is that you can copy and paste the generated frames directly into your Figma design file and make minor adjustments yourself. This gives you more control over the result, rather than speeding everything up and leaving total control of the output to AI.
The same principle applies everywhere. Whether you're building a writing assistant, generating interface components, or using AI to audit your design system, the quality of what comes out of it depends entirely on the quality of what goes into it.
The future of design is not rushed, it is decided
These two days have reinforced a conviction I already held, but which I now perceive more clearly: our role as designers is not diminished because of AI. It becomes more critical.
The tools are getting faster. The results are becoming more convincing. And that means that the decisions we make upstreamRegarding structure, accessibility and intention, these factors carry more weight than ever.
AI can help us do more. It can free up time for the thinking that really matters. But it won't decide to what mattersShe will not do it.advocate for your usersShe will not object when a design choice excludes someone.
That part still rests on us.
→ To learn more about AI and UX: ux-republic.com/blog
Aki Matsunaga, Senior Product Designer at UX-Republic
