What does AI confidence actually look and feel like?
And what would someone do or say differently once they have it?
A research project at Berghs, run together with four other UX designers. Between us we interviewed nine designers about how they actually use AI. Access to AI tools was never the problem. What people struggled with was knowing when to trust AI, when to question it, and when not to use it at all. We set out to understand what AI confidence really is, and why it matters right now.
Understand how designers experience AI adoption day to day, identify the emotional and cognitive barriers, explore how people evaluate and trust AI output, and map the gap between learning the tools and developing judgment.
Nine designers, aged 27 to 48, from student to senior level, working in a bank, a government organisation, an agency, as a freelance consultant and in in-house product teams. We deliberately sampled across the adoption spectrum: skeptics, occasional users and enthusiasts.
Semi-structured interviews, 30 to 60 minutes, built around ten themes from current behaviour to desired future. Each interview was clustered into its own affinity map, then themes were compared across all nine.
Desk research, two of the nine interviews, affinity mapping and insight writing.
The question wasn't whether designers can use AI. It was whether they know when to trust it, when to question it, and when to leave it alone.
The gap we were looking for was emotional and cognitive, not technical.
What we asked in the interviews.
- Definition. What is AI confidence, and what behaviours signal it in real work?
- Barriers. At what moments, and in which tasks, does uncertainty arise?
- Trust & judgment. How do people decide whether an AI output is good, correct or usable?
- The gap. What separates knowing how to prompt from knowing when to trust the result?
- Learning. How do people learn AI, and how do peers and culture shape their confidence?
- Context. What support helps, and when do people feel safe to experiment?
- Outcomes. What changes in someone's work once they become more confident with AI?
AI is everywhere, but real fluency is rare.
The gap isn't interest. It's support, guidance, and room to practice.
Source: Google × Ipsos, “The Path to AI Fluency”, 2026
'AI fluent' in that study means using AI regularly and understanding what it can and can't do, not expertise in any one tool.
Confidence is low, even among users.
Only a small share feel sure they get the most out of AI. Many use it without knowing how to use it well.
Guidance is missing.
Most people get no formal training or clear guidance. The problem isn't motivation, it's missing support structures.
Learning is social.
People turn to colleagues rather than formal systems. Fluency grows where experimentation is supported.
What really stands in the way.
Emotional gap.
Fear, shame, identity threat
- People find AI useful, but worry about how their data is used.
- Many fear that AI will take their jobs.
- Some feel AI works against careful decision-making, and are disappointed by it.
- AI can make people less willing to speak up or admit mistakes.
Cognitive gap.
Skills ≠ judgment
- Using AI well takes new skills: writing good prompts and judging the results.
- Organizations haven't yet changed how they think to work well with AI.
- The biggest risk is believing you understand AI before you really do.
HOW THE TWO GAPS FEED EACH OTHER
Fear blocks practice → no practice → no judgment → more fear.
Nine designers. Every one of them drew a line somewhere.
We deliberately talked to people across the adoption spectrum, from a freelancer who won't let AI near final work to a designer who calls it a superpower. They disagreed about almost everything except one thing: where the human has to stay in charge.
- Confidence means knowing when NOT to use it. Nobody defined confidence as tool skill. They defined it as judgment, and several drew a sharp line between feeling confident with AI and actually making better decisions.“AI confidence, to me it means knowing when not to use it.”UX designer, government organisation
- Polished output is the risk. A fluent, well-formatted answer gets less scrutiny than it deserves. Participants described having to learn not to let a confident answer outweigh their own analysis.“It's so easy to just throw something in there and get something that looks good out.”UX designer
- Judgment is built by use, not by reading about it. The most experienced users described an internal map built from failures: what works, what breaks, how to steer it. That map only comes from using the tool and watching it fail.“You still need to apply your own judgment. You cannot give everything away to AI.”UX designer & developer
- Everyone draws a line, in different places. Early exploration, yes. Final deliverables, no. The line moved depending on role and risk, but every participant had one.“I don't think I would use it for final deliverables.”Visual & product designer
- The time saving isn't always real. Several described AI slowing them down: re-prompting, correcting, or being handed so many options that deciding got harder.“I feel like I ended up spending way more time trying to get it right than just doing it myself.”Visual & product designer
- Support stops after the launch. Workshops at rollout, then nothing. People learn alone, through trial and error, and several said they'd never openly discussed how colleagues actually use AI.“Mostly by myself.”UX designer, on how she learned AI
WHAT DIDN'T MATCH THE RESEARCH
One thing our interviews contradicted: the research says fluency brings both productivity and confidence. We met people who felt faster but less sure.
Not a smarter assistant. A more honest one.
- Memory that carries over. Work is scattered across separate chats and sessions. People wanted something that holds the project context instead of starting from zero every time: “a continuous system of memory and context, not a series of isolated interactions.”
- Visible uncertainty. Several asked for the AI to evaluate its own answers and show how sure it is, rather than delivering everything in the same confident voice.
- Checkpoints, not one big answer. On complex tasks, people wanted the tool to pause and confirm it understood before producing: to ask first, then execute.
- Challenge, not agreement. The most repeated wish across interviews: an AI that pushes back. One participant wanted “a colleague that asks annoying questions and challenges assumptions.”
AI can help at every step. Judgment is still needed at every step.
The eight phases are a standard UX process, not something that emerged from the interviews. We used it as a frame and marked, phase by phase, what AI can take on and where a human still has to decide.
Our workflow map, straight from the Figma board.
Four habits that help people use AI with confidence.
- Intentional use. Regular, deliberate AI use in real tasks, not occasional experiments.
- Knowing the limits. A clear sense of when AI is appropriate, and when it is off-limits.
- Double-checking. A habit of verifying outputs before acting on them.
- Owning the decision. It feels like augmentation, not replacement, and you can explain every AI-assisted decision.
AI fluency isn't learned in a course. It's built by doing.
Research backs up what we saw: fluent people don't know more, they use AI more, testing it in real work. The biggest barrier isn't motivation, but missing time, guidance and a safe place to experiment. Teams that learn together get there faster.
Source: Harvard Business Publishing, “Learning Through Experimentation”
That's what I take into my own work: judgment comes from practice, not from more tools.
Our own user interviews with UX designers · Google × Ipsos, The Path to AI Fluency (2026) · Harvard Business Publishing, Learning Through Experimentation · Worklearning.ai, AI Fluency in Learning and Development · NN/g on evaluating AI-generated designs · Master's thesis on how UX designers evaluate and override AI-generated designs · Articles and community discussion on human-in-the-loop AI workflows