These reflections draw on several UX research studies I have conducted this year across different industries, each broadly aimed at optimising the value of AI features.
During recent UX testing of AI prototypes across different sectors, I’ve observed users engaging with AI-powered assistants designed to support decisions and simplify complex tasks.
What stood out was how easily trust could be won, or lost, depending on how the AI offer was presented, its design, and its behaviour.
1. Users Didn’t Want AI to Guess
Across multiple studies, users tended to trust AI that felt genuinely curious about their needs. When an AI reached a conclusion too quickly, users often questioned whether it truly understood them, even when it turned out to be correct.
They responded better when it felt as though the AI was thinking with them, not for them.
It was the difference between a rushed answer and a considered one, and that subtle distinction had a big impact on perceived intelligence and trust.
2. Clarity Built Trust
Across studies, users were more comfortable when it was made clear what the AI was doing and why. When its role or purpose felt vague, people struggled to judge whether to trust it. Users wanted to understand, in simple terms, what the assistant would help with and what to expect next.
Names and descriptions that were straightforward and human-centred helped people relax, while technical or abstract labels created more distance. The more clearly the AI defined its purpose, the more credible it felt.
3. Tone Shaped Credibility
Across studies, tone played a bigger role in trust than many expected. Users tended to respond well when the AI language felt clear, confident, yet grounded. When the tone became overly familiar or self-assured, some questioned its authenticity and took it less seriously.
The wording and attitude of a message could completely change how intelligent the AI seemed. People preferred a tone that signalled competence without assuming too much — confident, but not overbearing.
4. Design Cues Shape Trust
The format of the interaction had a noticeable impact on how users felt about the AI. Subtle differences in layout and flow changed whether users perceived it as approachable or mechanical; a reminder that presentation can influence trust as much as content.
5. Control Created Comfort
Across multiple studies, trust often depended on whether people felt informed and in control. When they clearly understood what was happening and could influence the outcome, their confidence in the AI increased. But when the process felt ambiguous or decisions seemed to happen automatically, hesitation set in.
Users appreciated systems that kept them involved; tools that felt collaborative rather than directive. They liked smart support, but they still wanted to steer the experience themselves.
In a financial tool, this sensitivity to control became even more pronounced. Participants needed clearer confirmation before any automated action that might affect their money, usage, or costs. It reinforced a broader truth: in important areas, people are comfortable with AI assistance, but only when they remain the final decision-maker.
6. Novelty Wasn’t Value
Across multiple projects, features that appeared clever but didn’t reflect real behaviour often prompted scepticism. Users quickly recognised when something was for show rather than genuine usefulness.
In contrast, tools that solved small, everyday frictions were seen as intelligent and worthwhile. The same pattern appeared in other contexts; from smart-home apps to shopping assistants, users could tell when a feature added little practical value.
Novelty captured some attention, but usefulness sustained trust.
7. AI Still Has to Be Usable
Not every issue I observed was uniquely about AI. Users could still struggle with familiar UX problems: understanding what to do next, completing key tasks, interpreting feedback, finding controls or recovering when something went wrong.
AI-specific questions around trust and expectations matter, but they sit inside the wider product experience. An AI feature can be useful and credible in principle and still fail because the surrounding interaction creates unnecessary friction or confusion.
8. What This Means for Design Teams
From these sessions, a few principles keep surfacing:
- Curiosity built confidence.
When systems explored before concluding, users saw them as more thoughtful and intelligent. - Transparency fostered trust.
People valued brief explanations of how suggestions were formed rather than silent accuracy. - Tone mattered.
Language that was warm and factual created reassurance, while overly casual or promotional wording reduced credibility. - Design cues shaped perception.
Subtle signals of ongoing exchange made interfaces feel more intelligent and supportive. - Control created comfort.
Users were most confident when they could see options and make adjustments along the way. - Value outweighed novelty.
Practical features that saved time or clarified decisions built more trust than clever extras. - Usability still mattered.
Users needed to understand what to do, complete important tasks without unnecessary friction and recover when either the interface or AI behaved unexpectedly.
9. From Intelligence to Permission
The challenge wasn’t simply to make AI appear smarter; it was to help products earn permission.
That permission lived in the small details: the phrasing of a question, the confidence of a tone, the transparency of a hand-off.
These issues emerged across AI user testing studies, from retail to home energy, and proved remarkably consistent.
A trustworthy AI didn’t need to sound human. It just needed to be clear about what it was doing.
Even so, every AI behaved and was perceived differently depending on its context and audience.
An assistant in retail didn’t evoke the same reactions as one embedded in a transactional flow. That’s why broad generalisations are interesting but not a guide.
What mattered was observing AI in realistic contexts, with the right people, and noticing how tone, transparency, and control played out in practice. Because the truth about AI trust wasn’t found in theory, it appeared, quietly, in the moment a real user either leaned in… or stepped back.
I can help
I use moderated research with real users to uncover how people understand, use and respond to AI products, including usability problems, misplaced trust, unmet expectations, workarounds and areas for improvement.
User testing can be useful from early concepts and prototypes through development and into live products.
Focused projects start from £2,750 + VAT.
AI Product User Testing: https://userfy.co.uk/testing-ai-products/
Email: phil.randall@userfy.co.uk
Call: 07712669935
Phil Randall (Owner at Userfy)
www.userfy.co.uk
