Field Notes
What I keep learning about how people come to trust AI
For the past few years, most of my research has been about one question: when do people trust an AI system enough to rely on it, and what makes that trust break? Across studies of assistant memory, AI research tools, developer platforms and personalisation, with participants in India and the US, the same patterns keep coming back. These are the ones I now bring into every study.
- 01
Show the plan before the work
People trust what they can inspect. When an AI tool laid out its plan before acting, users felt in control, even when they changed nothing. When a personalisation feature let users see exactly which apps it would connect, most looked first and then agreed to nearly everything. The look was the point.
In practice: I test the moment before commitment as closely as the output itself.
- 02
Invisible work reads as broken
Silence is the fastest way to lose a user. A long wait with no sign of progress felt like failure, even when the result was good. A memory that didn't visibly update felt like a memory that didn't work. Connected data that never showed up in an answer felt like data handed over for nothing.
In practice: I look for every moment where the system is working and the user can't tell.
- 03
Accurate isn't the same as useful
An answer can be correct and still worthless. Reports built from hundreds of sources still read as generic. A personal summary could describe someone perfectly and still feel "like something an astrologer would say." Users don't reward accuracy alone. They reward answers that move them forward.
In practice: I separate "was it right?" from "did it help?" in both tasks and analysis.
- 04
Relevance is the real intelligence
Trust broke less often on what a system knew than on what it chose to use. Assistants brought a hobby into a job application, trusted a location guess over a city the user had stated many times, or treated one passing question as a defining trait. Knowing a lot about someone matters less than knowing what matters right now.
In practice: I code failures by why the system chose the wrong signal, not only by whether the answer was wrong.
- 05
Ask, don't assume
People wanted AI to ask them questions, one at a time and in conversation. When a detail was missing, a single clarifying question beat a confident guess every time. And when the system got something wrong, people corrected it in the next message, not in a settings page, and expected it to remember.
In practice: I design tasks that leave gaps on purpose, to see whether the system notices.
- 06
Consent is granular, and cultural
Trust is never all or nothing. People happily shared some things and guarded others: photos, finances, health, other people's information. Where that line falls depends on context. In India, a cloud drive often holds Aadhaar and PAN copies, which changes what "connect your files" really asks of someone.
In practice: I map sensitivity source by source, and check what everyday tools actually mean in each market.
How I work
Over time, not in one sitting.
Trust builds and breaks across days, so I favour diary studies that follow people through weeks of real use.
Their tasks, not mine.
Wherever possible, participants bring their own prompts and problems, so we test the product against real life, not a script.
Watch the hesitation.
Ratings tell you what people concluded. Pauses, retries and workarounds tell you why.
Context first.
Who someone is, where they live and how they store their life shapes what trust means. I recruit and analyse with that in mind.
Good AI research doesn't ask whether people like the product. It asks whether they would let it into their lives, and on what terms.