Qualitative human-data collection

Numbers do not travel on their own. An insight moves between people when it carries a story, one person's experience that the data then supports, and it has to survive that retelling without degrading from one telling to the next. Alongside the measurement spine I want a way to hear people directly: short, in the moment, in their own voice. Not to replace the scaled reading, but to colour it and stress-test what it claims.

Short-form, voice-first, in-the-moment interviewing is neither new nor mine. It has been studied for decades and is sold commercially in several forms, some of them well funded. The edge is not the idea. It is how the idea is applied to get the best possible result for its purpose, and the rigour and honesty wrapped around it. Here is how I would apply it.

Reducing a person to a survey row strips out most of what makes their experience worth knowing. Instead of one long questionnaire I would ask a short series of purposeful, semi-structured questions and let people answer out loud. Think of it as a voice survey. The written survey has had a century to settle into a form people recognise and trust; the voice equivalent has not, and that gap is the opportunity.

Voice, not video, on purpose. People guard their face more than their voice, and a camera invites performance. Audio lowers self-consciousness and keeps attention on what is being said rather than how it looks. The obvious risk, transcription quality, is manageable by design: a short warm-up where the person reads the question back settles them in and confirms a clean transcript before the real questions begin.

The interviewer is removed. Plain written questions on a screen, so the person reflects on their own experience rather than managing an interviewer or the medium. This is a deliberate choice against AI-moderated interviewing. Conversational AI can probe for elaboration at scale, and early work suggests the depth can approach a human interviewer (Barari et al., 2025). But a machine asking how you feel creates a dissonance that some people quietly dislike, whether or not they say it, and that can contaminate what you learn. It is not an experience worth giving people when a well-built question on a screen does the same job and leaves them alone with their own thoughts.

The structure stays light. Simple, rule-based branching is enough: if someone describes a good experience, ask them to say more about it; if a poor one, follow that instead. There is no need for a fully responsive AI conversation, and not building one keeps the focus on capturing honest data rather than perfecting a chat or managing all the ways it can go wrong. Where AI earns its place, in transcription, structuring, and analysis, it stays in the background. There is a strong and often well-founded wariness of AI in research, and no reason to put it in front of the person.

Speaking is harder than ticking a box, and that is the point. People hear themselves think and reflect as they talk, so I would give them more time to consider rather than rushing them. Much of the craft is in the warm-up: making people comfortable enough to speak, rewarding them properly for it, and being honest about the exchange. The evidence on incentives is clear that small, prepaid, direct payments beat large promised rewards (Church, 1993; Singer & Ye, 2013), and the low-grade anxiety of whether you will actually be paid is part of what makes research feel grim. Ask for more, pay better, pay directly.

The richest version is repeated and in the moment rather than a single sitting. Experience sampling and diary methods ask people to reflect inside their own lives over days or weeks, which reduces the distortion of memory and shows how feeling actually shifts (Csikszentmihalyi & Larson, 1987; Stone & Shiffman, 1994; Bolger, Davis, & Rafaeli, 2003). Short, frequent voice pulses fit that shape, and the repetition is itself a quality check, because real engagement is visible over time.

Where this sits in the method is as a triangulation layer. It colours the scaled signal with real voices and tests whether findings hold up against how people actually talk. It does not need a perfect panel to be useful. Purposive, honestly-described sampling is the right tool when the point is to reach people who do a particular thing (Baker et al., 2013). The aim is a rounder set of signals, and a result that can be handed to someone else and still mean something.

None of this is built yet, and the honest answer is that the best form will come from running the experiments, not from arguing them out in advance. That is the work, and it is the half I find most interesting.