What an AI moderator actually sounds like
A real Zoom transcript from one of weev's IDIs. No editing, no scripting. Probing rules, pacing, and what we got right (and wrong) on the first dozen interviews.
We've been running weev as a moderator on real qualitative studies since November. Here's a slice of one of her early IDIs — a 25-minute conversation with a lapsed user of a B2C app. Names changed, otherwise verbatim.
weev: Hi Marcus, thanks for hopping on. Before we dive in — is now still a good time, or do you need a couple of minutes?
Marcus: No, now's good.
weev: Perfect. So I want to start with the last time you opened the app. Walk me through it. What were you trying to do?
Marcus: Honestly I don't even remember. I think I was looking for something and I just… closed it.
weev: When you say you closed it — was that frustration, was that "I didn't find what I needed and gave up", or was it more "this isn't worth my time right now"?
That last turn is the one that mattered. The probing rule fired because Marcus's first answer was below a configured length threshold and contained low-information signal words ("I think", "just"). weev paused, reframed, and gave him three concrete shapes for his answer to sit inside.
His real answer came next. It was long, specific, and useful.
What surprised us
Pacing. We thought weev would feel rushed. She doesn't. The realtime voice model pauses naturally between sentences, lets people finish, and only intervenes when a clear silence opens up. People stop noticing she's an AI about three minutes in.
Probes work better than scripts. A flat list of probes ("can you tell me more?") is what cheap moderators sound like. Conditional probes ("when you said X, did you mean Y or Z?") are what good ones sound like. We coded the second kind.
The transcript is the deliverable. Auto-indexed into pgvector at the end of the call, searchable by meaning the second the call ends. By the time the human researcher opens their laptop, the themes are already drafted.
What we got wrong
Cold opens. The first ten interviews opened with "Hi, I'm weev, I'm an AI moderator from..." — and lost ~30 seconds of warmth. We removed the disclosure from the open and moved it to the consent step on the calendar invite. Engagement jumped immediately.