super interesting. We are working on new cohorts in the refugee support movement in the wake of the Ukrainian refugee influx. Think your model here can help clarify some stuff!

Looks like a super useful application.

If you want to share your story--or share what you'd say to the people responsible for the current changes in the US--you can do so at the link below. And please share widely! It's essential that we document the human toll.

tinyurl.com/thehumantoll

We draw on prior efforts of @felixchopra.bsky.social,
@kunkakom.bsky.social, @ihaal.bsky.social, Friedrich Geiecke and Xavier Jaravel.

Thanks to my co-authors Semra, Tobias and @andbjn.bsky.social and to the @villumfonden.bsky.social foundation and @cphsodas.bsky.social who have both founded the AI-interview project.

Of course, there are big challenges to this kind of experimental comparison 🧵

1) Data quality in qual interviews isn’t standardized
2) It’s rarely operationalized as quantitative measures
3) Chat-based interviews ≠ in-depth, face-to-face interviews
4) Small N

Important limitations to keep in mind.

📄 Read the preprint: osf.io/k7f89

If you're working on or curious about AI-interview methods, feel free to reach out (DMs open)!

This is part of a broader project developing the AI-interviewer, led together with @annarogers.bsky.social

Why this matters 🧵

1) AI-interviewers can yield reasonable data quality, supporting their promise for large-scale qualitative research.

2) As opposed to other AI-interview studies, we used a locally hosted, open-weight model—boosting data security, transparency & reproducibility.

Key results 🧵:
🧠 Human-led: longer answers per question
🕐 AI-led: longer interviews overall (faster question delivery)
📋 No dif in response quality (specificity + relevance)

LLMs do pretty well—
But responses are short compared typical qualitative interviews.