AInterviewer: A Platform for Designing and Conducting AI-led Qualitative Interviews
The paper introduces AInterviewer, an open-source, multi-agent platform that combines the structured control of survey software with the flexibility of locally hosted Large Language Models to enable secure, reproducible, and standardized AI-led qualitative interviews.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you want to ask a hundred people deep, personal questions about their lives. Traditionally, you'd need to hire a hundred skilled interviewers, pay them, and hope they all ask the questions in the exact same way. That's expensive and hard to manage.
Recently, scientists tried using "AI robots" (Large Language Models) to do this interviewing. But most of these robots are like black boxes: they are owned by big companies, you can't see how they think, and you can't control exactly what they say. This makes it hard to trust the results or keep the data secret.
Enter "AInterviewer."
Think of AInterviewer as a custom-built, open-source interview factory. Instead of handing the whole job over to a single, unpredictable AI robot, the researchers built a team of specialized AI workers who follow a strict, transparent script.
Here is how it works, using some everyday analogies:
1. The Blueprint (The Interview Guide)
Before the interview starts, a human researcher draws up a detailed blueprint, called an Interview Guide.
- The Main Questions: These are the "must-ask" questions, like the main courses of a meal. They must be asked to everyone to ensure fairness.
- The Probes: These are the "follow-up questions." Imagine a detective asking, "Can you tell me more about that?" The system is smart enough to generate these on the fly based on what the person just said, keeping the conversation natural.
2. The Assembly Line (The Multi-Agent Pipeline)
This is the secret sauce. Instead of one AI doing everything, AInterviewer uses a team of three specialized AI agents working together like a well-oiled machine:
- The Traffic Cop (Classification Agent): Before asking a question, this agent checks the conversation history. It asks: "Has this already been answered?" or "Did the person refuse to answer?" If the answer is yes, it signals the team to move on or change tactics.
- The Translator (Reformulation Agent): If the person has already talked about a topic, this agent rephrases the next question so it doesn't sound repetitive. It's like a host saying, "Since we just talked about your job, let's dive deeper into how you use AI there," rather than just asking, "Do you use AI at work?" again.
- The Curious Listener (Probing Agent): This agent listens to the answer and generates the perfect follow-up question to get more details, just like a skilled human interviewer would.
3. The Safety & Control Features
- Local Hosting: You can run this system on your own computer (like a local server) instead of sending data to a big cloud company. This is like keeping your diary in a locked safe in your house rather than posting it on a public billboard. It ensures privacy and security.
- Open Source: The code is free for everyone to see and check. There are no hidden tricks. If you want to change how the "Traffic Cop" works, you can tweak the code yourself.
4. The "Test Drive" (Pilot Testing)
Before sending the interview to real people, researchers can run a simulation. They can set up "fake" AI interviewees with different personalities to test if the questions make sense. It's like a dress rehearsal for a play before the audience arrives.
5. What Did They Find? (The Pilot Study)
The team tested this system against real human interviewers with 40 students.
- Length: The AI interviews were actually longer in total because the AI never got tired or needed a coffee break to think of the next question.
- Quality: The answers were just as relevant and specific as those given to humans.
- The Catch: While the AI was great at keeping the conversation going, the researchers note that human interviewers still had a slight edge in getting very long, detailed answers to specific questions. However, the AI was close enough to be very promising.
The Bottom Line
AInterviewer is a tool that lets researchers conduct large-scale, deep conversations with people using AI, but with the control of a survey and the flexibility of a chat. It solves the problem of "black box" AI by making the process transparent, secure, and customizable, all while following the best rules of social science research.
Important Note: The authors are clear that this is a new method. It is currently text-only (no voice yet), and it is not ready for interviewing vulnerable people (like those in crisis) because it lacks specific safety guards for those situations. It is a powerful tool for research, but it is still a work in progress.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.