AI-Led Interview Toolkit for Qualtrics

An interview led by a large language model, run inside an ordinary Qualtrics survey. This is the module we built for Is Money Overrated? Misperceived Satisfaction from Income (Perez-Truglia and Rubião, NBER Working Paper No. 35423), packaged so it can be dropped into someone else’s survey.

Repository · Qualtrics template (.qsf) · Implementation guide (PDF)

What is in the toolkit

AI_Interview_Module.qsf — a minimal importable Qualtrics survey containing only the interview module: the situation question, three conversation rounds, decision generation, the validation step, and two example outcome questions. It contains no credentials; the OpenAI key is the placeholder YOUR_OPENAI_API_KEY.

The implementation guide — setup, customization, what happens when things fail, a data dictionary, and the four interview prompts reproduced verbatim. It is written for researchers who are comfortable with Qualtrics but have not worked with language models before.

The prompts — the four system prompts as plain text files, so you can read or edit them without opening the survey.

build_qsf.py — the script that regenerates the template from our fielded survey. It is there for reproducibility and is not needed to use the toolkit.

What you need

A Qualtrics account that allows JavaScript on questions (most academic licenses do), an OpenAI API key, and a small budget for API calls.

Basic use requires no programming. You import the template once, paste your key into a single field in the Survey Flow, and edit the interview prompts as plain text. The model name and all four prompts are Survey Flow embedded-data fields, so changing the interview does not mean touching any JavaScript.

Building your own

If you want to construct a Qualtrics survey like this from scratch rather than adapt ours, the QSF Builder skill writes .qsf files directly — including the question JavaScript for API calls, embedded data, and display logic.

Citation and credits

If you use the module, please cite the paper it comes from:

Perez-Truglia, Ricardo, and Rafael M. Rubião. “Is Money Overrated? Misperceived Satisfaction from Income.” NBER Working Paper No. 35423, July 2026.

To pass messages between the survey and the model, the template adapts the Qualtrics–JavaScript code of Costello, Pennycook, and Rand (2024), “Durably Reducing Conspiracy Beliefs Through Dialogues with AI,” Science 385(6714), eadq1814. The interview protocol, prompts, and choice-generation step are our own.