Dialogues with AI and Misinformation Discernment

This repository contains code, processed data artifacts, analysis scripts, and study interface materials for the CHI 2026 paper:

Dialogues with AI Reduce Beliefs in Misinformation but Build No Lasting Discernment Skills
Anku Rani, Valdemar Danry, Paul Pu Liang, Andrew Lippman, Pattie Maes
DOI: https://doi.org/10.1145/3772318.3790656

Project Summary

This project investigates whether conversational AI helps people learn to detect misinformation, or mainly improves in-the-moment performance while assistance is available.

Across a month-long longitudinal study, participants completed repeated sessions with three tasks:

• Before AI: independent authenticity judgments
• With AI: judgments supported by dialogue with an AI assistant
• After AI: independent judgments on new items right after AI assistance (when the AI was removed).

The paper reports a central pattern: strong immediate gains during AI-assisted interaction, but no durable improvement in independent discernment and a decline in post-assistance unassisted performance over time.

Repository Contents

.
├── analysis/
│   ├── preprocess.ipynb
│   ├── conversation_analysis.ipynb
│   ├── plot_llm_as_judge.ipynb
│   ├── classifiers_llm/
│   │   ├── llm_as_a_judge.py
│   │   └── classifier_prompts.py
│   ├── nlp/
│   └── stats/
│       ├── analysis.R
│       ├── analysis_persuasive.R
│       └── final_persuasive_outputs_R/
├── data/
│   ├── process_data.py
│   ├── raw/
│   └── processed/
├── study_interface/
│   ├── README.md
│   ├── Dialogues with AI app/user-interface/
│   └── Phase wise Data/
├── requirements.txt
└── LICENSE

Folder Guide

• data/raw/: raw study exports and source tabular files
• data/processed/: processed analysis-ready datasets and derived outputs
• analysis/stats/: R scripts for primary inferential analyses and robustness outputs
• analysis/classifiers_llm/: LLM-as-a-judge pipeline for conversation strategy labeling
• analysis/*.ipynb: notebook-based preprocessing, exploration, and plotting
• study_interface/: participant-facing web app and phase-wise media/data assets

Reproducibility Setup

1) Python Environment

From the repository root:

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

requirements.txt includes core dependencies used in this repository (OpenAI SDK, dotenv, pandas/numpy, Jupyter, matplotlib/seaborn).

2) R Environment

The R scripts install missing CRAN packages automatically on first run.

Reproducing Core Analyses

A) Persuasive-condition primary analyses (main paper outputs)

Rscript analysis/stats/analysis_persuasive.R

Primary output directory:

• analysis/stats/final_persuasive_outputs_R/

This script produces:

• participant fixed-effects linear probability model results with HC2 robust errors
• planned within-week contrasts (with-before, after-before)
• trend and week-difference tests for before-only and after-only outcomes
• truth-split robustness summaries
• Wilson interval summaries and a figure-ready PDF

B) Two-condition comparison pipeline

Rscript analysis/stats/analysis.R

Output directory:

• analysis/stats/output/

This script estimates condition-by-week-by-phase effects and exports adjusted means, contrasts, difference-in-differences, and growth tests.

C) LLM-as-a-judge conversation classification

1. Create a root .env file with your OpenAI key:

OPENAI_API_KEY=your_key_here

2. Run the classifier:

python analysis/classifiers_llm/llm_as_a_judge.py \
  --data data/processed/conversations \
  --classifiers all \
  --output data/processed/llm_as_judge_results

The classifier reconstructs conversations from phase CSV files and labels strategy categories defined in analysis/classifiers_llm/classifier_prompts.py.

D) Data processing utility

python data/process_data.py

This utility merges and harmonizes phase-level files and writes a merged CSV under data/processed/.

E) Notebooks

Use Jupyter to run exploratory and plotting notebooks:

jupyter notebook

Key notebooks are in:

• analysis/preprocess.ipynb
• analysis/conversation_analysis.ipynb
• analysis/plot_llm_as_judge.ipynb

Study Interface (Web App)

The participant interface is under:

• study_interface/Dialogues with AI app/user-interface/

Run locally:

cd "study_interface/Dialogues with AI app/user-interface"
npm install
npm start

Build/test:

npm run build
npm test

Additional details are documented in study_interface/README.md.

Data and Ethics Notes

• This repository is intended for research transparency and replication.
• Use the data in accordance with institutional, legal, and ethical requirements.

Citation

If you use this repository, please cite the paper:

@inproceedings{rani2026dialogues,
  title     = {Dialogues with AI Reduce Beliefs in Misinformation but Build No Lasting Discernment Skills},
  author    = {Rani, Anku and Danry, Valdemar and Liang, Paul Pu and Lippman, Andrew and Maes, Pattie},
  booktitle = {Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26)},
  year      = {2026},
  articleno = {792},
  pages     = {1--26},
  doi       = {10.1145/3772318.3790656},
  url       = {https://doi.org/10.1145/3772318.3790656}
}

License

See LICENSE for repository licensing details.