← Lab

Lab / Active Research

Hallucination Detection

When and why does AI-assisted writing fabricate citations in submitted and accepted academic papers?

Active Research

Research Questions

RQ1 — Expertise

Are hallucinated citations more likely when an author cites outside their primary domain of expertise?

RQ2 — Domain Velocity

Do fields with faster publication cycles produce higher hallucination rates than slower-moving fields?

RQ3 — Location

Do hallucinated citations cluster in particular sections of a paper (Related Work vs. Methods vs. Discussion)?

Pipeline

The project uses a four-phase automated and semi-automated pipeline targeting a stratified random sample of ~300–400 papers across fields with differing publication velocities.

0
Collection

Stratified sampling and PDF acquisition across target venues. Semi-automated with seeded random sampling for reproducibility.

1
Extraction

Citation extraction from PDFs; CrossRef verification; GPTZero AI-content scoring. Fully automated.

2
Coding

Manual coding of author expertise and citation characteristics via an interactive interface. ~20 min/paper.

3
Analysis

Hypothesis testing and visualization. Logistic regression on citation-level data; mixed models accounting for paper-level clustering.

Stack

Python CrossRef API GPTZero API Quarto Statistical Analysis PDF Extraction

Pre-print forthcoming. Citation details will be added upon publication.