Dataset

Dataset for the studyThis repository contains the replication dataset, analysis scripts, and experimental materials for the research paper <b>"Improving Phishing Resilience with AI-Generated Training: Large-Scale Evidence on Prompting, Personalization, and Duration"</b>.Study OverviewThe associated study investigates the effectiveness of Large Language Models (LLMs) in generating tailored phishing awareness training. Through two controlled experiments involving 480 participants, the research examines how Prompting Strategies (Study 1), Personalization, and Training Duration (Study 2) influence user detection accuracy, recall, and F1-scores against realistic phishing emails.Repository ContentsThis archive provides all necessary resources to reproduce the statistical analyses and replicate the experimental design:<b>📂 Datasets:</b> Anonymized and cleaned participant data (CSV format) for Study 1 ($N=80$) and Study 2 ($N=400$), including performance metrics (Pre/Post) and psychometric profiles (BFI-2-XS, StP-II-B, TEIQue-SF).<b>📂 Analysis Code:</b> Python scripts using <i>pandas</i> and <i>pingouin</i> to replicate the Mixed ANOVA, t-tests, and correlation analyses reported in the paper.<b>📂 Experimental Materials:</b> The complete set of email stimuli (genuine vs. phishing) developed using the NIST Phish Scale, including screenshots and HTML templates.<b>📂 LLM Prompts:</b> The exact system and user prompts used to generate the training modules via OpenAI's model, covering four prompting strategies (Direct-profile, Few-shot, Table-based, Guideline-based) and personalization logic.<b>📂 Instruments:</b> Details on the psychometric scales and the ad-hoc training reaction questionnaire used in the study.Methodological NoteThe data collection was conducted on the Prolific platform with approval from the Ethics Committee of the University of Bari. All Personal Identifiable Information (PII) has been removed to ensure participant anonymity.