Towards Responsible Generation of Synthetic Datasets with Large Language Models

Synthetic datasets generated through large language models are emerging as a promising resource to support artificial intelligence applications, particularly for machine learning, when real data are scarce (e.g., genuine emotion expressions instead of acted ones), sensitive (e.g., personal data subject to privacy concerns), difficult to obtain (e.g., online grooming conversations, which law enforcement cannot release), or ethically problematic (e.g., pain data, that cannot be elicited in laboratory settings). While useful, their generation poses significant methodological and ethical challenges, e.g., ensuring similarity to real data, avoiding bias amplification, and guaranteeing transparency through proper documentation. This paper proposes a structured methodology for generating synthetic datasets, emphasizing ontology-driven design, multi-model prompt testing, iterative human-in-the-loop supervision, semantic and feature-level validation, and modality-aware documentation. A case study based on the Plutchik wheel of emotions illustrates the feasibility of our approach.