The increasing realism and availability of AI-generated content, particularly through tools like ChatGPT and Stable Diffusion, pose growing challenges in academia and healthcare. Students may submit theses partially or entirely written by AI, while synthetic medical reports threaten the reliability of documentation in clinical environments. This PhD project addresses the urgent need for robust and interpretable methods to distinguish AI-generated from human-authored content.

The core objective is to develop a multimodal detection system that integrates Natural Language Processing (NLP), Computer Vision, and Explainable AI (XAI). On the textual side, we analyze linguistic and stylometric features such as perplexity, lexical diversity, part-of-speech patterns, and readability indices using transformer-based classifiers. On the visual side, we detect AI-generated imagery using frequency-domain analysis, CNN-based classifiers, and techniques like PRNU and Error Level Analysis. The system introduces an AI Probability Score to move beyond binary classification, providing nuanced assessments of authenticity.

A custom dataset of pre-2022 academic reports and anonymized medical records, paired with AI-generated counterparts, will be constructed for robust training and evaluation. Explainability is a central theme, with techniques such as SHAP, Grad-CAM, and LIME ensuring that users can understand and trust the model’s decisions.

This research contributes to content integrity in sensitive domains, offering a scalable, transparent tool for universities, hospitals, and publishers to verify document authenticity and address emerging threats from synthetic media.