The Role of Artificial Intelligence in Detecting Social Engineering and Phishing Attacks: Opportunities, Limitations, and Ethical Considerations

Abstract

Phishing and social engineering remain leading causes of cybersecurity breaches, with the FBI reporting 859,532 complaints and $16.0 billion in losses in 2024. This literature review examines how artificial intelligence (AI) – including machine learning (ML), natural language processing (NLP), deep learning (DL), and large language models (LLMs) – can enhance detection of phishing/social engineering. We compare AI-based techniques to traditional rule-based and heuristic methods, analyzing detection performance, adaptability, false positives/negatives, computational demands, and resilience to adversarial manipulation. Key findings show that AI classifiers (e.g. CNNs, RNNs, BERT) often achieve very high accuracy (e.g. 98–99% on standard email corpora) and can adapt to evolving attacks, but they are vulnerable to adversarial rephrasing and suffer from explainability and data bias issues. Ethical concerns include user privacy (inspecting personal communications), transparency (black-box models), and model bias (especially in cross-cultural contexts). We illustrate these concepts with ScamLens AI app combining NLP analysis of text and URLs with vision-based screenshot scanning, producing risk indicators and explanations. Finally, we identify research gaps – such as limited real-world evaluation, outdated datasets, and the need for phishing-specific LLMs – and outline future directions (adversarial robustness, federated training, multimodal analysis, and human-AI collaboration).