Generative-AI-Aware Phishing Detection with Human-in-the-Loop Security Analytics: An Illustrative Cybersecurity Evaluation

This article presents a simulated empirical assessment in Computer Science, Cybersecurity, and Human Factors. The evaluation used 8,400 simulated training messages, 1,800 validation messages, and 1,800 hold-out messages, including 600 generative-AI-assisted phishing examples in the hold-out set. A simulated corpus of 12,000 email messages was created with equal proportions of benign business communication, conventional phishing, and generative-AI-assisted phishing. Messages varied in urgency, personalization, link obfuscation, credential requests, tone, and contextual plausibility. Four detection workflows were evaluated: rules plus lexical features, transformer classifier, LLM- assisted security triage, and LLM-assisted triage with analyst confirmation for medium-confidence cases. No operational attack infrastructure or real victim data were used. The human-in-the-loop workflow achieved the strongest simulated balance between detection and false-positive control, with an F1 score of 0.94 and a 2.8% false- positive rate. Pure LLM triage detected subtle AI-generated persuasion better than lexical baselines but produced more false alarms on legitimate high-urgency corporate messages. Analyst review of medium-confidence cases reduced these errors while keeping simulated review workload below 18% of all messages. The numerical values are synthetic and are used solely to demonstrate a plausible empirical design, comparative analysis, and journal-style reporting workflow. No real participants, operational institutions, commercial facilities, or field sites were studied. The analysis emphasizes current research trends, multi-metric evaluation, practical trade-offs, and the importance of validation before deployment.