Phishing remains the most prevalent cybersecurity threat facing organisations today, driven increasingly by AI-generated content that bypasses traditional detection systems. Existing research on NLP-based phishing email detection suffers from persistent methodological gaps: methods are rarely compared under consistent, controlled conditions, and critical deployment metrics such as false-positive rate and inference time are seldom reported alongside accuracy.

This project conducts a comparative evaluation of four NLP approach families for phishing email detection, classical machine learning with TF-IDF and linear classifiers, lightweight deep learning with TextCNN and BiLSTM, transformer-based fine-tuning with DistilBERT, and semantic hashing with SimHash, evaluated independently on two datasets: MeAJOR corpus and a Kaggle dataset. All models are trained and tested using a consistent 60/40 stratified split, with performance measured across accuracy, F1-score, false-positive rate, and per-email inference time.

Results show that DistilBERT achieves the highest accuracy and lowest false positive rate across both datasets, recording 98.96% accuracy and an FPR of 0.0080 on MeAJOR and 99.48% accuracy with an FPR of 0.0043 on Kaggle, while Linear SVM emerges as the strongest classical model with sub-millisecond inference and above 98% accuracy on both datasets. SimHash consistently underperforms, confirming the fundamental limitations of fingerprint-based detection against linguistically diverse phishing content. The Kaggle LLM-generated email subset reveals a clear vulnerability in surface-pattern models: Multinomial Naïve Bayes collapsed from a human-written F1 of 0.7976 to an LLM F1 of 0.0597 on the Kaggle per-source breakdown, while most other models maintained or improved performance on LLM-generated content, suggesting that the vocabulary-dependence of bag-of-words models is uniquely vulnerable to AI-generated phishing. DistilBERT's contextual representations proved more robust to this shift, suggesting that semantic modelling is becoming increasingly important as AI-generated phishing scales. The project contributes a fully reproducible evaluation framework, novel FPR and inference time measurements, and evidence-based deployment recommendations under varying operational constraints.