Phishing has emerged as one of the fastest-growing cyber threats, with recent reports highlighting a sharp surge in attacks and billions of dollars in annual financial losses. Traditional defenses are inadequate, driving research into machine learning and deep learning solutions. This survey reviews 16 state-of-the-art approaches, covering machine learning, deep learning, and hybrid methods such as Random Forest, XGBoost, CNNs with attention, recurrent models, and advanced hybrids using BERT, Graph Neural Networks, and Large Language Models. While many achieve over 95% accuracy on benchmarks, challenges remain in adversarial robustness, dataset imbalance, computational efficiency, explainability, and cross-domain generalization. We highlight key research gaps in adversarial defense, privacy-preserving detection, zero-day identification, and deployment optimization, and outline future directions including federated learning, cross-platform detection, explainable AI, continuous adaptation, and lightweight edge-ready architectures. This survey synthesizes current methods, datasets, and metrics while offering guidance for building robust and trustworthy phishing detection systems against evolving threats.