Discriminative Artificial Intelligence is a class of AI methods that learn a mapping from inputs to labels (or targets), or directly model the conditional distribution P(y | x), commonly used for classification, ranking, and anomaly detection. Generative methods, on the other hand, primarily model data distributions, joint distributions P(x, y), or conditional distributions P(x | y), and can generate samples based on these models. The main difference between the two lies in their modeling objectives, and cannot be simply summarized as "whether or not they can generate content."

Early perceptrons and support vector machines belong to this lineage. In 2001, Ng and Jordan systematized the distinction between the two in their NIPS paper 《On Discriminative vs. Generative Classifiers: A Comparison of Logistic Regression and Naive Bayes》: generative models model the joint distribution P(x, y), while discriminative models directly model the conditional distribution P(y | x). Under the logistic regression and Naive Bayes models and experimental conditions compared in that study, generative methods converge faster in the small-sample regime, while discriminative methods may achieve lower asymptotic error as data increases. These are theoretical and experimental conclusions under specific models and assumptions, and should not be simplified into a "generative for small data, discriminative for big data" law applicable to all models.

Historically, spam filtering, credit scoring, and image classification, as well as today's bidirectional encoders like BERT and DeBERTa, are representative examples of the discriminative approach. In 2026, Kasliwal and Singh presented a modern trade-off for the same task in an arXiv preprint 《Multi-Label Topic Assignment via LLM Distillation: A Comparative Analysis of Generative vs. Discriminative Student Models》: for e-commerce multi-label topic labeling tasks, discriminative models are more accurate on structured product reviews, whereas even the smallest 1B generative model can outperform them in complex multi-turn dialogues; when the label scale expands to 112 topics, the Macro-F1 of the discriminative baseline drops by approximately 35%. This also explains the common pattern of "large model annotation + small model discrimination." Scenarios that involve "making judgments," such as risk control, quality inspection, medical diagnostic assistance, and retrieval ranking, remain the primary domain of discriminative models.