An AI model is a computational model that has been designed or trained to perform tasks associated with artificial intelligence, typically by representing patterns, relationships, rules, probabilities, or transformations that can be applied to inputs. In modern machine learning, the term usually refers to a trained parameterised system whose parameters have been learned from data using an optimisation procedure.
An AI model can take inputs such as text, images, audio, numerical measurements, sensor data, or structured records and produce outputs such as classifications, predictions, rankings, generated content, embeddings, decisions, or control actions. The word "model" is important because the system is a mathematical or computational representation of relationships relevant to a task rather than a complete intelligent entity. Models range from simple linear and logistic regression models to decision trees, support-vector machines, neural networks, probabilistic models, reinforcement-learning policies, and large multimodal systems.
A model normally has a structure, parameters, inputs, outputs, and an inference procedure. The structure defines the mathematical form of the model, while parameters are numerical values that determine its behaviour. Training or fitting estimates parameter values from data according to an objective. For example, a linear regression model can learn coefficients describing a relationship between input variables and a numerical target. A neural network contains many learned weights and biases. A decision tree contains learned or selected split rules.
A probabilistic model can contain parameters describing distributions or conditional relationships. Once trained, inference applies the model to new input data. Some models can also be adapted after initial training through fine-tuning, parameter-efficient methods, calibration, or other procedures. The distinction between training and inference is fundamental: training changes model parameters, while ordinary inference normally uses the established parameters to produce outputs.
AI models can be trained using supervised, unsupervised, self-supervised, or reinforcement-learning methods. Supervised learning uses examples paired with target outputs, such as labelled images for classification. Unsupervised learning seeks structure without explicit labels, such as clustering or density estimation. Self-supervised learning constructs training signals from the input data itself and is central to many large foundation models. Reinforcement learning trains a policy or value-related model through interactions with an environment and reward signals. Different training objectives create different capabilities.
A model trained to predict the next token is not automatically equivalent to one trained to classify sentiment, estimate a physical quantity, or control a robot. Pretraining and post-training can produce multiple stages of model development, with later stages changing how the original learned representations are used.
AI models can be deterministic or probabilistic at the application level. A deterministic model produces the same output for the same input and state, assuming identical computational conditions, while probabilistic models represent uncertainty or generate outcomes according to distributions. Even deterministic neural networks can appear variable if sampling or other stochastic inference mechanisms are placed around them. Classification models may output class labels or probability-like scores, but these scores should not automatically be interpreted as perfectly calibrated probabilities.
Regression models produce numerical estimates, while ranking models order candidates according to learned relevance. Generative models produce samples or sequences. Embedding models convert inputs into vector representations useful for similarity search, retrieval, clustering, recommendation, and other tasks. A deployed AI application can combine several model types rather than relying on a single model.
Model performance is normally evaluated using data that was not used directly to fit the model. A training set is used to learn parameters, a validation set can guide model selection and tuning, and a test set provides an independent evaluation under a defined methodology. Metrics depend on the task. Classification may use accuracy, precision, recall, F1 score, ROC-AUC, or calibration measures. Regression can use mean absolute error, mean squared error, or related measures.
Generative systems require more complex evaluation involving human judgements, automated metrics, factuality, safety, or task-specific criteria. A model can perform well on a benchmark yet fail in production because of distribution shift, data drift, unusual inputs, measurement differences, or changes in user behaviour. Evaluation must therefore match the actual operational environment and the consequences of mistakes.
Model architecture and model parameters should be distinguished. Architecture describes the computational arrangement, such as a transformer, convolutional network, recurrent network, decision tree, or ensemble. Parameters are the learned numerical values within a particular trained instance of that architecture. Two models can use the same architecture but have different parameters because they were trained on different data or with different objectives.
A model checkpoint is a saved state of model parameters and associated information that allows a trained model to be restored. A model version may change architecture, parameters, training data, preprocessing, or other components. A foundation model is trained broadly enough to support multiple downstream tasks, often through prompting, fine-tuning, adapters, retrieval, or task-specific heads. A model is therefore not simply its architecture and not simply its training dataset; its behaviour emerges from the interaction of architecture, parameters, data, training process, inference method, and surrounding system.
AI models have limitations that depend on the model class and application. They may encode biases or spurious correlations from training data, overfit to examples, fail under distribution changes, or produce poorly calibrated uncertainty. Neural models can be vulnerable to adversarial inputs, prompt injection, data poisoning, or other attacks depending on their architecture and deployment. Large models can require considerable computational and memory resources. Small models may be cheaper and faster but have less capacity or coverage for some tasks.
Interpretability varies: a small decision tree may be directly inspectable, while a large neural network may require specialised mechanistic or feature-level analysis. Model compression, quantisation, pruning, distillation, sparsity, and parameter-efficient adaptation can reduce deployment costs or alter performance characteristics. Responsible deployment therefore requires considering accuracy, robustness, security, privacy, cost, latency, maintainability, and the consequences of incorrect outputs.
An AI model should not be confused with an AI algorithm, AI application, dataset, or complete AI system. An algorithm is a procedure for performing computation or learning, while a model is the resulting computational representation used to make predictions or generate outputs. A dataset is the collection of examples or observations used for training, evaluation, or other purposes. An application combines models with software, interfaces, databases, business logic, authentication, monitoring, and other components.
A model can be embedded in a larger AI system and may rely on preprocessing, retrieval, external tools, or post-processing. The term "AI model" is also broader than "machine-learning model" in some contexts because symbolic or rule-based AI systems can be considered models, although contemporary usage often refers to learned models. Understanding these distinctions is essential when discussing what a system can actually do, where its knowledge comes from, how it was trained, and which component is responsible for a particular behaviour.
