Artificial Intelligence Fundamentals

Artificial intelligence is not one algorithm, one product, or one kind of output. It is a broad family of machine-based systems. These systems take input and infer how to produce predictions, content, recommendations, or decisions for an objective.

That definition gives you a useful boundary. An AI system can recommend a song, classify an image, forecast demand, generate text, or choose an action. The output may affect software, a business process, or a physical device. The amount of autonomy varies by system.

The practical question is not, "Does this look intelligent?" Ask what objective the system serves, what input it receives, what output it produces, and what that output influences.

The system is larger than the model

An AI model maps input to output. An AI system surrounds one or more models with data collection, application logic, interfaces, monitoring, people, and operating controls.

This distinction matters. A model can score well in isolation while the deployed system fails its users. Input may differ from training data. An interface may hide uncertainty. A process may give an output more authority than it deserves. You evaluate the complete system in its intended context, not only the model.

Use this five-part mental model:

Objective — the result the system is designed to pursue.
Input — data, rules, prompts, sensor readings, or other signals.
Model — the learned or hand-built representation used for inference.
Output — a prediction, content, recommendation, or decision.
Context — the people, process, environment, and consequences around the output.

AI is broader than machine learning

Machine learning is a set of techniques that builds models from training data. The model learns patterns or relationships instead of receiving an explicit instruction for every case.

Machine learning is part of AI, but it is not the whole field. Knowledge-based AI uses explicit logical or probabilistic representations to infer outputs. Some systems combine learned models with rules, search, planning, or human decisions.

This gives you two broad ways to encode behavior:

Explicit knowledge — people describe rules, variables, and relationships.
Learned patterns — training adjusts a model from examples, feedback, or interaction.

Neither approach is automatically better. The right choice depends on the task, available data, required transparency, operating conditions, and cost of error.

How machine learning learns

Machine learning methods differ mainly in the signal used during training.

Supervised learning trains on labeled examples. Each example includes input features and a known target called a label. Classification predicts a category. Regression predicts a number.

Unsupervised learning looks for structure in data without known answers. Clustering groups similar examples, but the model does not know the human meaning of each group.

Reinforcement learning learns a policy from rewards or penalties received after actions in an environment. The policy defines a strategy for choosing actions.

Generative AI creates content such as text, images, audio, video, or code. Generative models learn patterns in training data and produce new output from user input. Generative AI is one part of the AI landscape, not a synonym for all AI.

Training and inference

Training is the build-time process that adjusts a model from data or feedback. Inference is the use-time process that applies a model to new input and produces output.

For supervised learning, a common path is:

Define the task and the label you need.
Collect and prepare representative examples.
Split data so evaluation uses examples not used for training.
Train the model to connect features with labels.
Evaluate predictions against known answers.
Deploy the complete system.
Monitor performance and context after deployment.

The data-model boundary is central. A large dataset can still omit important conditions. A model tested on familiar examples can perform poorly on new conditions. Evaluation must represent the intended use.

Models you will encounter

A model is a structure plus parameters that maps input to output. Different structures suit different tasks.

Linear models learn weighted relationships.
Decision trees route examples through learned conditions.
Neural networks use connected layers of learned units. A deep neural network has more than one hidden layer.
Large language models estimate likely token sequences from context and use that capability for text tasks.

These names describe model families, not complete systems. The same model family can support many applications, and one application can combine several models.

Where AI fits

AI is useful when a system must infer an output from complex input and the required behavior cannot be captured well by a small, stable rule set. Representative uses include recognition, anomaly detection, forecasting, personalization, interaction support, optimization, and reasoning over knowledge structures.

Do not start with a model. Start with the decision or outcome. Establish what success means, what errors matter, who is affected, and whether a deterministic method already solves the problem.

AI is a poor fit when the correct behavior can be expressed as clear rules, when representative data is unavailable, or when errors cannot be controlled to an acceptable level. A model prediction is not a guarantee. It is an inference that must be evaluated in context.

Trustworthiness is part of the design

NIST groups trustworthy AI characteristics into seven connected areas: valid and reliable; safe; secure and resilient; accountable and transparent; explainable and interpretable; privacy-enhanced; and fair with harmful bias managed.

These characteristics can conflict. More transparency may expose private or proprietary information. A more interpretable model may trade accuracy against another model in a specific task. You make those tradeoffs explicitly for the intended context.

Risk work continues across the lifecycle. Define responsibility, map the context and impacts, measure performance and trustworthiness, and manage prioritized risks. Monitoring matters because data, users, environments, and system behavior can change after deployment.

Where this course stops

This course gives you the map: systems and models, AI approaches, learning types, training and inference, common uses, and trustworthiness. It does not teach model implementation, production operations, prompt engineering, or a specific framework.

Your next path is ordered. Learn supervised learning and evaluation first. Then study data quality, generalization, and overfitting. Add neural networks and generative models. Finish with production monitoring, security, fairness, and risk management.