What Is Artificial Intelligence?

8. Expanded lesson: concepts, boundaries and examples

8.1. What makes a system an AI system?

There is no single definition accepted by every research community. In this course, use an engineering definition: an AI system receives information about a task and its environment, processes that information using rules, search, learned parameters or a combination of these, and produces an output intended to support or perform an intelligent task. The definition describes observable function; it does not imply consciousness, human-like understanding or independent goals.

When you assess a system, state its boundary. A model alone is not the whole product. The deployed system may include sensors, data validation, preprocessing, a model, business rules, a user interface, a human review step, monitoring and feedback. A failure can originate in any of these components. For example, a correct model can still produce a harmful recommendation if the input units are wrong or a safety rule is missing.

8.2. How the main approaches differ

Rule-based systems execute conditions explicitly written by people: if a measured value exceeds a threshold, raise an alert. They are often transparent and useful when rules are stable, but become difficult to maintain when exceptions multiply. Expert systems represent domain knowledge through facts and rules, sometimes using an inference engine to derive conclusions. They can make reasoning traces inspectable, yet their knowledge base requires careful maintenance.

Search and planning explore possible states and actions to find a path that satisfies a goal or minimizes a cost. A route planner, scheduler or game-playing program may use search without learning its policy from examples. Machine learning estimates parameters or decision rules from data. Deep learning is a family of machine-learning methods based on multi-layer neural networks that learn internal representations. These approaches can be combined: a product may use a neural model, explicit constraints and a search procedure.

8.3. Different tasks require different outputs

• Recognition: detect or identify an object, sound, event or pattern.
• Classification: assign an input to one of a set of categories.
• Prediction and forecasting: estimate a future or unknown value.
• Anomaly detection: flag observations that differ from expected behaviour.
• Recommendation: rank or suggest options for a user or process.
• Generation: produce new text, images, audio, code or other structured output.

Do not confuse the task with the method. Classification can be implemented with rules, a decision tree, logistic regression or a neural network. The task describes what output is needed; the method describes how the system obtains it.

8.4. Traditional programming and machine learning

In conventional programming, a developer writes the rules that transform inputs into outputs. In supervised machine learning, the developer supplies examples with inputs and expected answers; a training algorithm adjusts model parameters to reduce a defined loss. This does not mean that the machine discovers truth automatically. Results depend on how examples were collected, which target was chosen, how the loss represents the real objective and whether evaluation data represent future use.

8.5. Narrow AI and general intelligence

Most deployed AI is narrow: it performs within a bounded task, domain or operating envelope. Strong performance in one domain does not establish broad transfer to unrelated tasks. Artificial General Intelligence (AGI) is a debated research concept with no universally accepted operational test. A fluent response or a successful benchmark alone is insufficient evidence of general intelligence.

8.6. Worked example: a building energy assistant

Imagine a system that helps an engineer review building energy use. It receives meter readings, weather, schedules and equipment states. Validation checks missing values and units. A forecasting model estimates demand. A rule engine checks operational limits. The interface presents a recommendation and its evidence; an engineer decides whether to act. If the forecast is inaccurate, investigate sensor quality, data alignment, model drift and operating assumptions before blaming the model alone. This example shows why system-level evaluation and explainability matter.

8.7. Why this foundation matters for Explainable AI

Explainable AI (XAI) studies methods for understanding and communicating how an AI-based system produces outputs and where its limitations lie. Before asking why a model made a prediction, identify the model, its inputs, the target, the surrounding rules and the decision being supported. An explanation of one model output is not automatically an explanation of the whole product, nor proof that the decision is correct or fair.

8.8. Practical task

Choose an AI-enabled product. Describe its input, output, user, operating environment and consequence of error. Identify which parts may use explicit rules, search, machine learning or human judgement. Then list two failure modes outside the model itself and explain how you would detect them. State what evidence would be needed before calling the product reliable.

9. Deep dive: approaches and task types

9.1. Rule-based systems

A rule-based system represents decisions as explicit conditions and actions. For example: if a measured pressure exceeds a specified limit, issue an alarm; if a required sensor is unavailable, stop automatic control and request inspection. The advantage is that individual rules can be reviewed and traced. The limitations are equally important: rules cover only situations anticipated by their authors, conflicting rules need resolution, and a growing exception list becomes difficult to maintain. Rules are often appropriate for clear constraints and safety interlocks, but they should not be mistaken for a system that has learned from experience.

9.2. Expert systems

An expert system encodes specialist knowledge in a knowledge base and applies an inference mechanism to facts. A diagnostic system may combine symptoms, equipment states and known failure patterns to produce possible causes. Its explanation may show which facts and rules contributed to a conclusion. However, the quality of the result depends on the completeness and currency of the knowledge base. An expert system does not automatically become more knowledgeable when new events occur; its knowledge must be reviewed and updated.

9.3. Search and planning

Search methods examine possible states or action sequences. A planner may represent a problem as a set of states, legal actions and a goal. It then explores alternatives, sometimes using a cost function or heuristic to prioritize promising paths. Examples include route planning, timetabling, resource allocation and game playing. Search can produce a valid plan without a neural network or a statistical training dataset. The challenge is controlling the size of the search space and ensuring that the representation captures the real constraints.

9.4. Machine learning and deep learning

Machine learning is useful when examples contain patterns that are difficult to encode as a manageable list of rules. A model can estimate energy demand from weather and occupancy, classify a transaction as suspicious or rank documents by relevance. Deep learning uses layered neural networks and is often effective for images, audio, text and other high-dimensional data. It may require substantial representative data, compute and monitoring. In many practical systems, learned predictions are constrained by explicit rules and reviewed by people.

9.5. Examples of AI tasks

Recognition answers “what is present?”—for example, detecting a person or a crack in an image. Classification assigns a category, such as normal/fault or one of several document types. Prediction estimates an unknown value, while forecasting estimates a future value using information available now. Anomaly detection highlights unusual readings that deserve investigation; unusual does not automatically mean faulty. Recommendation ranks candidate actions or content. Generation creates a new response, drawing, audio segment or code fragment. One product may combine several tasks in a single workflow.

9.6. Where AI is applied—and what must be checked

In engineering, AI may support predictive maintenance, energy forecasting, image inspection and document review. In healthcare, it may assist image triage or risk estimation, but clinical decisions require appropriate validation and professional oversight. In finance, it may detect suspicious transactions or estimate credit risk, raising questions about bias, accountability and appeal. In education, it may recommend learning materials or provide feedback, but the quality of the learning objective and the student's privacy remain important. The domain changes the acceptable error rate, the evidence required and the level of human control.

9.7. Why AI is not the same as human thought

A system can produce a useful result without having human experience, intentions or consciousness. A classifier maps inputs to labels according to learned or programmed structure; a language model generates likely continuations from its training and context. Fluent output should not be treated as proof that the system understands a situation in the same way a person does. Ask what the system was trained to do, what evidence it uses, where it fails and whether its output can be checked.

9.8. Mini-experiment: three ways to detect excessive energy use

1. Rule-based: alert when daily consumption exceeds a fixed threshold. It is easy to explain, but a single threshold may ignore weather and occupancy.
2. Machine learning: predict expected consumption from temperature, area and occupancy, then investigate a large difference between actual and predicted use. It can adapt to patterns in examples, but depends on representative data and careful evaluation.
3. Hybrid: use a model to estimate expected consumption, then apply explicit safety limits and require human review before an operational action. This can combine flexibility with transparent constraints.

For each approach, identify one false alarm and one missed event. Decide what data would be needed to compare them fairly. The purpose is not to select a winner from a toy example, but to see how task definition and consequences shape the system design.