Foundations, History and Scope of Artificial Intelligence

Introduction

Artificial intelligence (AI) is one of the most influential areas of contemporary computing. It affects search engines, recommendation systems, medical diagnosis, autonomous vehicles, financial services, education, manufacturing, public administration and scientific research. Despite its visibility, AI is not a single technology or a single algorithm. It is a broad field concerned with designing computational systems that can perceive their environment, represent information, reason, learn, communicate and act toward goals.

A sound understanding of AI begins with its foundations. Students must distinguish AI from ordinary automation, understand how the field developed, recognize its principal paradigms and evaluate both its capabilities and limitations. This chapter presents AI as a scientific and engineering discipline, explains its historical development, and introduces important distinctions such as narrow versus general AI, symbolic versus statistical AI, and predictive systems versus autonomous agents.

Learning Objectives

By the end of this chapter, students should be able to:

1. Define artificial intelligence and explain its central terminology.

2. Describe major historical milestones and the causes of AI winters and later resurgence.

3. Identify the disciplines that contribute to AI.

4. Distinguish symbolic, statistical and hybrid AI paradigms.

5. Differentiate narrow AI from artificial general intelligence.

6. Evaluate AI capabilities, limitations and risks.

7. Explain how modern AI systems are developed and governed.

8. Analyze an applied AI scenario from technical, organizational and ethical perspectives.

What Is Artificial Intelligence?

There is no universally accepted definition of AI. Definitions differ because they emphasize different aspects of intelligence. A useful working definition is:

> Artificial intelligence is the study and engineering of computational systems that perceive or receive information, represent knowledge, learn or reason, and produce actions or decisions directed toward specified objectives.

This definition includes both traditional rule-based programs and modern machine-learning systems. It also emphasizes that AI involves more than producing outputs that appear intelligent. An AI system must be designed to perform a task involving some aspect of perception, reasoning, learning, decision-making or language use.

Four influential perspectives are commonly used to classify definitions of AI:

• Acting humanly: constructing systems whose behavior resembles human behavior, as in the Turing test.
• Thinking humanly: modeling human cognitive processes, including memory, problem solving and language understanding.
• Thinking rationally: using formal logic or valid inference to derive appropriate conclusions.
• Acting rationally: selecting actions expected to achieve goals under available information and constraints.

The rational-agent perspective is particularly useful in engineering. An agent receives observations from an environment and chooses actions. Its performance depends not only on the action itself but also on the objective, uncertainty, available resources and consequences of decisions. A rational agent is not necessarily all-knowing or always successful; it chooses an appropriate action based on the information and computational resources available to it.

Core Terminology

Several terms are fundamental to AI:

Agent: An entity that perceives an environment through sensors or data inputs and affects that environment through actions or outputs.

Environment: The external world, software context or organizational setting in which an agent operates.

State: A representation of relevant information about the current situation of an environment.

Action: An operation selected by an agent, such as recommending a product, moving a robot or rejecting a transaction.

Goal: A desired condition or outcome.

Utility: A numerical or qualitative representation of how desirable an outcome is.

Knowledge representation: The methods used to encode facts, concepts, relationships, rules or uncertainty so that a system can use them.

Inference: The process of deriving conclusions from available information.

Learning: The improvement of a system’s performance through experience, data or feedback.

Model: A computational representation that captures patterns, relationships or decision procedures relevant to a task.

Training: The process of fitting or configuring a model using data, examples or rules.

Inference: The use of a trained or configured model to produce an output for a new input. Inference is distinct from training.

Feature: An input variable or measurable property used by a model.

Label: The target value or category associated with an example in supervised learning.

Dataset: A collection of examples used for training, validation or testing.

Autonomy: The ability of a system to operate with limited direct human control.

Generalization: The ability to perform effectively on new cases rather than only on examples seen during training.

Explainability: The extent to which the reasons for a system’s output can be understood by people.

Alignment: The degree to which an AI system’s objectives and behavior correspond to legitimate human intentions and values.

These terms are related but not interchangeable. For example, a system may be highly accurate without being autonomous, and it may be autonomous without being general-purpose.

The Intellectual Foundations of AI

AI emerged from several disciplines rather than from computer science alone.

Mathematics and Logic

Mathematics provides formal languages for representing uncertainty, optimization and relationships. Logic supports symbolic reasoning, while probability supports reasoning under incomplete or uncertain information. Optimization provides methods for selecting model parameters or actions that improve an objective.

Computer Science

Computer science contributes algorithms, data structures, programming languages, databases, computational complexity, computer architecture and software engineering. AI is constrained by the time, memory, energy and hardware required to compute an answer.

Statistics and Data Science

Statistics supplies methods for estimating patterns from data, quantifying uncertainty and evaluating hypotheses. Modern machine learning relies heavily on statistical concepts such as sampling, estimation, generalization and error measurement.

Neuroscience and Cognitive Science

Research on the brain and human cognition has inspired neural networks, attention mechanisms, memory models and theories of perception and learning. However, an artificial model inspired by biology is not necessarily a faithful model of the human brain.

Linguistics

Natural language processing draws on syntax, semantics, pragmatics and discourse analysis. Language technologies must address ambiguity, context, variation and the relationship between words and the world.

Psychology and Philosophy

Psychology contributes theories of learning, judgment and perception. Philosophy raises questions about knowledge, consciousness, reasoning, responsibility and the meaning of intelligence. These questions remain important because technical performance alone does not determine whether an AI system is socially acceptable.

Economics, Control and Operations Research

Economics studies preferences, incentives and strategic behavior. Control theory concerns the behavior of dynamic systems and feedback. Operations research contributes optimization, scheduling and decision analysis. Together these disciplines support planning and action in complex environments.

Historical Development

Early Foundations: Logic, Computation and Cybernetics

The intellectual roots of AI predate electronic computers. Formal logic developed methods for manipulating symbols according to rules. In the nineteenth century, Charles Babbage and Ada Lovelace articulated ideas related to programmable computation. In the twentieth century, mathematical logic and computability theory clarified what could be computed.

Alan Turing’s 1950 paper, “Computing Machinery and Intelligence,” asked whether machines could think and proposed an operational conversation-based test rather than relying on an abstract definition of thought. Norbert Wiener’s work on cybernetics emphasized feedback and control in biological and mechanical systems. These ideas helped establish the possibility of intelligent behavior through computation.

The Birth of AI as a Field

The term “artificial intelligence” is generally associated with the Dartmouth workshop proposal of 1955, prepared by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon for a 1956 research project. The proposal expressed the optimistic view that learning and intelligence might be described precisely enough for machines to simulate them.

Early research explored theorem proving, game playing, symbolic problem solving and natural language. Programs such as the Logic Theorist and the General Problem Solver demonstrated that computers could manipulate symbols and solve restricted classes of problems. Research laboratories also developed early robots and systems for planning and perception.

Symbolic AI and Expert Systems

From the 1950s through the 1970s, much AI research followed the symbolic paradigm. Intelligence was represented through explicit symbols, rules, logical statements and search procedures. A system might represent the rule “if a patient has symptom X and test result Y, consider diagnosis Z,” then apply this rule to a case.

Expert systems became prominent during the 1970s and 1980s. They encoded specialized knowledge in a knowledge base and used an inference engine to derive conclusions. Their success in narrow domains demonstrated the commercial value of AI, but their performance depended heavily on carefully acquired knowledge and controlled environments.

AI Winters

An AI winter is a period of reduced funding, public interest and institutional confidence in AI. AI winters occurred partly because early claims exceeded actual capabilities. Systems that worked well in demonstrations often failed when confronted with ambiguous language, large search spaces, incomplete knowledge or changing environments.

The first major decline followed concerns during the 1970s that machine translation and other projects had not met ambitious expectations. Another downturn occurred in the late 1980s and early 1990s. Expert systems were expensive to maintain, difficult to update and vulnerable to errors outside their designed domain. Specialized hardware markets also weakened.

AI winters were not caused by a single technical failure. They resulted from a combination of unrealistic promises, limited computing power, insufficient data, difficult problems, poor deployment economics and changes in research priorities. The history illustrates an important professional lesson: demonstrations and predictions should be evaluated separately from robust, real-world performance.

Statistical and Machine-Learning Resurgence

From the 1990s onward, statistical machine learning became increasingly influential. Instead of specifying every rule manually, researchers developed algorithms that inferred patterns from examples. Decision trees, support vector machines, probabilistic models and ensemble methods achieved strong results in classification, prediction and speech processing.

Several developments supported renewed progress:

• faster and cheaper computation;
• large digital datasets;
• improved algorithms and optimization methods;
• specialized hardware, including graphics processing units;
• internet-scale data collection and distribution;
• improved software libraries and engineering practices.

In 1997, IBM’s Deep Blue defeated world chess champion Garry Kasparov under match conditions, demonstrating the power of specialized search and evaluation. In 2012, a deep neural network achieved a major improvement in large-scale image classification, helping initiate the modern deep-learning expansion. In 2016, AlphaGo defeated professional Go player Lee Sedol through a combination of neural networks, search and reinforcement learning.

Generative and Foundation Models

Recent AI systems can generate text, images, audio, video and software code. Large models are trained on broad datasets and adapted to many tasks. Such systems are often called foundation models because they can serve as a base for multiple applications.

Their capabilities include summarization, translation, question answering, code assistance and multimodal interaction. However, fluent output does not guarantee factual accuracy, sound reasoning or understanding. Contemporary AI therefore combines impressive practical performance with continuing research problems involving reliability, bias, interpretability, data governance and control.

Major AI Paradigms

Symbolic AI

Symbolic AI represents knowledge explicitly through rules, logic, semantic networks, ontologies or planning structures. Its strengths include transparency, precise manipulation of abstract concepts and reliable behavior when rules are complete and applicable. It is useful in configuration, formal verification, scheduling, compliance reasoning and domains with stable expert knowledge.

Its limitations include the difficulty of manually encoding common sense, brittleness when inputs differ from expected forms and the challenge of handling noisy perceptual data. A rule-based system may know that a bird typically flies yet require many exceptions for penguins, injured birds and artificial birds.

Statistical AI and Machine Learning

Statistical AI learns relationships from data. In supervised learning, examples contain inputs and target outputs. In unsupervised learning, the system seeks structure without predefined labels. In reinforcement learning, an agent learns through interaction and feedback related to rewards or objectives.

Statistical systems are often effective for perception and prediction, especially where large representative datasets are available. Their limitations include dependence on data quality, sensitivity to distribution changes, difficulty explaining complex decisions and the possibility of learning undesirable correlations.

Neural and Deep Learning

Neural networks consist of interconnected computational units organized into layers. Deep learning uses networks with multiple learned layers to construct increasingly abstract representations. It has transformed image recognition, speech processing, language modeling and scientific prediction.

Deep models can learn useful representations without manually designed features, but they generally require substantial data and computational resources. They may also be difficult to interpret and can produce confident errors.

Hybrid AI

Hybrid AI combines symbolic and statistical methods. A neural model may interpret images or language, while a symbolic component applies constraints, rules or planning. A learning system may propose actions that a verification module checks before execution.

Hybrid systems are attractive because many practical tasks require both pattern recognition and structured reasoning. For example, an industrial maintenance system may use machine learning to detect abnormal sensor patterns and a rule-based safety layer to prevent unsafe recommendations.

Narrow AI and Artificial General Intelligence

Narrow AI, also called weak AI, is designed for a limited task or domain. Examples include spam filters, route planners, face-recognition systems and recommendation engines. A narrow system may exceed human performance on its specific task while lacking broad understanding outside it.

Artificial general intelligence (AGI) refers to a hypothetical or research goal involving flexible, general-purpose intellectual abilities across many domains. There is no universally agreed test or definition of AGI. Human-level performance on selected benchmarks would not by itself establish general intelligence, because a system might exploit the structure of a benchmark without possessing robust transfer, common sense or autonomous learning.

The distinction is important for responsible communication. Present systems can be highly capable without being generally intelligent. Claims about future AGI involve technical, philosophical and governance uncertainties and should not be confused with measured performance on narrow tasks.

What AI Can and Cannot Do

AI is particularly effective when:

• the task has a clear objective;
• relevant data are available;
• training examples represent deployment conditions;
• performance can be measured;
• the operating environment is sufficiently stable;
• errors have manageable consequences or can be reviewed.

AI is less reliable when tasks require deep contextual judgment, causal understanding, open-ended common sense, rare-event reasoning or moral responsibility. A model may identify statistical associations without knowing whether one event causes another. It may also fail under distribution shift, when real-world inputs differ from training data.

Common limitations include:

1. Data limitations: missing, biased, outdated or incorrectly labeled data.

2. Generalization failures: poor performance on unfamiliar populations or conditions.

3. Distribution shift: changes in behavior, equipment, language or environment after deployment.

4. Spurious correlation: reliance on an irrelevant feature that happens to predict the target.

5. Hallucination or fabrication: generation of plausible but unsupported content.

6. Opacity: difficulty determining why a complex model produced an output.

7. Security vulnerabilities: adversarial inputs, data poisoning, model theft and prompt manipulation.

8. Resource demands: substantial energy, hardware, data and specialized expertise.

9. Automation bias: excessive human trust in machine outputs.

10. Objective mismatch: optimizing a measurable target that does not fully represent the real goal.

These limitations mean that AI should be treated as part of a sociotechnical system. The system includes people, procedures, data pipelines, interfaces, organizational incentives and legal obligations.

A Framework for Evaluating an AI System

Before adopting an AI application, an organization should ask:

1. What decision or task is being supported?

2. Who is affected by errors and who is accountable for outcomes?

3. What data are used, and are they relevant, lawful and representative?

4. What performance metric is appropriate, and what trade-offs does it hide?

5. How will the system behave under unusual or changing conditions?

6. Can a qualified person review, override or appeal an output?

7. How will performance, fairness, security and drift be monitored?

8. What is the fallback procedure if the system fails?

Accuracy alone is rarely sufficient. In a medical or credit context, false positives and false negatives may have different costs. In a safety-critical application, reliability, traceability and fail-safe behavior may be more important than small gains in average accuracy.

Illustrative Case Study: AI-Assisted Loan Screening

Background

A fictional regional bank, Northbridge Bank, considers using a machine-learning system to prioritize small-business loan applications for human review. The proposed system would estimate the probability that an applicant will repay a loan using historical application and repayment data.

Problem

The bank wants faster processing and consistent decisions. However, historical data may reflect unequal access to credit. If the model learns these patterns, it could disadvantage applicants from particular communities. A highly predictive system could therefore still be inappropriate.

Analysis

The bank identifies several risks. First, the target variable—repayment—does not fully represent creditworthiness because previous applicants may have received different loan amounts or terms. Second, some variables may act as proxies for protected characteristics. Third, economic conditions may change, making historical relationships unreliable. Fourth, applicants need understandable reasons for adverse decisions.

Approach and Solution

The bank establishes a multidisciplinary review involving credit officers, data scientists, legal specialists, compliance staff and community representatives. It defines a limited purpose: prioritizing applications for review rather than automatically rejecting them. The team documents data sources, removes unjustified variables, tests performance across relevant groups and evaluates error rates rather than relying only on overall accuracy.

The system produces an explanation suitable for human review, such as insufficient documented cash flow or high existing debt obligations. A credit officer can override the recommendation with a recorded justification. The bank also establishes monitoring for data drift, periodic fairness testing, an appeals process and a manual fallback.

Outcome

In this illustrative case, the bank deploys the system only after a pilot and retains human authority over final decisions. Processing time decreases, but the bank rejects the idea that automation alone solves fairness or accountability problems.

Learning Points

The case demonstrates that AI adoption is a governance and process-design problem as well as a modeling problem. The choice of target, data, threshold, review procedure and accountability structure can matter as much as the algorithm. A technically accurate model may still be unsuitable if its objective is incomplete or its consequences are unacceptable.

Ethical, Legal and Governance Considerations

AI systems can affect privacy, equality, employment, safety, freedom of expression and access to essential services. Responsible AI therefore includes:

• lawful and proportionate data collection;
• privacy protection and secure handling of personal information;
• testing for discriminatory outcomes;
• transparency about automated processing;
• human oversight appropriate to the risk;
• documentation of data, models, limitations and decisions;
• mechanisms for correction, complaint and appeal;
• cybersecurity and incident response;
• environmental and resource considerations.

Regulatory approaches increasingly classify AI applications according to risk. High-impact systems may require stronger documentation, testing, monitoring and human oversight than low-risk recommendation tools. Exact legal duties vary by jurisdiction, so professionals must consult current legislation and organizational policies rather than assuming that a general ethical principle is legally sufficient.

Contemporary Best Practices

A responsible AI development lifecycle commonly includes:

1. Define the problem and determine whether AI is necessary.

2. Identify stakeholders, harms, constraints and success criteria.

3. Obtain and document appropriate data.

4. Establish a baseline, including a non-AI or simpler method.

5. Train and validate models using suitable procedures.

6. Test accuracy, robustness, fairness, privacy and security.

7. Conduct human-centered usability and failure-mode analysis.

8. Deploy gradually with monitoring and rollback procedures.

9. Record incidents, update the system and reassess its purpose.

Professionals should avoid treating a model as a finished product. Deployment changes the data, incentives and behavior surrounding a system. Continuous evaluation is therefore essential.

Chapter Summary

Artificial intelligence is the study and engineering of computational systems that perceive information, represent knowledge, learn or reason, and act toward goals. Its foundations include mathematics, logic, computer science, statistics, neuroscience, linguistics, psychology, philosophy, economics and control theory.

The field developed through symbolic reasoning, search, expert systems, statistical learning, neural networks and hybrid approaches. Its history includes periods of optimism and disappointment. AI winters resulted from excessive expectations, limited resources, difficult problems and weak deployment economics. Modern resurgence has been supported by data, computing power, improved algorithms and engineering infrastructure.

Symbolic AI emphasizes explicit rules and logic; statistical AI learns patterns from data; deep learning uses multilayer neural networks; and hybrid AI combines learned representations with structured reasoning or constraints. Most current systems are narrow AI, capable within defined tasks but not equivalent to artificial general intelligence.

AI can improve prediction, perception, automation and decision support, but it remains vulnerable to biased data, distribution shift, spurious correlations, security attacks, opacity and objective mismatch. Effective AI practice therefore requires technical evaluation, human oversight, documentation, monitoring, fairness analysis, privacy protection and accountability. The central professional lesson is that AI is a sociotechnical capability whose success depends on the interaction of algorithms, data, people, institutions and governance.

Key Terms

Artificial intelligence: The study and engineering of computational systems that perform tasks involving perception, reasoning, learning or action.

Agent: A system that perceives an environment and selects actions.

Environment: The context or world in which an agent operates.

Machine learning: Methods that enable systems to learn patterns or decisions from data or experience.

Symbolic AI: AI based on explicit symbols, rules, logic and structured representations.

Statistical AI: AI that uses statistical models to infer patterns and make predictions.

Deep learning: Machine learning using multilayer neural networks to learn representations.

Hybrid AI: An approach combining statistical or neural methods with symbolic reasoning or constraints.

Narrow AI: AI designed for a specific task or restricted domain.

Artificial general intelligence: A proposed form of AI with broad, flexible abilities across domains.

Knowledge representation: Encoding facts, concepts, relationships and rules for computational use.

Inference: Deriving conclusions or outputs from information and a model or rule system.

Training: Fitting or configuring a model using data, examples or feedback.

Generalization: Performing effectively on new cases beyond training examples.

Distribution shift: A change between the data used to develop a system and data encountered after deployment.

Explainability: The extent to which system outputs and their reasons can be understood.

AI winter: A period of reduced confidence, funding and activity in AI research or development.

Autonomy: Operation with limited direct human control.

Alignment: Correspondence between an AI system’s behavior and legitimate human objectives or values.

Sociotechnical system: A system formed by the interaction of technology, people, processes and institutions.

Review Questions

Short Answer Questions

1. Define artificial intelligence using the rational-agent perspective.

2. What is the difference between training and inference?

3. Identify two disciplines that contribute to AI and explain their contributions.

4. Distinguish symbolic AI from statistical AI.

5. What is an AI winter, and why does it occur?

Long Answer Questions

1. Explain the major stages in the historical development of AI from symbolic systems to foundation models.

2. Compare narrow AI and artificial general intelligence, including the limitations of current narrow systems.

3. Discuss the strengths and weaknesses of symbolic, statistical and hybrid AI.

4. Explain why data quality, distribution shift and objective design are central to AI reliability.

5. Describe a responsible AI development lifecycle and explain the purpose of each stage.

Analytical / Application Questions

1. A hospital wants to use an AI system to prioritize patients for follow-up. Analyze the data, performance, fairness, oversight and governance questions that should be addressed before deployment.

2. A predictive-maintenance model performs well during testing but poorly after a factory changes its equipment. Identify the likely technical cause and propose a monitoring and remediation plan.

3. Design a hybrid AI solution for an autonomous warehouse robot. Explain which functions should use statistical learning, which should use symbolic rules, and how safety constraints should be enforced.

Multiple Choice Questions with Answers

Question 1

Which perspective defines AI primarily in terms of selecting actions that are appropriate for achieving goals?

A. Acting rationally

B. Thinking humanly

C. Thinking symbolically

D. Acting randomly

Correct Answer: A. Acting rationally

Question 2

What is the main characteristic of narrow AI?

A. It possesses human-like abilities in every intellectual domain.

B. It is designed for a specific task or restricted domain.

C. It cannot use data to improve performance.

D. It must be implemented entirely with hand-written rules.

Correct Answer: B. It is designed for a specific task or restricted domain.

Question 3

Which approach represents knowledge explicitly through rules, logic or ontologies?

A. Deep learning

B. Statistical sampling

C. Symbolic AI

D. Unsupervised clustering

Correct Answer: C. Symbolic AI

Question 4

Which factor strongly contributed to the modern resurgence of AI?

A. The elimination of all data-quality problems

B. Reduced availability of computing resources

C. Large datasets and improved computational hardware

D. The replacement of algorithms by manual decision-making

Correct Answer: C. Large datasets and improved computational hardware

Question 5

What is distribution shift?

A. A change between development data and data encountered after deployment

B. The division of a dataset into exactly two equal parts

C. The conversion of symbolic rules into legal regulations

D. A method for increasing model opacity

Correct Answer: A. A change between development data and data encountered after deployment

Question 6

Which statement best describes hybrid AI?

A. It uses only human experts and no computational models.

B. It combines learned statistical methods with symbolic reasoning or constraints.

C. It is restricted to image classification.

D. It avoids both data and explicit knowledge.

Correct Answer: B. It combines learned statistical methods with symbolic reasoning or constraints.

Question 7

Why can a highly accurate loan-screening model still be inappropriate?

A. Accuracy is never useful in financial applications.

B. Historical data and the target variable may encode unfairness or incomplete definitions of creditworthiness.

C. Machine-learning models cannot process numerical data.

D. Human review always produces less reliable decisions.

Correct Answer: B. Historical data and the target variable may encode unfairness or incomplete definitions of creditworthiness.

Question 8

Which term describes improvement through data, examples or feedback?

A. Learning

B. Compilation

C. Encryption

D. Rendering

Correct Answer: A. Learning

Question 9

What is a principal limitation of modern generative AI systems?

A. They can produce fluent but unsupported or factually incorrect content.

B. They are incapable of producing language.

C. They operate only when every rule is manually specified.

D. They never generalize beyond a single stored example.

Correct Answer: A. They can produce fluent but unsupported or factually incorrect content.

Question 10

Which practice is most appropriate for a high-impact AI system?

A. Deploying it without monitoring once testing is complete

B. Using overall accuracy as the only evaluation measure

C. Providing oversight, documentation, monitoring and an appeal or correction process

D. Removing all human responsibility for outcomes

Correct Answer: C. Providing oversight, documentation, monitoring and an appeal or correction process