Artificial Intelligence (AI)

Artificial intelligence (AI) is the field of computing concerned with developing systems that can perform tasks involving capabilities such as learning, reasoning, perception, language, planning and problem-solving. An AI system uses models, rules or other computational methods to infer outputs from inputs, including predictions, recommendations, decisions or generated content. The Organisation for Economic Co-operation and Development (OECD) defines AI systems in terms of this ability to infer outputs, while recognizing that autonomy and adaptation after deployment vary by system.

In everyday use, AI helps computers interpret a photograph, predict equipment failures, detect suspicious payments, translate speech, recommend a route or produce a draft document. These tasks use different techniques. Some systems reason over explicit rules or search possible solutions; others learn statistical relationships from data. A text-generating chatbot is one example of AI, not the definition of the entire field.

The same term is used for three different objects: the scientific discipline, an AI model or algorithm, and a deployed AI system that combines models with data, software, people, interfaces and operating controls. A model's apparent intelligence does not, by itself, prove that the complete application is reliable, safe or useful in a particular environment. Machine learning, deep learning and generative AI belong within the broader field, but none is synonymous with artificial intelligence.

Meaning and scope

The word intelligence in artificial intelligence refers to computational abilities and observable performance, not proof of human-like thought, awareness or consciousness. The field brings together computer science, mathematics, statistics, knowledge representation, cognitive science and engineering. Its goals and methods have changed over time; there is no single algorithm that defines all AI.

An AI system generally has a task, an environment and inputs. It may extract a pattern from an image, estimate a probability, choose among possible plans or generate text. Whether its output should be trusted depends on the specific population, input quality, operating conditions and consequences of error. A model trained to identify manufacturing defects is not automatically qualified to interpret medical images; a conversational model capable of drafting answers is not necessarily connected to current records or authorized to change them.

It is also important to distinguish AI from ordinary automation. Some traditional programs follow predetermined steps without inference or learning. Some AI systems produce an assessment but do not act. Others combine AI output with conventional software, formal rules, workflow permissions and human approval. A system can be automated without using AI, and AI can be useful without automatically executing decisions.

A quick explanation

A company might use AI to predict which industrial machines require inspection. Sensors provide measurements; a trained model estimates failure risk; the maintenance application compares that estimate with operating constraints; and a technician decides whether to schedule work. The prediction is an AI output. Data collection, permission to create work orders, maintenance decisions and later performance checks belong to the full system. This distinction explains why a promising model test is not the same as a proven operational result.

What artificial intelligence systems do

AI systems transform inputs into outputs that can inform or alter activity in the physical or digital world. The OECD's system definition explicitly encompasses predictions, content, recommendations and decisions. These output categories provide a more useful description of actual applications than a simple distinction between smart and non-smart software.

Output or function	What the system produces	Illustrative application
Recognition and classification	A label, category, detected object or extracted information	An image model identifies a visible defect in a manufactured component.
Prediction and estimation	A forecast, probability, score or numerical estimate	A system estimates equipment-failure risk from sensor data.
Recommendation and optimization	A suggested action, ranking, allocation or plan	A logistics application proposes delivery routes subject to constraints.
Content generation	New text, software code, images, audio or other content	A generative AI assistant drafts a response to a customer inquiry.
Decision or action	A selected response or operation within an authorized environment	A monitored industrial controller adjusts a process within defined limits.

These functions are not mutually exclusive. A single application may recognize a condition, estimate a risk, recommend an intervention and initiate an approved action. Such a sequence requires more than a model: it also needs interfaces, reliable data, authorization rules and a way to detect when the system has behaved incorrectly.

How artificial intelligence works

There is no single mechanism by which all AI operates. Depending on the task, an AI system may use explicitly represented knowledge, search and planning, learned statistical models or combinations of these methods. A useful way to understand the process is to follow the route from a defined objective through inputs and inference to an output and, when authorized, an action or decision.

Part of the system	What happens	Practical question
Objective and context	A task and its operating boundaries are defined.	What result is needed, for whom, and with what consequences if it is wrong?
Inputs	Data, records, images, sensor readings, human knowledge, rules or prompts become available to the system.	Are these inputs relevant, current, authorized and sufficiently representative?
Inference or reasoning	An algorithm or model applies learned parameters, explicit rules, search or other computations.	What method produces the output, and what assumptions does it make?
Output	A prediction, classification, recommendation, generated item or selected action is produced.	How can the output be tested and interpreted?
Use and feedback	An application, person or process acts on or reviews the result; monitoring detects changes and errors.	Who has authority to act, and how will failures be corrected?

Knowledge representation and reasoning

In symbolic AI, information is represented explicitly as facts, objects, rules, relationships or logical statements. An inference system may draw conclusions from these representations. An expert system, for example, can use a knowledge base and formal rules to support troubleshooting or a structured diagnosis. Such systems can be inspectable when their rule base is manageable, but acquiring and maintaining accurate knowledge can be costly. Not every fixed business rule is an AI technique; its classification depends on the task and inference involved.

Search, planning and optimization

AI research also studies how systems explore possible states and choose actions. Search algorithms examine alternatives; planning methods connect actions to goals and constraints; optimization methods look for high-performing solutions under specified conditions. For example, a scheduling system may compare many feasible maintenance sequences and select one that respects staffing and equipment constraints. These methods can be combined with machine learning or operate without a learned neural model. Knowledge representation, automated planning and optimization are related but distinct research areas within the wider technical landscape.

Learning from data

Machine learning fits a model using data, examples or interaction, rather than requiring all useful relationships to be entered manually. During training, model parameters may be adjusted to improve performance against a learning objective. During inference, the trained model processes new inputs. A fraud-review model, for example, may estimate which new transactions merit investigation based on patterns in previously examined cases.

Supervised, unsupervised, self-supervised and reinforcement learning describe different sources of learning signals; they do not form a single ladder of intelligence. A deployed model does not necessarily update itself after every interaction. Model updates, external record retrieval and saved user history are different mechanisms and should not be confused.

Neural networks, deep learning and language models

Artificial neural networks use connected computational units with adjustable weights. Deep learning employs multiple processing layers to learn representations relevant to a task. Modern systems use these methods for language, images, speech and many other tasks, but their suitability depends on data, evaluation and computational constraints.

A large language model typically learns patterns in sequences and can generate or process text. Many modern language models employ the transformer architecture introduced in 2017, which uses attention mechanisms to represent relationships among input elements. A foundation model is a broadly trained model that may be adapted to several tasks. Neither term describes the whole of AI, and a model's generated explanation should not automatically be treated as a reliable account of its own internal computation.

From inference to action

A model may produce a score, sentence or plan; a separate application decides how it is displayed or used. Such an application may check authorization, retrieve current records, validate an output, request a human decision and record an outcome. An AI agent may combine a model with tools and a control loop to pursue a specified objective, but its permissions, connected systems and safeguards come from system design. The ability to generate a response is different from authority to perform an action.

Main approaches and branches of AI

The field is divided along several different dimensions. Some names describe technical methods, others describe tasks or application areas, and still others describe systems built around AI. They should not all be treated as mutually exclusive branches of one taxonomy.

Area	Role in artificial intelligence	Relationship to the wider field
Symbolic AI and expert systems	Represent explicit knowledge and reason over rules, constraints or logical structures.	Method-oriented approaches that can operate independently or combine with learned models.
Search, planning and optimization	Explore alternative actions, reason about constraints or optimize a defined objective.	Computational methods often combined with knowledge representations and learned models.
Machine Learning	Learns patterns or decision functions from data and experience.	A major subset of AI, not the whole field.
Deep Learning	Uses multilayer neural networks to learn representations and perform tasks.	A subset of machine learning.
Natural language processing	Analyzes, transforms or generates human language.	An application and research area that can use rules, statistics or deep learning.
Computer Vision	Analyzes images and video to recognize, locate, measure or interpret visual information.	A task domain employing multiple AI techniques.
Generative AI	Produces text, images, audio, code or other content from learned representations.	A capability class, prominently implemented with machine learning.
Reinforcement Learning	Learns action policies through interaction and feedback, often framed as reward.	A learning paradigm relevant to control, games, robotics and some model-training methods.
AI agents and autonomous systems	Connect perception, planning, tools or actions to pursue objectives within an environment.	A system or application pattern that may use several underlying AI methods.
Robotics	Combines sensing, control and physical action in machines.	An engineering field that may use AI but also includes conventional control and automation.

This classification helps avoid misleading statements such as all AI is machine learning, all deep learning is generative or all robots are AI. Machine learning and deep learning form a methodological hierarchy; language processing, computer vision and robotics cut across methods; generative AI is primarily distinguished by output; autonomous agents are distinguished by system behavior.

Types of AI by capability and autonomy

Two other classifications are common but answer different questions.

Specialized and general-purpose capabilities

Specialized or narrow AI refers to systems developed and evaluated for particular tasks, even when a model can perform many related tasks. Many deployed AI applications are specialized through their training, configuration, tools, operational restrictions or testing. Broadly capable foundation models may support numerous tasks, but versatility on some tests does not demonstrate unrestricted competence across all situations.

Artificial general intelligence (AGI) is a debated concept for AI with broad, adaptable cognitive capabilities comparable to or beyond those of humans across a wide range of activities. There is no single universally accepted operational test for declaring AGI. Artificial superintelligence is a hypothetical concept describing capabilities surpassing humans across many areas. Neither term is a necessary category for describing an ordinary enterprise AI deployment.

Autonomy and adaptability

AI systems can range from tools that only make recommendations to systems permitted to take actions without case-by-case approval. Their level of autonomy depends on design, connected capabilities, authority and context. Separately, a system's ability to adapt after deployment concerns whether and how its behavior changes. These are distinct attributes; the OECD definition explicitly notes that both vary among AI systems.

Autonomy should therefore be described in terms of observable permissions and safeguards: what a system can access, what it may change, which actions require authorization and how it can be interrupted. Assertions that an AI system thinks for itself provide little useful information about those conditions.

Components of an operational AI system

The exact architecture depends on the task. Nevertheless, an enterprise system typically brings together several components that have different owners and distinct failure modes.

Component	Purpose	What can go wrong
Objective and operating context	Define the task, users, environment, acceptable outcomes and limits of authority.	A technically capable model optimizes an objective that does not match the actual business need.
Data or represented knowledge	Supply observations, training examples, documents, rules or contextual information.	Missing, stale, biased, improperly licensed or inaccessible information distorts results.
Algorithms and models	Perform inference, learning, search, generation, ranking or planning.	Performance fails on unfamiliar conditions, or outputs are unreliable or poorly calibrated.
Applications, tools and infrastructure	Connect models to interfaces, databases, business processes, computing resources and permitted actions.	Access control, reliability, integration or dependency failures produce operational harm.
Evaluation and monitoring	Check expected performance, adverse outcomes, drift, security and continued fitness for purpose.	A model passes a laboratory test yet fails in real workflows or changing conditions.
People, accountability and controls	Assign ownership, approvals, oversight, incident response and the ability to correct or stop use.	No one can explain, challenge or reverse a consequential automated action.

A model can be purchased from a provider, built internally or incorporated as a feature of existing software. The organization still needs to know the complete system boundary to judge confidentiality, dependencies, accuracy, legal obligations and business value. These concerns are especially important when an AI capability is embedded in a product that was not originally bought or classified as an artificial intelligence system.

AI system lifecycle

Artificial intelligence has a lifecycle that extends beyond model training. The stages below describe recurring work and decision boundaries, not a mandatory linear method or a substitute for a full implementation process.

1. Define the problem and context. Establish whose problem is being solved, the decision or service affected, the permitted role of automation, relevant users and foreseeable harms.
2. Assess data and feasibility. Determine whether suitable information, lawful rights, operating conditions and an appropriate technical approach exist.
3. Develop, configure or acquire. Build or obtain a model and design the surrounding workflow, access controls and integration.
4. Evaluate before release. Test task performance, robustness, security, failure cases and human ability to understand or intervene where necessary.
5. Deploy with controls. Assign operational owners, approvals, user training, monitoring and rollback arrangements.
6. Monitor in operation. Measure business results and adverse outcomes; respond to data drift, model or vendor changes and new threats.
7. Retire, replace or revise. Withdraw a system that is no longer appropriate, and manage records, access, dependent processes and affected users.

Learning and feedback do not end at deployment. A system can change because the environment changes even if the model parameters do not. Conversely, a provider may update a model while the organization's application and workflow remain nominally unchanged. Both situations create reassessment obligations; the relevant process is governed by AI governance and the organization's change and risk management arrangements.

Applications of artificial intelligence

Artificial intelligence is used across information-intensive, decision-intensive and some physical processes. An application should be classified by its actual task, decision consequences and operating evidence rather than by industry branding. The examples below illustrate use categories; they do not imply that every deployment is effective or suitable.

Business or public-service setting	Representative AI use	Critical context or limitation
Customer operations	Classify service requests, suggest responses, translate messages or summarize interactions.	Require appropriate escalation, privacy controls and checks on incorrect advice.
Financial services	Detect unusual transactions, estimate risks and support fraud review.	False positives, fairness, explanation and regulatory obligations may be consequential.
Manufacturing and maintenance	Analyze images for defects or estimate equipment-failure risk from sensor patterns.	Poor data quality or changing equipment conditions can undermine performance.
Healthcare	Assist image interpretation, administrative documentation or clinical research.	Clinical validity, human responsibility and applicable medical regulation are essential.
Logistics and supply chains	Forecast demand, support routing or optimize resource allocation.	Results depend on forecast quality, constraints, disruption handling and operational feasibility.
Cybersecurity	Identify anomalies, prioritize alerts or assist analysts with triage.	Attackers can manipulate inputs; overreliance may hide critical events.
Knowledge and document work	Search, classify, extract, summarize or generate document drafts.	Source completeness, citations, access rights and output review remain important.
Public administration	Triage requests, detect inconsistent records or assist caseworkers.	Due process, accessibility, transparency and the right to contest decisions can be central.

Applications frequently blend AI with existing enterprise software and human expertise. For example, an AI-enhanced service desk may use conventional workflow rules for routing, a classifier for request types and a language model for drafting. The completed service depends on the quality of the entire design, not just the model used in one step.

Why organizations use AI

Organizations use AI when its pattern recognition, prediction, generation, search or reasoning capabilities can improve a defined activity or enable a new one. Possible outcomes include faster access to information, better identification of anomalies, more accurate forecasts, improved accessibility, more responsive service and new scientific or commercial capabilities. The benefit is conditional: integration, human review, system reliability, data rights and the cost of errors can change the result.

The Stanford Institute for Human-Centered Artificial Intelligence's 2026 AI Index tracks technical performance, adoption, economic effects and responsible AI separately. That separation matters. A technology may improve rapidly on selected benchmarks while organizational benefits, public trust or safety practices follow a different path. The report also documents substantial variation among tasks and continuing failure on some evaluations despite strong results on others.

For an organization, AI creates value only to the extent that it improves an outcome relative to a credible alternative, after allowing for implementation, monitoring, infrastructure and oversight. A system that performs a task faster but introduces rework or unacceptable errors may produce no net gain. Likewise, a scientific model can make an important research contribution without immediately yielding a commercial product or clinical benefit.

Documented uses and what the evidence actually shows

Real-world examples show that different forms of AI evidence establish different claims. A public inventory can confirm that a use exists; a regulatory authorization relates to a defined intended use; a scientific database can demonstrate technical output and research access; and a workplace study can quantify a measured outcome in its observed setting. These forms of evidence should not be treated as interchangeable.

Documented example	What is established	What it does not establish by itself
Protein-structure prediction: AlphaFold	Google DeepMind and EMBL's European Bioinformatics Institute provide the AlphaFold Protein Structure Database, which makes more than 200 million predicted protein structures available for research.	A prediction for every protein is not a laboratory-confirmed structure, and database availability is not evidence of a cure or treatment outcome.
Authorized medical devices	The United States Food and Drug Administration maintains a public list of AI-enabled medical devices authorized for marketing under applicable review pathways. The FDA reported more than 1,600 such devices as of September 2026.	An authorization is specific to a device and its intended use; it does not prove that every deployment improves clinical outcomes or that every AI medical product appears on the list.
Government procurement analytics	The U.S. General Services Administration's published 2025 AI use-case inventory identifies Acquisition Analytics as a deployed classification application for government spending records.	Inventory inclusion does not by itself verify measured cost savings, error reduction or performance across agencies.
Customer-support assistance	A peer-reviewed study of 5,172 support agents reported an average 15% increase in issues resolved per hour after access to an AI conversational assistant, with heterogeneous effects across experience levels.	The estimate applies to the studied setting and its productivity measure; it cannot be generalized to all AI tools, occupations or organizations without further evidence.

These examples span scientific discovery, regulated devices, public administration and workplace assistance. They also illustrate why the question Does AI work? needs a more precise object: which system, task, users, baseline, observation period and definition of success? A capability demonstration, an approved deployment and a measured outcome should remain separate claims.

Limitations of artificial intelligence

AI can perform very well on some tasks and remain unexpectedly fragile on others. Technical progress should therefore be assessed task by task, rather than by treating one strong benchmark as a guarantee of broad competence. Important limitations depend on the method and use context:

• Errors and unreliable outputs. A model may misclassify an input, invent a plausible statement, misinterpret context or produce inconsistent answers.
• Generalization limits. Performance on training data, a benchmark or a pilot may not carry over to another population, location or operating process.
• Dependence on data and knowledge. Incorrect labels, missing information, changing conditions and poorly defined rules can produce unreliable outputs.
• Incomplete explanations. Some AI techniques make decisions that are difficult to interpret or justify to an affected person. An explanation generated by a model is not necessarily a faithful account of its internal mechanism.
• Costs and physical constraints. Data preparation, computing, integration, supervision, energy consumption and maintenance may outweigh a narrow benefit.
• Limits of autonomy. Producing a recommendation is different from having the authority or ability to execute it safely.

These weaknesses do not affect all AI systems equally. A deterministic rules engine and a large generative model have different error patterns, while an automated physical controller may introduce direct safety consequences absent from a text drafting assistant. Evaluation must use the particular system's purpose and exposure as its reference point.

Risks, safety and responsible AI

Risks arise from both technical behavior and how AI changes decisions, information flows and human accountability. The NIST AI Risk Management Framework is a voluntary resource for identifying and managing AI risks; the OECD AI Principles provide internationally agreed guidance on trustworthy and human-centric AI. Neither replaces sector-specific legal requirements or local evaluation of a particular use.

Relevant risks include:

• Accuracy and reliability: wrong outputs, unsupported confidence, fragile behavior and changes in performance over time.
• Bias and unequal effects: data or design choices may systematically disadvantage groups or fail to represent the people affected.
• Privacy and confidentiality: sensitive or protected information may be collected, exposed or reused contrary to policy or law.
• Security and manipulation: malicious inputs, model or data tampering, unauthorized tool use and compromise of connected systems.
• Transparency and contestability: affected people may be unable to tell that AI was involved, understand a decision or obtain correction.
• Intellectual property and information integrity: generated content may raise ownership, licensing, provenance or misinformation concerns.
• Safety and systemic effects: errors can have physical, financial, public-service or organizational consequences that extend beyond the model itself.

Risk treatment should be proportional to the application. A summary of a public meeting and an AI-supported eligibility decision have different consequences, evidentiary requirements and oversight needs. Highly consequential decisions generally require more stringent validation, human accountability and procedures for challenge and correction. Responsible AI describes related objectives and practices; AI governance concerns the structures through which those objectives and obligations are made operational.

AI governance, standards and regulation

Organizations govern artificial intelligence through assigned decision rights, acceptable-use rules, system inventories, risk assessments, testing, procurement controls, monitoring, incident handling and independent review as appropriate. This governance spans business, technology, legal, data, privacy, security and risk functions; accountability cannot be assigned simply by naming a technical model owner.

Several standards and frameworks help establish a shared language and operating discipline:

• NIST AI RMF 1.0: a voluntary risk management framework organized around the functions Govern, Map, Measure and Manage. Its version and applicability should be checked when it is used.
• ISO/IEC 42001:2023: an international standard specifying requirements for an organizational artificial intelligence management system, including its establishment, operation and continual improvement. It is a management-system standard, not a certification that any particular model's answer is correct.
• ISO/IEC 22989:2022: a terminology standard describing concepts used in artificial intelligence, useful for distinguishing system, model and related technical terms.
• OECD AI Principles: nonbinding intergovernmental principles addressing human rights, transparency, robustness, safety and accountability, updated in 2024.

Regulation is jurisdiction-specific. The European Union's AI Act, for example, establishes obligations and prohibitions that depend on a system's category, provider or deployer role and applicability dates. Its implementation timetable extends beyond 2026 for some high-risk applications; a general AI definition page should not be treated as a legal applicability determination.

AI governance is related to, but distinct from, AI strategy and AI roadmaps. Strategy defines the intended role and value of AI; a roadmap sequences investments and implementation; governance determines who can authorize and control the use of AI, on what evidence, and with what accountability.

Measuring AI performance and value

An evaluation should distinguish technical performance, operational results and harm or control outcomes. A high score on a technical benchmark does not automatically establish that an AI application improves a live business process. Likewise, a faster workflow is not necessarily better if its error rate or exposure rises.

Evidence level	Illustrative measures	Interpretation caution
Model or technical task	Accuracy, precision, recall, calibration, error rates, robustness, latency or task success, as applicable.	Compare with a relevant baseline and test population; one benchmark does not describe all settings.
Operational service	Human review rate, exception volume, rework, time to completion, service reliability and escalation outcomes.	Check whether new work or displaced effort has merely moved elsewhere.
Business value	Quality-adjusted throughput, customer outcomes, avoided loss, cost to serve or other purpose-specific results.	Measure net effects after integration, monitoring, infrastructure and human oversight costs.
Safety, rights and controls	Adverse incidents, privacy events, unresolved complaints, fairness testing, audit findings and override effectiveness.	Low reported incident counts may reflect poor detection rather than low underlying risk.

Metrics must be defined before interpreting results: denominator, population, measurement window, comparison group, acceptable error and decision consequence matter. When causes are uncertain, observed changes should not be presented as proven benefits of the AI system. Where possible, monitoring should also capture model versions, input shifts and changes to the surrounding business workflow.

Illustrative example: AI-assisted service request routing

Consider a hypothetical company receiving thousands of internal IT requests every month. Staff currently read each request, identify its category, decide which support team should receive it and request missing details. The organization introduces an AI-assisted routing function within its IT service management application.

The system uses a trained classifier to suggest categories and routing teams from request text and historical labeled records. A separate language component may propose a clarification question when essential information is missing. The application checks access permissions and required fields, and routes ambiguous or high-impact cases to a human service agent. It does not acquire unlimited authority to approve sensitive changes merely because it can infer a category.

Before deployment, the team evaluates model performance on representative requests, including uncommon incidents and requests from different business units. After deployment, it compares correctly routed cases, rework, elapsed time and user outcomes against a suitable baseline. Incident records, overrides and model or vendor updates become part of monitoring. The example illustrates the distinction between an AI technique (classification), an AI-enabled application (service routing) and the surrounding process that makes a result usable. No numerical productivity claim is implied by this hypothetical case.

Relationship to other technology and management concepts

Artificial intelligence sits within a broader technical and organizational environment. The following distinctions are important for understanding related pages without making this entry an independent implementation guide for each one.

Related concept	Distinction from artificial intelligence
Machine Learning	A major AI method that learns from data; AI also includes knowledge-based reasoning, search and other techniques.
Deep Learning	A machine-learning approach based on multilayer neural networks, not a synonym for every AI system.
Generative AI	A capability class that generates content, whereas AI includes prediction, classification, recommendation and action.
Large language model	A model type especially associated with language tasks; it can be a component of larger AI systems.
AI agent	A system or application that pursues objectives and may use tools; a model alone does not define its permissions or actions.
Robotic process automation	Frequently automates predefined business steps; it may incorporate AI but is not inherently an AI method.
Robotics	Concerns physical machines, sensing and actuation; some robots use AI while others rely on conventional control.
Artificial general intelligence	A debated hypothetical category of broadly general capability, distinct from a deployed task-focused AI system.
AI Strategy	An organizational set of choices about why and where to use AI, rather than the AI techniques themselves.
AI Governance	The rules, authority and oversight applied to development, acquisition and use of AI systems.

Development of artificial intelligence

Artificial intelligence developed through several overlapping research traditions: logic and symbolic reasoning, search and planning, statistical learning, neural networks, robotics and human-computer interaction. It has not advanced through a single predetermined series of stages, and ideas considered obsolete in one period may later become useful when combined with newer techniques.

Period or event	Importance for the field
1950: machine intelligence as a research question	Alan Turing's Computing Machinery and Intelligence helped frame questions about whether machine behavior could be evaluated in relation to intelligence.
1955-1956: Dartmouth research project	The proposal by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon and the 1956 workshop helped establish artificial intelligence as the name of an organized research field.
1970s-1980s: knowledge-based systems	Researchers and organizations developed expert systems and logic-based approaches that demonstrated the value and maintenance challenges of explicit knowledge representations.
1990s-2010s: learning at scale	Statistical learning and later deep neural networks benefited from improved algorithms, computing and large data resources, expanding applications in vision, speech and language.
2017: transformer architecture	The transformer research paper introduced an attention-based architecture that subsequently became influential for large-scale language modeling.
2020s: foundation models and wider deployment	Broadly trained models and accessible generative AI applications expanded public and organizational use, while increasing scrutiny of data use, reliability, security, energy demand and accountability.

These milestones mark important developments rather than a complete history. A strong performance result on a chess game, coding test or laboratory benchmark does not by itself demonstrate general human-level intelligence; it demonstrates competence under the particular evaluation conditions. Concepts such as artificial general intelligence and superintelligence remain debated and should be distinguished from the performance and limitations of deployed systems.

Frequently asked questions

What is artificial intelligence in simple terms?

Artificial intelligence allows a computer or machine to make predictions, recognize patterns, work with language, reason over information or produce content using computational methods that approximate capabilities associated with intelligence. Its usefulness depends on the task and the system in which the technique operates.

Is artificial intelligence the same as machine learning?

No. Machine learning is an important part of AI in which models are fitted from data. AI also includes explicit knowledge representations, reasoning, search, planning and optimization methods that need not be based on modern deep learning.

Are generative AI and large language models the same as AI?

No. Generative AI emphasizes content production; a large language model is a particular type of model. AI is the wider field, encompassing prediction, classification, planning, perception, recommendation and other capabilities. A model may be incorporated into many different applications.

How does AI differ from ordinary automation?

Conventional automation can run predetermined workflow steps. AI supplies inference, reasoning or learned pattern-based capabilities that may be used by an automated system. An AI output can also remain advisory, with a person making the final decision. Neither intelligence nor permission to act follows automatically from the presence of automation.

Does AI learn from every interaction?

Not necessarily. A deployed model may stay unchanged until it is retrained or updated. A service might remember a conversation or retrieve current records without changing the model itself. The answer depends on the application and its model-management arrangements.

Can AI be accurate on one task and unreliable on another?

Yes. Performance depends on the data, task, test population and operating conditions. A strong result in a laboratory, on a benchmark or in one department does not guarantee reliable performance elsewhere. Meaningful evaluation compares the proposed use with a suitable baseline and checks consequential errors.

Can AI make decisions autonomously?

Some AI-enabled systems are designed to recommend actions; others can take defined actions through connected software or machines. Their actual autonomy depends on authorized permissions, operating context, safeguards and oversight, not simply on a model's ability to generate an answer.

Will AI replace human jobs?

AI can automate, support or change particular tasks. Employment effects depend on adoption, how work is reorganized, complementary skills, demand and other economic conditions. Evidence from one use case or a model benchmark is insufficient to predict overall job displacement.

See also

• Machine Learning
• Deep Learning
• Artificial Neural Network (ANN)
• Natural Language Processing (NLP)
• Computer Vision
• Generative AI
• Large Language Model (LLM)
• AI Agent
• AI Strategy
• AI Governance
• AI Roadmap
• Artificial General Intelligence (AGI)

Further reading

• CIO Index: Artificial Intelligence Resource Library