What Exactly Is AI?

Defining Artificial Intelligence: Beyond the Hype

At its foundational level, artificial intelligence (AI) refers to engineered computational systems designed to perform tasks that typically require biological intelligence, such as visual perception, natural language interpretation, decision-making, pattern recognition, and adaptive problem-solving. Rather than indicating an artificial mind possessing consciousness, intent, or self-awareness, AI encompasses a broad spectrum of mathematical frameworks, statistical algorithms, and computational logic. These systems ingest digital inputs—ranging from structured tabular records and pixel arrays to raw text and audio waveforms—and produce actionable inferences, classifications, predictions, or synthetic generations based on learned or programmed associations. What Is Artificial Intelligence (AI)? artificial intelligence - Glossary | CSRC

Technical consensus across international standard-setting bodies clarifies that an AI system is an artifact governed by software and data. According to the Organisation for Economic Co-operation and Development (OECD), an AI system is a machine-based system that, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. This operational view treats intelligence not as a metaphysical essence, but as a demonstrable set of functional capabilities: processing information, generalizing beyond exact training examples, adapting to variable operating environments, and optimizing toward defined mathematical objectives without requiring step-by-step human programming for every possible permutation. artificial intelligence - Glossary | CSRC Recommendation of the Council on Artificial Intelligence ISO AI Concepts and Terminology Standard

Computer science textbooks and academic philosophy commonly organize AI definitions across two distinct dimensions: thinking versus acting, and matching human performance versus adhering to ideal rationality. Russell and Norvig’s classic taxonomy distinguishes between systems that think like humans (cognitive modeling), systems that act like humans (the behavioral criterion exemplified by the Turing Test), systems that think rationally (formal logic and deductive reasoning), and systems that act rationally (computational agents designed to maximize expected utility). Modern engineering predominantly centers on rational agency—constructing computational systems that calculate and execute the actions most likely to achieve an intended objective, given available data and bounded computational resources. What Is Artificial Intelligence (AI)? Artificial Intelligence - Stanford Encyclopedia of Philosophy

The Core Mechanics: How AI Actually Works

To understand what AI actually is, one must differentiate between the two foundational paradigms that have defined the discipline since its inception at the 1956 Dartmouth workshop: symbolic AI and statistical machine learning. While early research treated intelligence as a manipulation of explicit linguistic rules and logic trees, modern deployments rely almost entirely on continuous numerical optimization over high-dimensional data spaces. What Is Artificial Intelligence (AI)? Artificial Intelligence - Stanford Encyclopedia of Philosophy

Symbolic Systems vs. Statistical Learning

The classical era of artificial intelligence—often termed "Good Old-Fashioned AI" (GOFAI) or symbolic AI—relied on deterministic, hand-crafted rules, ontologies, and formal deductive logic. In an expert system designed to diagnose infectious diseases, for example, human domain specialists had to explicitly transcribe every medical rule: IF fever > 38.5 AND white_blood_cell_count > 12000 THEN consider bacterial_infection. While symbolic systems are fully interpretable, they fail when encountering noisy real-world data, unmodeled exceptions, ambiguous language, or complex sensory input like images. What Is Artificial Intelligence (AI)? Artificial Intelligence - Stanford Encyclopedia of Philosophy

Modern AI inverts this methodology through machine learning (ML). Instead of human engineers programming explicit instructions for mapping an input $X$ to an output $Y$, engineers build mathematical architectures containing millions or billions of adjustable numerical parameters, known as weights and biases. The algorithm is exposed to extensive datasets, learning the underlying statistical function directly through numerical optimization. Rather than following rigid logic, the system discovers correlation structures across multidimensional feature spaces. What Is Artificial Intelligence (AI)? ISO AI Concepts and Terminology Standard

Traditional Programming:
  [Data] + [Rules / Program]  ----->  [Output / Answers]

Machine Learning (Statistical AI):
  [Data] + [Output / Answers] ----->  [Trained Model / Rules]

The Optimization Loop: Data, Weights, and Loss Functions

At run time, what appears to be "reasoning" or "intuition" inside a modern neural network is a high-speed matrix calculation governed by a loss function and an optimization algorithm:

1. Forward Propagation: The model accepts an input vector (such as an array of token IDs or pixel values) and passes it through successive layers of mathematical operations. Each connection multiplies the input by a weight, sums the values, adds a bias, and applies a non-linear activation function (such as ReLU or GeLU).
2. Error Measurement (Loss Function): The model produces a candidate output. A mathematical metric, such as cross-entropy loss or mean squared error, computes the difference between the model's generated output and the ground-truth target.
3. Credit Assignment (Backpropagation): Using the chain rule of calculus, the algorithm calculates the partial derivative of the error with respect to every single parameter in the network, establishing which weights contributed to the error and in what direction.
4. Parameter Adjustment (Gradient Descent): An optimizer nudges the weights down the gradient slope by a small increment governed by a learning rate. Over iterations across massive datasets, the network minimizes its average error, effectively shaping a high-dimensional surface that maps inputs to probable outputs. What Is Artificial Intelligence (AI)? Artificial Intelligence - Stanford Encyclopedia of Philosophy

The Spectrum of Capabilities: Narrow, General, and Superintelligence

Discussions surrounding artificial intelligence frequently conflate what software can do today with theoretical constructs of what future machines might do. Formal taxonomies distinguish between three discrete developmental tiers of machine capability. ISO AI Concepts and Terminology Standard Artificial Intelligence - Stanford Encyclopedia of Philosophy

Every commercially available and operational system today belongs to the category of Artificial Narrow Intelligence (ANI). Even Large Language Models (LLMs), which exhibit broad surface-level versatility across translation, summarization, and coding, remain domain-bounded statistical engines. They operate strictly within the realm of textual and multi-modal token prediction without cross-domain cognitive integration, physical embodiment, or genuine autonomous intentionality. What Is Artificial Intelligence (AI)? ISO AI Concepts and Terminology Standard

What AI Is Not: Clearing Common Misconceptions

Understanding what artificial intelligence is requires establishing what it is not. Pop-culture depictions and anthropomorphic marketing have contributed to several pervasive fallacies regarding machine cognition. What Is Artificial Intelligence (AI)? Artificial Intelligence - Stanford Encyclopedia of Philosophy

Consciousness and Sentience

An AI model possesses no inner experience, subjective awareness (qualia), feelings, or selfhood. When a conversational agent outputs the sentence "I am happy to assist you," it does not experience the emotion of happiness or the desire to assist. The sentence is generated because the statistical probabilities across its parameters determine that those specific tokens form an appropriate, expected completion to the user prompt. Simulating the linguistic markers of emotional intelligence or intentionality is computationally distinct from experiencing them. What Is Artificial Intelligence (AI)? Artificial Intelligence - Stanford Encyclopedia of Philosophy

"The question of whether a computer can think is no more interesting than the question of whether a submarine can swim."
— Edsger W. Dijkstra

Deep Semantic Understanding

Human language understanding integrates syntactic structure with embodied sensory experience, causal models of reality, and social context. In contrast, modern generative models operate as sophisticated pattern-matching systems. While they encode intricate representations of grammar, factual relationships, and discursive styles, they do not verify external reality or maintain an ontological belief system. When an AI generates a plausible sounding but factually incorrect statement—an event known as a hallucination—it has not committed an act of deceit; it has merely sampled tokens from a probability distribution that diverged from empirical fact. What Is Artificial Intelligence (AI)? artificial intelligence - Glossary | CSRC

The Database Fallacy

Conversely, AI is not simply a search engine or a relational database retrieving pre-written answers. A deep learning model does not store copies of its training data in an organized archive. Instead, it compresses features, relationships, and linguistic structures into billions of numerical parameters. When an image generator produces an illustration of an astronaut on horseback, it is not pasting existing photographs together; it synthesizes a completely new pixel distribution guided by learned statistical correlations of "astronaut," "horse," and compositional styles. What Is Artificial Intelligence (AI)? ISO AI Concepts and Terminology Standard

Major Learning Paradigms and Architectural Foundations

The algorithmic universe of artificial intelligence is divided into operational paradigms based on how models acquire data and receive training signals. Modern applications often chain several of these paradigms together into complex computational pipelines. What Is Artificial Intelligence (AI)? ISO AI Concepts and Terminology Standard

Supervised, Unsupervised, and Reinforcement Learning

                       ┌─────────────────────────┐
                       │   Learning Paradigms    │
                       └────────────┬────────────┘
         ┌──────────────────────────┼──────────────────────────┐
         ▼                          ▼                          ▼
┌─────────────────┐        ┌─────────────────┐        ┌─────────────────┐
│   Supervised    │        │  Unsupervised   │        │  Reinforcement  │
│    Learning     │        │    Learning     │        │    Learning     │
├─────────────────┤        ├─────────────────┤        ├─────────────────┤
│ Labeled Data    │        │ Unlabeled Data  │        │ Reward Signals  │
│ (Input -> Label)│        │ (Clustering,    │        │ (Trial & Error  │
│ Classification, │        │  Latent Space)  │        │  in Dynamic     │
│ Regression      │        │ Dimensionality  │        │  Environments)  │
└─────────────────┘        └─────────────────┘        └─────────────────┘

• Supervised Learning: The algorithm is supplied with paired training data consisting of inputs and corresponding ground-truth labels $(x_i, y_i)$. Tasks include image classification (tagging chest X-rays as healthy or diseased) and numeric regression (estimating housing values based on square footage and location). What Is Artificial Intelligence (AI)? ISO AI Concepts and Terminology Standard
• Unsupervised Learning: The algorithm processes raw, unlabeled data $(x_i)$ to discover hidden structure, underlying clusters, or low-dimensional representations without human intervention. Techniques such as principal component analysis (PCA), k-means clustering, and autoencoders fall into this category. What Is Artificial Intelligence (AI)? ISO AI Concepts and Terminology Standard
• Reinforcement Learning (RL): An autonomous agent interacts with an environment, observing states, selecting actions, and receiving scalar rewards or penalties. Through trial and error, the agent learns an optimal policy $\pi$ to maximize cumulative reward over time. RL is the engine behind game-playing systems (such as chess and Go engines) and robotics pathfinding. What Is Artificial Intelligence (AI)? Artificial Intelligence - Stanford Encyclopedia of Philosophy
• Self-Supervised Learning: The dominant paradigm underlying modern foundation models. The system creates its own supervisory signals from raw, uncurated data by masking portions of the input and training the model to predict the missing segments (for instance, predicting the next word in a sequence or the missing patch of an image). What Is Artificial Intelligence (AI)? ISO AI Concepts and Terminology Standard

The Transformer and Attention Mechanisms

The contemporary leap in natural language processing and multimodal AI stems from the Transformer architecture. Prior neural architectures, such as Recurrent Neural Networks (RNNs) and Long Short-Term Memory networks (LSTMs), processed sequential information step-by-step, which created computational bottlenecks and limited their ability to retain long-range context. What Is Artificial Intelligence (AI)?

The Transformer replaced recurrent loops with the self-attention mechanism. This allows a network to process entire sequences in parallel, computing mathematical affinity scores between every word (or token) in a sentence simultaneously. In the phrase "The animal didn't cross the street because it was too tired," the attention mechanism calculates an affinity weight showing that the token "it" correlates strongly with "animal," whereas in "because it was too wide," "it" connects to "street." By scaling this mechanism across dozens of layers and hundreds of billions of parameters, models construct rich syntactic, semantic, and contextual representations of human knowledge. What Is Artificial Intelligence (AI)? Artificial Intelligence - Stanford Encyclopedia of Philosophy

Practical Implications, Boundaries, and Governance

Because artificial intelligence has transitioned from theoretical computer science into pervasive infrastructure, defining what AI is also requires identifying its operational constraints and failure modes in human society. artificial intelligence - Glossary | CSRC Recommendation of the Council on Artificial Intelligence

Algorithmic Limits and the Black-Box Problem

Modern deep neural networks function as non-linear, high-dimensional statistical representations. While their mathematical foundations are known, their internal decision pathways can involve trillions of parameters, making it difficult to pinpoint exactly why a model reached a specific output. This lack of mechanistic interpretability—known as the black-box problem—poses severe challenges in high-stakes environments such as criminal justice sentencing, credit evaluation, and medical interventions. artificial intelligence - Glossary | CSRC ISO AI Concepts and Terminology Standard

Furthermore, AI models are inherently vulnerable to systemic weaknesses:

• Data Bias and Algorithmic Drift: Models internalize and amplify historical biases present in their training datasets. If training corpuses overrepresent specific demographics or contain systemic inequities, the predictive outputs will reflect and reinforce those disparities.
• Brittleness and Adversarial Exploitation: Unlike biological cognitive systems that apply common sense when conditions shift, statistical models can fail unpredictably when presented with out-of-distribution data or adversarial inputs—minute, imperceptible perturbations in an image or text that cause a model to misclassify an object with absolute confidence.
• Computational Footprint: Modern training runs require clusters of specialized accelerators, consuming megawatts of electrical power and vast water reserves for data center cooling, tying AI progress to physical and environmental resource limits. artificial intelligence - Glossary | CSRC Recommendation of the Council on Artificial Intelligence

Regulatory and Standards Frameworks

To mitigate these risks without stalling technical innovation, international governing bodies and standards organizations have established formal risk-management frameworks. Standards such as ISO/IEC 22989 clarify terminology and establish uniform methods for assessing model robustness, functional safety, and lifecycle management. Similarly, institutions like the U.S. National Institute of Standards and Technology (NIST) and the OECD maintain frameworks centering on trustworthy AI. These frameworks require that artificial intelligence deployments demonstrate validity, explainability, privacy protection, fairness, and human oversight. artificial intelligence - Glossary | CSRC Recommendation of the Council on Artificial Intelligence ISO AI Concepts and Terminology Standard

Ultimately, artificial intelligence is neither magical consciousness nor a trivial database search. It is an advanced computational discipline that applies high-dimensional statistical optimization, linear algebra, and data-driven modeling to solve complex predictive and analytical tasks. While current systems can mimic human fluency and surface-level reasoning with remarkable precision, they remain mathematical artifacts: purpose-built engines transforming inputs into outputs through the mechanics of probability and computation. What Is Artificial Intelligence (AI)? Recommendation of the Council on Artificial Intelligence Artificial Intelligence - Stanford Encyclopedia of Philosophy

I'll check authoritative definitions before writing.## A working definition, and why the question is slippery

What exactly is AI? In the sense used by most governments, standards bodies and engineering teams today, artificial intelligence is a machine-based system that infers, from the input it receives, how to generate outputs — predictions, content, recommendations or decisions — that can influence physical or virtual environments, with different systems varying in how much autonomy they have and whether they continue to adapt after deployment. That wording comes from the OECD's definition of an "AI system," revised in November 2023 and adopted almost verbatim as Article 3(1) of the EU AI Act, which makes it the closest thing to a consensus definition currently in circulation.

That definition is useful precisely because it is boring. It says nothing about thinking, understanding, consciousness or human-likeness. It describes a behaviour: a computational system that derives its own mapping from inputs to outputs rather than having that mapping fully written out by a programmer. Regulators converged on this framing because they needed a line they could actually draw around real products.

But the question "what is AI really" usually contains two separate questions, and conflating them is the main source of confusion:

1. An engineering question — which techniques and systems does the term cover?
2. A philosophical question — is any of it genuinely intelligent?

The first has a reasonably settled answer. The second does not, and probably cannot be settled by looking at the machines alone.

Where the term came from

"Artificial intelligence" was coined in a 1955 funding proposal to the Rockefeller Foundation written by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon, which asked for a two-month, ten-person study at Dartmouth College the following summer. The proposal rested on a deliberately sweeping conjecture: that every aspect of learning or any other feature of intelligence could in principle be described precisely enough for a machine to simulate it. The workshop ran for roughly six to eight weeks in the summer of 1956 and is generally treated as the founding event of the field; McCarthy reportedly chose the label partly for its neutrality, to distinguish the work from automata theory and from Norbert Wiener's cybernetics.

Two things about that origin still shape the term. First, it was a name for an aspiration, not a description of a technology — which is why the word has always promised more than any given system delivers. Second, the field's earliest working system was symbolic, not statistical: Newell and Simon's Logic Theorist, brought to Dartmouth, proved theorems from Whitehead and Russell's Principia Mathematica by searching through logical derivations.

The four classic framings

The standard textbook treatment, from Russell and Norvig's Artificial Intelligence: A Modern Approach, organises definitions along two axes: whether you care about thought or behaviour, and whether you measure success against human performance or against an ideal standard they call rationality — a system is rational if it does the right thing given what it knows.

The textbook advocates the fourth cell, and modern practice has largely followed it. An agent is simply something that perceives and acts; AI becomes the study and construction of agents that act well given their goals and information. This framing is favoured because rationality is a clearly defined, general standard, whereas human behaviour is the product of a specific evolutionary history and is not an obvious benchmark for optimality. It also happens to be the framing most amenable to mathematics: decision theory, optimisation and reinforcement learning all formalise "acting rationally" directly.

Notice what this does to the philosophical question. The dominant engineering definition of AI deliberately sidesteps whether the system thinks, because behaviour is measurable and inner experience is not.

The moving-target problem

Part of why "what actually is AI" feels unanswerable is that the category empties itself as it succeeds. Optical character recognition, spam filtering, route planning, chess engines, speech transcription and machine translation were all frontier AI research; once reliable, they became ordinary software. The sentiment is widely quoted as McCarthy's lament that "as soon as it works, no one calls it AI anymore," though the attribution is loose — the idea appears in many forms from the 1970s onward, and the specific phrasing was attributed to McCarthy posthumously rather than traced to a publication of his. Larry Tesler's sharper version, dubbed "Tesler's Theorem" by Douglas Hofstadter, is that AI is whatever hasn't been done yet.

This "AI effect" cuts both ways. It makes the field look like it never delivers, while also making the current label a poor guide to capability: a vendor calling something AI in 2026 tells you about marketing conventions more than about mechanism.

What is actually inside the box

Strip away the branding and contemporary AI systems are built from a small number of ingredients.

Symbolic and knowledge-based approaches. Facts and rules are represented explicitly, and the system derives conclusions by search or logical inference. Expert systems, planners, constraint solvers and theorem provers live here. These dominated the field's first several decades and still underpin a great deal of production software.

Machine learning. Rather than specifying the rules, you specify an objective and let an optimisation procedure fit the parameters of a model to data. Conceptually the system learns an approximation to some mapping $f: X \to Y$ — from image to label, from sentence to translation, from state to action — by adjusting parameters to reduce error on examples. The main families are supervised learning (labelled examples), unsupervised and self-supervised learning (structure or prediction tasks derived from the data itself) and reinforcement learning (reward from interaction). European Commission guidance on the AI Act treats exactly these techniques as the typical way a system "infers how to generate outputs."

Deep learning and foundation models. Deep learning is machine learning with large multi-layer neural networks. Large language models are deep networks trained mainly to predict continuations of text, then shaped further by post-training. The OECD's own explanation of its revised definition uses this as the key example of implicit objectives: a model like ChatGPT has objectives that were never explicitly programmed but acquired partly through imitation of human-written text and partly through reinforcement learning from human feedback. A self-driving system that follows traffic rules without "knowing" the underlying goal of protecting lives is the other example they give.

The nesting matters for reading claims accurately. All deep learning is machine learning; all machine learning is AI; but not all AI is machine learning, and generative models are one branch rather than the whole field.

A second distinction worth holding onto is training versus inference. Training is the expensive, usually one-off process that produces a model; inference is running that fixed model on new input. Most deployed systems do not learn from you in real time. Adaptiveness — continuing to evolve after deployment, as with a recommender that tracks your taste or speech recognition that tunes to your voice — is a property some systems have and many do not. Systems may be trained once, periodically or continually.

Where the boundary falls — and the awkward cases

If AI is defined by inference rather than by complexity, a lot of software that feels clever is not AI, and some unglamorous statistical tooling is. The European Commission, in guidance issued to interpret Article 3(1), decomposed the definition into seven elements — machine-based; varying autonomy; possible adaptiveness; explicit or implicit objectives; inferring how to generate outputs; outputs such as predictions, content, recommendations or decisions; capable of influencing environments — and identified categories of systems falling outside it. "Machine-based" was clarified to mean integration of hardware and software, computationally driven. Critically, adaptiveness is not mandatory, and inference should be read as primarily a property of how the system was built rather than something it must do live.

Those exclusions reflect the EU's interpretive guidance rather than a universal scientific taxonomy, and they are specific to that regulation.

Other frameworks draw slightly different lines. The US NIST AI Risk Management Framework — voluntary, sector-neutral guidance published in January 2023, organised around the functions Govern, Map, Measure and Manage — uses an "engineered or machine-based system" formulation closer to the OECD's pre-2023 wording, and covers traditional machine learning, generative AI and AI features embedded in larger products alike.

Narrow systems, general ambitions, and what remains contested

Everything deployed today is best described as narrow or task-bounded, even when the task range is wide: a language model handles an enormous variety of text tasks but was still produced by optimising a specific objective over a specific corpus. "Artificial general intelligence" names a hypothesised system with broad, human-level competence across domains; there is no agreed operational test for it, no agreed timeline, and claims about its proximity should be treated as contested forecasts rather than facts.

The deeper disputes are genuinely open. Whether a system that produces fluent, apparently reasoned output understands anything, whether internal representations constitute concepts, and whether any computational system could have experience — these are live arguments in philosophy of mind and cognitive science, not questions resolved by benchmark scores. What can be said without controversy is narrower: current systems have no verified subjective experience, their competence is uneven in ways that do not mirror human competence, and strong performance on a benchmark does not license inferences about general capability.

So the honest answer to "what is AI exactly" has three layers. Operationally, it is the OECD/AI Act definition: machine-based systems that infer outputs influencing the world. Technically, it is a toolbox — search, logic, statistical learning, neural networks — unified more by shared history and shared engineering problems than by any single mechanism. Rhetorically, it is a frontier label that keeps relocating. When the word appears in a product description, the useful follow-up questions are concrete: what does the system take as input, what is the output used for, was the behaviour learned or written, does it change after deployment, and who is accountable when it is wrong.

Definitions here are also jurisdiction- and time-dependent; the OECD text was revised on 8 November 2023 and the surrounding Principles again in 2024, and national implementations continue to diverge. Anyone using these definitions to determine legal obligations rather than to understand the technology should work from the current official text and seek qualified legal review.