The phrase artificial intelligence does enormous work in contemporary discourse. It covers chess engines, recommendation algorithms, self-driving cars, image generators, large language models, and a half-dozen capabilities that have not been built yet. Without disambiguation, many disagreements about “AI” turn out to be disagreements about which of these the parties have in mind.

The encyclopedia’s working vocabulary, after Russell and Norvig’s standard textbook treatment.1

Symbolic AI

The original conception, dominant from the 1950s through roughly the 1990s. Symbolic AI builds intelligent behavior from explicit rules: a knowledge base of facts and a system of inference rules that operate on them. Expert systems in medicine, deductive theorem provers, classical chess engines — these are symbolic AI. Their strengths: transparency (the rules are inspectable), correctness (the inference is verifiable), and stability (changes are predictable). Their weakness: brittleness. They handle exactly what they were designed for, and fail abruptly outside it.

Symbolic AI is not dead, but it is not where the energy is in 2025. Most production AI systems use it only in specific roles (constraint solvers, formal verification) within larger machine-learning pipelines.

Machine learning

The dominant paradigm since roughly the 1990s. Machine learning systems do not have explicit rules. They have a model architecture (a parameterized function) and a training procedure that adjusts the parameters based on data, to minimize a loss function. The trained system makes predictions or generates outputs by applying the learned function to new inputs.

Machine learning is broad. It covers linear regression, decision trees, support vector machines, random forests, and the many kinds of neural networks. The unifying property is learning from examples rather than from hand-coded rules. The strengths: it scales with data, it captures patterns too complex for hand-coding, and it generalizes from experience. The weaknesses: it requires data, it inherits the data’s biases, and it is typically opaque (you cannot inspect the learned model and report what it has learned).

Deep learning

A subset of machine learning using neural networks with many layers (“deep” networks). The architectural ideas are old; what changed in the 2010s was the combination of much more data, much more compute, and a few key algorithmic advances (ReLU activations, dropout, attention mechanisms, transformers) that made deep networks trainable at scale.

Deep learning powers most contemporary high-profile AI: image classification, speech recognition, machine translation, the foundation models that produce ChatGPT, Claude, Gemini, and so on. When recent commentary refers to “AI progress,” it usually means deep-learning progress.

Generative AI

A subcategory within deep learning. Generative AI systems produce novel outputs of a given type — text, images, audio, video, code — rather than merely classifying or predicting. The recent rise of generative AI is mostly a story of large language models (LLMs) and diffusion models, two architectural families that have, in different ways, learned to produce high-quality novel content.

Generative AI is what most lay discussion of “AI” in 2025 actually refers to. ChatGPT, Claude, Gemini, Midjourney, DALL-E, Stable Diffusion, GitHub Copilot, the deepfake tools — all generative. The encyclopedia’s interest is disproportionately in this category, because the generative capabilities are what produce the cognitive-effect literature this work is built on.

Large language model

A specific kind of generative AI: a deep neural network trained on a large corpus of text to predict the next token given preceding tokens. The training objective is simple — token prediction — and the resulting capabilities are surprisingly general: question answering, summarization, code generation, translation, creative writing, and a long tail of tasks the training did not explicitly target.

LLMs are the central technical artifact of the period this encyclopedia covers. Their strengths: fluent generation across domains, broad world knowledge encoded in weights, capacity for in-context adaptation. Their weaknesses: hallucination, lack of grounded reasoning, opacity, and the training-corpus biases discussed throughout Section C.

What “AI” does not mean here

A few clarifications worth making.

  • AI is not the same as automation. Automation existed long before AI; most automation is not AI. The relevant question is whether the system learns (machine learning) or generates (generative AI). A washing machine is not AI, however convenient it is.
  • AI is not the same as artificial general intelligence (AGI). AGI is the hypothetical capacity for cross-domain general reasoning at human level or above. Current AI systems do not have this. Whether they ever will, and on what timescale, is contested.
  • AI is not the same as the threat model. The popular phrase “AI” in alarmed contexts often refers to AGI-shaped fears that are not well- matched to current AI systems. The encyclopedia tries to keep these separate.

A small note on the encyclopedia’s usage

When the encyclopedia uses “AI” without qualifier, it usually means contemporary deep-learning systems, especially generative ones, and especially large language models. This is the engine of the cognitive effects the work analyzes. Where the discussion needs to be more precise — distinguishing recommendation systems from LLMs, say, or generative-image models from text models — the article specifies. Where it does not specify, the default is the general-purpose generative AI of the 2024–2025 period.

The next article (A.03) gives a parallel working vocabulary for human cognition; A.04 takes up what specifically changed at the 2024–2025 moment; A.05 introduces the augmentation vs. decline tension that organizes the rest of the encyclopedia.

Footnotes

  1. Russell & Norvig, 2020. The standard reference. ↩