A landscape, not a ladder

A familiar diagram places artificial intelligence in a large circle, machine learning inside it, deep learning inside machine learning, and generative AI at the center. The diagram contains a useful truth: deep learning is a family of machine-learning approaches, and many prominent generative systems use deep neural networks. But perfect nesting creates several misconceptions.

Generative models existed before the current deep-learning era and are not limited to neural networks. Deep learning is used for non-generative tasks such as classification and detection. AI also includes approaches that do not learn from data, including search, planning, optimization, and explicit rules. A business application may combine all of them.

Audiobook description

A broad shape labeled AI contains two large overlapping regions: explicit knowledge and search on the left, machine learning on the right. Deep learning sits inside machine learning. A band labeled generative approaches crosses part of the left and right regions. Several application icons sit across boundaries, showing that products combine methods.

The word AI is best treated as an umbrella for engineered systems that perform functions associated with perception, language, learning, reasoning, planning, recommendation, decision, content generation, or action. The umbrella is historically and institutionally shaped; its boundary changes. For practical work, the label should open a second question: which capabilities and methods are present here?

Two broad traditions: express the knowledge or learn the pattern

One long AI tradition asks people to express knowledge in a form a machine can manipulate. A rule might say that a claim above a certain value requires a second review. A knowledge representation might encode that Toronto is in Ontario and Ontario is in Canada, allowing a system to derive that Toronto is in Canada. A planner might search possible action sequences until it finds one that reaches a goal.

This family is often called symbolic AI because entities and relationships are represented with explicit symbols. Its strengths include traceable logic, direct use of expert knowledge, and predictable enforcement of clear constraints. Its weaknesses appear when rules become numerous, exceptions multiply, or relevant knowledge is difficult to articulate.

The other broad tradition asks a system to estimate useful patterns from data. Instead of writing every signal of fraud, developers give a learning algorithm examples and a performance objective. Training adjusts a model so that it performs better on the defined task. Tom Mitchell's classic formulation centers the relationship among task, experience, and performance.¹

Audiobook description

Two parallel lanes produce a claim-routing result. The symbolic lane begins with expert rules and a search or inference procedure. The learning lane begins with past examples and a performance measure, then trains a model. The lanes join before the final output, illustrating a hybrid system.

The contrast is not absolute. A learning system contains human choices about data, objectives, architecture, and thresholds. A symbolic system can use probabilities or learn some components. A modern language application may use a neural model to interpret a request, a search system to retrieve documents, explicit rules to enforce permissions, and a conventional program to execute an action.

Machine learning is organized by the signal available

Machine learning is not one algorithm. It is a large family of ways to improve performance through experience. A helpful first distinction asks what feedback the learner receives.

In supervised learning, examples are paired with target answers. Emails may be labeled spam or not spam. Transactions may be labeled fraudulent or legitimate. The learner estimates a relationship that can be applied to new cases.

In unsupervised learning, the data do not arrive with the desired answer for every example. The system may group similar customers, compress information, or identify unusual patterns. In self-supervised learning, the data provide their own training targets: hide part of an image and predict it, or hide a word and infer it from context. Much of modern foundation-model training uses objectives of this general kind.

In reinforcement learning, an agent takes actions, receives rewards or penalties, and learns a policy for choosing actions over time. The challenge is delayed consequence: an action that looks good now may lead to a poor result later.

Audiobook description

Three lanes begin with a learner. In the first, a teacher supplies examples with correct answers. In the second, parts of the data hide and reveal one another, allowing the learner to discover structure. In the third, an agent acts in an environment and receives rewards over time. Each lane labels the feedback signal available.

These categories can combine. A language model may be pretrained with a self-supervised prediction objective, then tuned on instruction-response examples, then further adapted using preference-based rewards. The resulting assistant did not learn in only one way.

Neural networks and the meaning of “deep”

An artificial neural network is a computational model built from connected units arranged in layers. Inputs are transformed through weighted connections and nonlinear operations. During training, the weights are adjusted so the network produces more useful outputs.

The biological name is historical inspiration, not a claim that the model reproduces a brain. Artificial units omit nearly all the complexity of living neurons, bodies, development, and social experience. Calling a model neural tells us about a family of computational structures, not about consciousness.

Deep learning refers to neural networks with multiple processing layers that can learn representations at several levels of abstraction. A vision system might move from pixel patterns to edges, shapes, object parts, and categories. A language system develops distributed representations useful for relationships among tokens and contexts. The 2015 review by Yann LeCun, Yoshua Bengio, and Geoffrey Hinton emphasized this multilayer representation learning.²

Audiobook description

Four nested boxes read AI, machine learning, neural networks, and deep learning. From the deep-learning box, three arrows travel outward to classification and prediction, generation, and reinforcement-learning agents. The nesting answers what it is; the arrows answer what it can be used to do.

Depth made it practical to learn complex representations from large amounts of raw or lightly processed data, but it is not automatically the best choice. For small structured datasets, a simpler model can be cheaper, easier to maintain, and equally accurate. Explicit rules remain appropriate for non-negotiable policy constraints. Method should follow the problem, evidence, and operating environment.

Predicting a label and generating a new sample

Another useful axis distinguishes models by what they learn and produce. A discriminative model learns a boundary or relationship useful for predicting a target from an input: Is this transaction likely fraudulent? Which product category fits this image? What demand should we expect next week?

A generative model learns enough about the structure or distribution of data to produce new samples or continuations with related characteristics. It may generate text, images, audio, video, molecular structures, or synthetic records. NIST's glossary describes generative AI as a class of models that emulates characteristics of input data to produce derived synthetic content.³

Audiobook description

The left lane sends an existing document through a model to one of three category boxes. The right lane sends a prompt through a model to a newly created document. Under the left is “which label fits?” Under the right is “what sample could plausibly follow?”

The boundary can blur. A language model generates by predicting a probability distribution over the next token. A generative model can support classification by comparing likely completions. A product may use generation only as an interface around a predictive system.

Evaluation differs sharply. A fraud classification can often be compared with later confirmed outcomes, though labels may be delayed or imperfect. A generated strategy memo has many possible acceptable answers. It requires evaluation of factual support, relevance, novelty, tone, policy compliance, and downstream use. The freedom that makes generation valuable also makes simple right-or-wrong scoring less adequate.

The product is usually a combination

Taxonomies are useful for learning and dangerous for procurement. A vendor may say “our solution uses generative AI” as though that phrase specifies the system. It does not tell you which model, what data flows through it, whether information is retrieved, which rules constrain output, what software executes actions, or who reviews consequential results.

Consider an assistant that handles an employee's request to book travel. A language model interprets the request. A retrieval system obtains policy. Explicit rules reject prohibited options. A search service queries availability. An optimizer balances time and cost. Conventional code calls booking systems. A human approves an exception. Logging and monitoring support oversight. The assistant's conversational surface is only the visible layer.

Audiobook description

A vertical stack begins with the employee conversation. Below are a generative model, retrieval and search, policy rules and optimization, booking software tools, and enterprise data. Human approval and governance appear as rails running alongside every layer, not as a final decoration.

The most useful inventory therefore records capabilities and components rather than chasing labels. For each system: identify the business purpose, model families, learning methods, data sources, external tools, autonomy, affected people, human decision points, and monitoring. This description remains useful even when a fashionable term changes.

Remember this

Audiobook description

Four rows form a worksheet. Goal asks what capability matters. Method asks whether rules, search, optimization, or learning produces behavior. Model family names the computational structure. System lists data, tools, people, and controls. The completed rows flow into one precise system description.

Five key terms

• Symbolic AI
• Machine learning
• Deep learning
• Generative model
• Hybrid system

Reflection questions

1. Which AI categories are used as synonyms in your organization, and what decisions become unclear as a result?
2. Choose one AI product. Which parts are rules, learned models, search, optimization, conventional software, and human judgment?
3. Where would a simpler method be more appropriate than a deep or generative model?