The Major Types of AI

On 1 January 2023 a new section of the California Vehicle Code came into force. It bars carmakers and dealers from naming or marketing a driver-assistance feature in language that would lead a reasonable person to believe the car can drive itself, and it requires them to explain what the feature does and what it cannot do¹.

The interesting part is how the law defines the feature it regulates. It does not mention cameras, radar, neural networks or any other technology. It borrows a definition from an engineering standard, SAE J3016, which sorts vehicles into six levels, from Level 0 (no driving automation) to Level 5 (full driving automation), according to one question: how the driving task is divided between the person and the machine². The law covers Level 2, where the system steers and brakes but the driver must supervise at all times.

The car industry had a vocabulary problem: “autopilot”, “self-driving” and “driver assist” meant whatever the speaker wanted them to mean. It solved the problem by separating one question from all the others. Business AI has the same vocabulary problem, with more words and less discipline. This chapter offers the same remedy.

Four questions hiding in one

Ask a room of managers how many types of AI there are and you will hear narrow AI, general AI, machine learning, deep learning, generative AI, computer vision, foundation models and agents. Every one of those terms is legitimate. They are not rivals on one list, because they answer different questions.

Capability asks what the system produces or does: perceive, understand, generate, act. Approach asks how it is built: written rules, machine learning, deep learning, a foundation model. Generality asks how broad its competence is, from one narrow task to many unrelated ones. Autonomy asks how much it does without a person: suggest, decide or act.

Serious classifications work this way. When the OECD built a framework to help regulators characterize AI systems, it did not publish a list of types. It described each system along five dimensions at once, including its data and input, its model and the task and output it performs, alongside the people affected and the economic context³. A single label cannot carry that much information, and nobody should expect it to.

This is why arguments about “what type of AI this really is” go nowhere. When a data scientist calls a system machine learning and a business sponsor calls it generative AI, both are right, because they are describing different properties. The useful question is not which label wins. It is which property the decision in front of you depends on.

What it can do: generative AI names an output

The capability lens is the one leaders use most, and What Exactly Is Artificial Intelligence? set out its family of six capabilities, from perceiving and understanding to generating and acting. One point belongs here, because it causes most of the confusion: generative AI is a capability label. It says the system produces new content, such as text, images, audio or code. It says nothing about how the system was built, what data it was given, how reliable it is or what it is connected to.

That gap matters in practice. A team that calls its project “generative AI” invites a review of the output: tone, accuracy, brand voice. Those checks are necessary. But the same project may also read customer records, write to a billing system or send messages, and the label gives a review board no reason to ask. As Generative AI Changes the Game explained, generative systems added a new kind of output to the enterprise. They did not replace the systems that predict, rank and detect, which still carry much of the value.

How it is built: the approach decides how it fails

The approach lens names the method. AI vs Machine Learning placed the main ones: written rules, machine learning, deep learning inside it, and foundation models trained broadly and adapted to many tasks⁴. Most of today’s generative AI runs on deep learning, usually on a foundation model, which is why the two lenses are so often confused. They remain different questions. A deep learning model can generate nothing at all and simply score credit risk; a generative system can be wrapped in written business rules.

For a leader, the approach lens answers a practical question: how will this system fail, and who will fix it? AI vs Automation showed that a written rule fails the same way every time and is fixed by changing the rule, while a learned model fails statistically and is fixed by changing its data or training. A foundation model adds a third pattern: much of what it knows came from someone else’s training, so some behavior can only be changed by the supplier.

Real systems mix approaches, and the best often do so on purpose. A learned model reads an unstructured document; a written rule enforces the policy limit; a person handles the exception. Describing such a system as “machine learning” is not wrong. It is incomplete in exactly the place where accountability sits.

How general it is: narrow by design, general by claim

The generality lens asks how broad a system’s competence is. At one end is narrow AI, very good within a defined task or domain. At the other is artificial general intelligence, or AGI, broadly competent across unrelated tasks.

Narrow does not mean weak. A 2024 paper by researchers at Google DeepMind, which reviewed nine prominent definitions of AGI, argues that performance and breadth must be measured separately. By that yardstick narrow systems such as AlphaFold, which predicts protein structures and earned its creators a share of the 2024 Nobel Prize in Chemistry, already reach superhuman performance, while the general chatbots of 2023 sat at the lowest level [@morris-2024-levels-of-agi; @nobel-chemistry-2024]. Foundation models have widened the range of a single system, because one model can be adapted to many tasks⁴. That breadth is real. It is still tested, and trusted, one task at a time.

One caution covers the far end of the line. There is no agreed definition of AGI and no agreed test for it, so the term in a sales pitch or a press release is a claim, not a specification. When you hear it, ask what was demonstrated, on which tasks and under what conditions. Superintelligence, AI far beyond people in most domains, is a hypothetical idea; no enterprise decision today depends on it. For the systems you can buy, the generality question is simpler and more useful: is the system scoped to the job, and has it been tested on the job, with your data, in your conditions?

How much it does on its own: autonomy is a setting

The autonomy lens asks how much the system does without a person. It is often left out of a proposal, and it often decides the risk. It is also the lens leaders most control, because autonomy is a design choice rather than a property of the model.

Google’s data center cooling shows the difference. In 2016 DeepMind trained deep neural networks on historical sensor data and used them to recommend settings for cooling equipment. Operators vetted the recommendations and applied them, and DeepMind reported a 40 percent cut in the energy used for cooling⁵. Two years later, the system was allowed to control the cooling directly, still under operator supervision. DeepMind reported average energy savings of around 30 percent. That is an average over months of live control, while the 40 percent came from trial runs of recommendations, so the two are not directly comparable. The move came with eight safety mechanisms, including checks that discard low-confidence actions, a second layer of verification in the local control system, and the ability of operators to switch the AI off at any moment and fall back to the existing automation rules⁶.

Capability, approach and generality hardly changed between the two stages. Autonomy did, and with it the engineering and the controls. The DeepMind researchers behind the AGI table make the same point in general form: the right level of autonomy need not be the highest that the model’s capability allows, and the choice should follow from the risk⁷. Agentic AI and Autonomous Actions, in Module 06, sets out the levels and how to choose among them. For now, the lesson is that a proposal that does not state its autonomy has left out its most consequential sentence.

Let the decision choose the lens

Technology teams need precise technical categories. Executives need the lens that serves the decision in front of them, and different decisions lean on different lenses.

Regulators have made the same choice. The EU AI Act does not sort systems by technique. Its high-risk list, Annex III, is organized by area of use, and one of its entries covers AI used as a safety component in managing and operating road traffic or the supply of water, gas, heating or electricity⁸. The Act also uses a generality lens of its own: providers of general-purpose AI models carry separate obligations, which have applied since August 2025⁹. A two-word label tells counsel neither what the system is used for nor whether a general-purpose model sits inside it.

Story: a name that answered the wrong question

The vocabulary problem in this chapter’s opening has a documented case. Tesla calls the combined use of two of its driver-assistance features, Traffic-Aware Cruise Control and Autosteer, “Autopilot”. Tesla itself characterizes the system as SAE Level 2: it provides steering, acceleration and braking within a specified driving environment, under the direct supervision of the driver¹⁰. Autosteer, the company told regulators, “is designed and intended for use on controlled-access highways”¹¹. Through the four lenses, then, the system keeps a car in its lane and at a distance from the car ahead; it is narrow by design; and on autonomy it is an assistant that a person must supervise at every moment. The name answers none of those questions accurately.

The US National Highway Traffic Safety Administration (NHTSA) examined the gap between the name and the system in an investigation that it upgraded to an engineering analysis in June 2022 and closed in April 2024. It analyzed 956 crashes reported up to August 2023. In 467 it found recurring patterns, including 211 crashes in which the front of the Tesla struck a vehicle or obstacle that an attentive driver would have had time to respond to; those crashes included 13 that were fatal, with 14 deaths. It concluded that a weak driver-engagement system was not appropriate for Autopilot’s “permissive operating capabilities”, and that the mismatch produced “a critical safety gap between drivers’ expectations of the L2 system’s operating capabilities and the system’s true capabilities”¹⁰. It also addressed the label directly. The term “Autopilot”, the agency wrote, “does not imply an L2 assistance feature, but rather elicits the idea of drivers not being in control”, while peers generally used words such as “assist”, “sense” or “team”¹².

The fixes came through the lenses the name had hidden. In December 2023, while not concurring with the agency’s analysis, Tesla recalled 2,031,220 vehicles and began sending a software update that added alerts and controls, including extra checks when Autosteer is used outside controlled-access highways and eventual suspension for drivers who repeatedly fail to stay engaged¹¹. Those are changes to the autonomy boundary and to the operating domain. The label followed. In December 2025 California’s Department of Motor Vehicles adopted a judge’s decision that Tesla’s use of “Autopilot” was misleading and violated state law; by February 2026 Tesla had stopped using the term in its California marketing and so avoided a suspension of its licenses¹³. Tesla has since asked a court to overturn the finding, as reported¹⁴.

Business AI rarely carries a name as loaded as Autopilot, but the mechanism is the same. A one-word or two-word label invites the people who use a system, and the people who review it, to fill in the lenses it leaves out, and they tend to fill in the autonomy lens generously. Four plain sentences, one per lens, leave less to the imagination.

What this means for leaders

The four lenses are not a technical taxonomy to memorize. They are a habit of description. A proposal described through all four is one a committee can judge; a proposal described with one fashionable label hides at least three answers, and the hidden answer is usually the one about autonomy.

Three practical consequences follow. First, ask for the description before the debate: four sentences, one per lens, written by the team that owns the system. Second, when two people argue about what type of AI something is, ask each which question they are answering; the argument usually ends there. Third, treat generality claims as claims. A system that is “general” in a demonstration is narrow in your business until it has been tested on your tasks.

Check yourself

1. There is one official list of the types of AI.MythFact
2. Generative AI describes what a system produces, not how it was built.MythFact
3. Narrow AI means weak AI.MythFact
4. One system can honestly be deep learning, rule-bound and narrow at the same time.MythFact
5. How much a system does on its own is fixed by the model it uses.MythFact
6. The EU AI Act classifies systems mainly by the technique they use.MythFact

Reflection: describe your own system in four sentences

What comes next

Across the four lenses, one approach sits beneath most of the systems in this chapter: the cooling controller, AlphaFold and nearly every generative tool learned their behavior from data rather than from written rules alone. The next chapter, How Machine Learning Learns, opens that box and shows how a machine actually learns from examples, and what can go wrong when it does.