What AI is: from AI to ML to deep learning to generative and agentic AI

Different kinds of AI raise different risks and obligations: a spam filter, a chatbot and an autonomous agent are not governed the same way. Governing a system starts with recognizing which kind you are dealing with. This topic gives you the vocabulary to do

Domain I · Understand what AI is and why it needs governance · about 10 min

Why this matters for the exam

Different kinds of AI raise different risks and obligations: a spam filter, a chatbot and an autonomous agent are not governed the same way. Governing a system starts with recognizing which kind you are dealing with. This topic gives you the vocabulary to do that.

What you need to know

AI, machine learning, deep learning, generative AI and agentic AI are often used as if they meant the same thing. They are related but different, and this topic explains what each one means and how they fit together.

What is artificial intelligence?

Artificial intelligence (AI) is technology that performs tasks we normally associate with human thinking, such as understanding language, recognizing patterns and making decisions. ChatGPT answering a question in plain English is AI. So is the software in a car that spots a hazard and warns the driver. The AIGP course defines AI two ways: as a field of computer science "dedicated to simulating intelligent behavior in computers," and as a broad term for an engineered system that uses computational techniques to perform or automate tasks. What is being simulated is human intelligence — the ability to think creatively, consider possibilities, and keep a goal in mind while making decisions.

There is no single fixed definition of AI. What counts as "intelligent" keeps moving as technology advances, much as the meaning of "personal information" has kept expanding in privacy law. The oldest working answer comes from Alan Turing, who proposed in 1950 that a machine could be called intelligent if people could not tell its answers from a human's. This is the Turing Test. The test measures whether the machine's output is convincing. It does not measure whether the machine actually thinks. The field got its name a few years later, at the 1956 Dartmouth Summer Research Project, widely treated as the founding event of AI as an academic discipline.

Definitions of AI vary, but most are assembled from the same seven elements. They are worth a look because later material is built from this same vocabulary — the OECD's classification framework later in this domain, and the definitions used by AI laws in Domain II:

• Technology — AI is built from algorithms and data structures.
• Intelligence — it mimics aspects of human intelligence, like reasoning.
• Autonomy — it can operate and make decisions without direct human oversight.
• Goal-directed — it is designed to achieve specific objectives.
• Outputs — it produces predictions, decisions, classifications or actions.
• Learning — it adapts to new inputs.
• Human interaction — its impact depends on how people use and respond to it.

What AI is used for

AI earns its place where it is faster or more accurate than people working at scale. It can outperform human accuracy in reading medical scans, and it can review case law far more broadly and quickly than a human could. Automating repetitive processing can also remove human error and bias from decisions. The same automation can just as easily encode bias and scale it, if the underlying data or design is flawed.

The course groups common uses into seven categories:

• Recognition — image, speech and facial recognition; defect detection.
• Forecasting — demand, weather, dynamic pricing.
• Event detection — fraud detection, cyber incident detection.
• Personalization — tailored customer experiences.
• Recommendation — products, content, clinical decision support.
• Interaction support — chatbots and virtual assistants.
• Goal-driven optimization — supply chain and routing optimization.

Why AI is everywhere now

AI is accelerating because several neighboring technologies matured at the same time. The first five supply data, computing power or supporting infrastructure; the last four are fields advancing together with AI:

• Cloud computing — on-demand, scalable computing power available to everyone, driving AI development and data processing.
• Mobile technology and social media — smartphones and social platforms create the massive data volumes AI models learn from.
• Internet of Things (IoT) — connected devices generate a constant stream of data that feeds AI models.
• Privacy-enhancing technologies (PETs) — AI drives demand for better privacy tooling, and benefits from it.
• Blockchain — provides trusted infrastructure for secure transactions, and in specific contexts can strengthen data privacy and security.
• Computer vision — lets machines interpret images and video, transforming health care, autonomous vehicles and robotics.
• Autonomous vehicles — push advances in perception, decision-making and adaptation to changing environments, and drive innovation in computer vision and edge computing.
• Autonomous weapons — raise distinct concerns about accountability, bias in targeting, and international stability.
• Augmented and virtual reality (AR/VR) — AR overlays virtual objects on the real world; VR immerses users in a fully simulated one.

The four parts of every AI system

Four terms appear in almost every AI law, standard and contract, so it is worth being precise about what each one means.

They fit together in a simple sequence. An algorithm is trained on data, and the result is a model. The model is the part that makes predictions or decisions. Wrap that model in the interfaces and infrastructure needed to use it in the real world, and you have an AI system. When a law regulates an "AI system," it means this whole package, not the model alone.

AI, machine learning, deep learning and generative AI

These terms sit inside one another: machine learning is a type of AI, deep learning is a type of machine learning, and generative AI is built on deep learning.

Agentic AI has a dashed border because it describes behavior (acting autonomously) rather than a way of building a system. Most of today's agents are built on generative models, so the boxes overlap. But an agent does not have to be generative: a reinforcement-learning system that acts without creating content still counts as agentic AI.

Machine learning (ML) is AI that learns from data instead of following rules a programmer wrote by hand. A machine learning system improves with experience: the more data it processes, the better it performs, without being reprogrammed for each task. Machine learning models are commonly split into two groups by what they learn. Discriminative models learn to tell categories apart — a random forest deciding whether an email is spam or not, for example. Generative models learn what the data itself looks like, well enough to produce new examples of it.

Deep learning is machine learning built on neural networks with many layers, loosely inspired by the human brain. It is particularly good with unstructured data such as photos, audio and free text, and it finds useful patterns on its own. Older methods needed a person to define the patterns. Networks learn through backpropagation with gradient descent: with each training pass, the network's internal weights are adjusted slightly to reduce prediction error, millions of times over. This takes large amounts of good data and serious computing power. It is also why deep learning is trained on GPUs rather than ordinary processors: GPUs run thousands of calculations in parallel, which is exactly what neural networks need.

Generative AI is deep learning used to create new content: text, images, video, audio or code. Having learned what cats look like, a generative model can produce a cat picture that never existed before — one that resembles its training data without copying it. ChatGPT (OpenAI), Gemini (Google), Copilot (Microsoft), Firefly (Adobe) and Claude (Anthropic) are all generative AI.

Agentic AI describes systems that pursue goals on their own with limited human oversight. Given an objective, an agent breaks it into sub-tasks and executes them, often improving through reinforcement learning. These systems are also called "AI agents." Because more autonomy means fewer moments where a human checks the work, agentic AI raises risks of its own, and Domain IV treats its deployment as a separate governance question.

Putting the vocabulary to work

Here is the vocabulary applied to an example. A bank's system reads loan applications and predicts which borrowers will default, improving as repayment histories accumulate. Is it AI? Yes. Machine learning? Yes: it improves with experience. Deep learning? The description does not say — that is an implementation detail. Generative AI? No: it predicts, it does not create content. Agentic? No: it scores applications; it does not act on its own. The same set of questions works on any system description.

Next up: how machines actually learn — supervised, unsupervised, reinforcement and semi-supervised training.

Practise this topic

Domain I is free in the app, including its practice questions and flashcards, with progress tracking and no card details.

Start studying free

Next: How machines learn: supervised, unsupervised, reinforcement and semi-supervised training
Back to the AIGP study guide.

AI Governance Study is an independent study aid. It does not represent a government entity: it is not affiliated with, endorsed by or authorised by any government, government agency or regulatory authority, and it does not provide government services or legal advice. Laws and frameworks are described in our own words — the official texts are listed at official sources. It is also not affiliated with, endorsed by, or sponsored by the IAPP. The AIGP name is used only to identify the exam this material helps you prepare for.