What is AI and what
are the different types?

The current AI boom has created a landscape filled with confusion and misconceptions about artificial intelligence. Even those working closely with technology often find it hard to keep pace with new discoveries and emerging trends.

AI has already embedded itself into many of the tools we use, sometimes in obvious ways, sometimes running quietly in the background. Most of us use it every day without really thinking about it. But how many of us can explain what AI actually is or how it works?

In this blog, we’ll show you how to distinguish between different types of AI and select the perfect model for any task.

What is AI?

At its most fundamental level, AI is exactly what it sounds like—technology built with the intention of mimicking human intelligence to some degree. This fascination with creating artificial intelligence isn’t new; from HAL 9000 and Frankenstein to the intricate automata of ancient civilisations, humanity has long found fascination in bringing intelligent life to inanimate objects.

The formal research into AI began in earnest during the 1950s with pioneers like Alan Turing laying the theoretical groundwork for machine intelligence. Since then, we’ve witnessed remarkable breakthroughs that have captured public imagination, such as Deep Blue’s chess victory over world champion Garry Kasparov.

Today, AI is rapidly changing how businesses approach problem solving and workflow automation, while allowing for never-before-possible solutions to specialised tasks. However, AI is an incredibly broad and ever-expanding group of technologies and in order to figure out which type of AI best suits your needs, first it’s important to understand the different types of and what makes each of them unique. In the next section, we’ll outline different AI architectures and explain when they’re the right choice for your task.

Different types of AI

Infographic
A high-level view of how different types of AI overlap and interact.

You may have heard the terms ‘AI’, ‘Machine Learning (ML)’, ‘Natural Language Processing (NLP)’, and ‘Generative (Gen) AI’ thrown about interchangeably. However, these different ‘families’ of AI models have key distinctions, overlaps, strengths, and weaknesses that are important to keep in mind when selecting which model to use.

Machine Learning (ML)

At its core, machine learning (ML) encompasses AI systems that learn from data rather than rely on pre-programmed instructions.

Consider the difference between a traditional toaster and a (hypothetical) ML-powered toaster. Your everyday toaster follows rigid programming—press the lever, wait for the timer, receive your toast. It’s reliable but inflexible. Now imagine a toaster equipped with “too burnt” and “still-just-bread” buttons. Initially, it produces random results, but with each piece of feedback, it adjusts its timing. After a week of this iterative process, it has learned your perfect toasting preference without anyone explicitly programming that knowledge into its system.

This iterative learning process forms the backbone of all machine learning models. Whether through simple human feedback, like the thumbs up and down buttons you see after ChatGPT responses, or sophisticated loss functions that measure output accuracy, these systems are capable of continuously refining their performance.

ML is present in a wide range of tools, from those that help rank or organise information, to systems that spot trends and patterns in large datasets. It has helped automate tasks that once depended entirely on human intuition and expertise.

Neural Networks and Deep Learning

In humanity’s pursuit of artificial intelligence that rivals or exceeds natural cognitive abilities, researchers turned to algorithms that mirror the shape of the human brain. The perceptron, conceptualised in 1943 and brought to life in 1957, is a simple mathematical function that acts as a single ‘neuron’. When these ‘neurons’ are combined in great quantities, they are able to learn increasingly complex representations of data. The logic is simple: more perceptrons (or, more ‘brain cells’) should entail more sophisticated data interpretation.

For a long time, limited computing power made it hard to train neural networks—the parameters of each perceptron need to be updated with each training iteration. (For example, a ‘small’ neural network with 5 layers of 100 perceptrons per layer, would be composed of over 40,000 parameters!) The rise of deep learning was made possible not just by improved algorithms, but by dramatic advances in computing hardware—especially GPUs—which enabled the training of large, layered neural networks on massive datasets, driving the AI breakthroughs we see today.

Neural networks are the building blocks upon which the other types of architectures in this blog are built, and they can be found in a wide range of tools, from image generators to large language models.

Natural Language Processing (NLP)

Natural language processing (NLP) is focused on understanding, analysing, and generating human language in all its complexity. From grammar and sentence structure to semantic meaning and speech recognition, NLP models tackle the full spectrum of linguistic challenges. These systems span an impressive range, from simple decision trees handling basic syntax to sophisticated models like GPT-4 with its estimated 1.8 trillion parameters.

Given language’s central role in human communication, NLP has become the category of AI technology that most people encounter daily. Whether you’re asking Siri a question, dictating notes on your phone, relying on auto-complete suggestions, or seeing those familiar blue-underline grammar corrections, you’re experiencing NLP at work.

Large Language Models (LLMs)

The type of NLP model you’re probably most familiar with is the Large Language Model (LLM). LLMs stand apart from traditional NLP approaches through their massive scale, employing billions (or trillions!) of parameters and vast training datasets to capture the intricate complexities of human communication. This ambitious scope stems from a fundamental design philosophy rooted in neural network architecture—if AI systems are modelled on the human brain and aim to learn human-like language capabilities, then matching the brain’s immense scale becomes essential. With approximately 86 billion neurons and 100 trillion synapses powering human cognition, even today’s most advanced LLMs are still working toward approaching this biological benchmark.

LLM-powered chatbots such as Claude and chatGPT are already proving indispensable across industries. Their practical impact is reshaping modern workflows, streamlining everything from content creation to complex problem-solving.

Generative AI (GenAI)

Generative AI, a subset of machine learning, is characterised by its capacity to produce novel outputs (such as text, images, audio, and video) by leveraging existing patterns found within its training data. Often called by its shorthand name, ‘GenAI’, this subcategory of AI can be tricky to define.

In short, GenAI encompasses models that generate ‘new’ outputs given an input. For example, you can ask (prompt) ChatGPT to “Write a poem about a duck”, and it will ‘generate’ new text (based purely on your prompt) – this is GenAI.

Conversely, using your face to unlock your phone is not GenAI, as it isn’t generating a tangible output – instead it is classifying your facial features and ensuring that they match the facial features stored when you initially set up Face ID.

Another way of looking at things… Encoders vs Decoders

Encoders

In many machine learning models, encoders serve as a crucial first step in the pipeline, converting raw data into standardised mathematical representations that models can process effectively. For instance, in a simple model an encoding of ‘duck’ could look like the following:

 

Table
An example of how an animal’s characteristics can be translated into numerical representations for machine learning models.

As you can see, by defining animals by various ‘features’, we’re able to create a mathematical representation of “duck” that the model can process. This standardisation process creates consistency within AI systems. Encoders ensure that all data entering the model follows the same format and structure, allowing the system to detect patterns, identify changes, and recognize meaningful differences across datasets.

Encoder models excel at taking varied and noisy input data, such as random free-text or images, interpreting it and transforming it into its essential information and features – i.e. they encode the data. A common type of encoder model is a classification model. Classification models take in random input data, and categorise it into predefined classes. So for example a simple, pre-trained classification model could be fed a series of free-text descriptions of different birds, and successfully predict the exact type of bird for each from a predefined list (such as pheasant; duck; pigeon; crow; etc.).

Decoders

In reverse, decoders perform the equally critical task of converting abstract numerical encodings back into human-readable information – i.e. they decode the data. The encoder-decoder architecture provides the computational flexibility that makes modern AI systems so versatile. Different models can incorporate encoders, decoders, both, or neither, depending on their specific applications and requirements.

Decoder-only models allow for a streamlined approach to language processing that has proven remarkably effective in GenAI. Unlike traditional encoder-decoder architectures, these models—including the GPT family—eliminate the encoder component entirely, processing inputs directly through their decoder layers. This architectural choice makes them well-suited for generative tasks, excelling at understanding context and creating new, coherent outputs – i.e. decoding a prompt and generating a comprehensive response.

By focusing solely on the generation process, decoder-only models can dedicate their full computational capacity to creating human-like text, which explains why they’ve become the foundation for most conversational AI systems and other GenAI tools.

Agents

So far, we’ve looked at models that generate or analyse outputs. Agentic AI systems distinguish themselves through their capacity to autonomously make decisions and execute actions using accessible tools and knowledge bases. While a standard LLM cannot answer real-time queries like “What is the NASDAQ currently valued at?” due to its reliance on historical training data, an agent processes such requests by leveraging available tools—performing web searches or accessing live market indices directly. These interactions build the agent’s knowledge base, enabling it to handle comparative queries such as year-over-year market analysis. The systems can operate continuously, responding to real-time triggers like automated trading instructions based on predetermined price thresholds.

Similarly, self-driving vehicles exemplify agentic AI through their control over multiple driving mechanisms, processing inputs from speed sensors, cameras, and GPS systems to make rapid decisions as conditions change. This dynamic responsiveness makes agentic AI systems particularly valuable for applications demanding complex analysis and immediate action execution in changing environments.

Why this matters

The AI landscape may seem overwhelming, but understanding these fundamental distinctions—from different types of AI to the specific capabilities of encoders, decoders, and agents—provides a practical framework for navigating your options. Each AI architecture serves distinct purposes, whether you need generative capabilities, analytical processing, or autonomous decision-making systems. You are now equipped with the knowledge you need to confidently understand what colleagues and industry professionals mean when terms like “agent” “GenAI,” or “LLM” enter the conversation. By matching your specific requirements to the appropriate AI type, you can cut through the confusion and select tools that genuinely enhance your workflow rather than add unnecessary complexity.

Curious about AI?

We’re creating a space to share insights, practical lessons from using it, and unpack the questions you’re already asking.

Sign up to get the latest straight to your inbox.