Saying “artificial intelligence” without specifying what type is like saying “vehicle” without distinguishing between a scooter and a plane. The label is useful for general conversation, but it obscures distinctions that matter when you want to understand what a system can do and what it cannot.

AI is not a single technique. It is a family of approaches with radically different mechanisms, capabilities and limitations. The map below is not exhaustive — the field is enormous — but it covers the types that appear in 95% of conversations about applied AI.

Why the distinction matters

When you know what type of AI a system uses, you understand what it needs to function (labelled data? an environment to interact with?), what it is inherently good at, and where it is likely to fail. You can evaluate claims about AI products more critically and make better decisions about when to use it and when not to.

Supervised learning

What it is: The model learns from labelled examples. It is shown an input and the correct answer, and learns to produce that answer for similar inputs.

How it works: Imagine a spam classifier. You show it thousands of emails already classified as spam or not spam. The model learns which features of the email correlate with each category. When a new email arrives, it applies what it has learned.

Real examples:

  • Spam detection
  • Medical image diagnosis
  • Speech recognition
  • Sentiment classification in reviews
  • House price prediction

Key limitation: It needs labelled data, and labelling data at scale requires time and money. If you do not have examples of what you want the model to learn, you cannot use this approach directly.

Unsupervised learning

What it is: The model learns without labels. It is not told what the correct answer is — it is asked to find structure in the data by itself.

How it works: Imagine thousands of news articles without categories. Unsupervised learning can group them into similar topics (clustering) without anyone telling it what categories exist. The model discovers the latent structure of the data.

Real examples:

  • Customer segmentation by behaviour
  • Anomaly detection (fraudulent transactions that “fall outside” the normal pattern)
  • Data compression
  • Topic discovery in large text collections
  • Recommendation systems

Key limitation: The groupings the model produces may not match the categories that are useful for your business. Interpreting the results requires human judgement.

Reinforcement learning

What it is: The model learns through interaction with an environment, receiving rewards or penalties for its actions. There is no static dataset — the model generates its own learning process by playing.

How it works: Think of an agent learning to play chess. You do not tell it what move to make in each position. You tell it that winning is worth +1 and losing is worth -1. The agent plays millions of games against itself, adjusts its strategy based on outcomes, and over time learns to play better than any human.

Real examples:

  • AlphaGo and AlphaZero (board games)
  • Physical robot control
  • Optimising cooling systems in data centres
  • Automated trading strategies
  • RLHF (Reinforcement Learning from Human Feedback): the method used to align language models like ChatGPT or Claude

Key limitation: It requires precisely defining what gets rewarded, and if that definition is slightly wrong, the agent can find unexpected ways to maximise the reward that are not the desired ones. This is the “reward specification” problem.

Generative AI

What it is: Models trained to produce new content — text, image, audio, video, code — that is coherent with the patterns learned during training.

How it works: Unlike supervised models that classify or predict a value, generative models learn the statistical distribution of their training data and can sample from that distribution to produce new examples. A language model learns how human text is structured; an image model learns what images look like. Then it can generate text or images that resemble those from training.

Real examples:

  • GPT-4, Claude, Gemini (text)
  • DALL·E, Midjourney, Stable Diffusion (image)
  • Sora, Runway (video)
  • Whisper + ElevenLabs (audio and voice)
  • GitHub Copilot (code)

Key limitation: Generative models have no access to “truth” — they generate what is statistically plausible, not what is factually correct. Hallucinations are a structural consequence of this mechanism, not a bug.

The full map

ARTIFICIAL INTELLIGENCE
│
├── Machine Learning
│   ├── Supervised → classifies, predicts with labels
│   ├── Unsupervised → clusters, finds structure without labels
│   └── Reinforcement → learns by acting, maximises reward
│
└── Generative AI
    ├── Language Models (LLM) → text
    ├── Diffusion Models → images
    └── Multimodal Models → text + image + audio

In practice, modern systems combine several approaches. Large language models use transformers (an architecture) trained in a supervised way on text, then refined with reinforcement learning from human feedback (RLHF). They are neither purely supervised nor purely reinforcement learning — they are hybrids.

Knowing which category a system falls into is not everything, but it is the first step to understanding what it can do, what it needs to function, and where it is likely to fail. That is what makes the map worth having.