AI vs Machine Learning vs Generative AI

Artificial intelligence — the outer ring

AI is the broad label for systems that perform tasks normally requiring human cognition: reasoning, perception, decision-making, planning.

It is deliberately wide. A rules engine that approves a loan against fixed criteria is AI. So is a chess program that searches moves. Neither learns anything, and both predate the current wave by decades.

This matters commercially: "AI-powered" is true of software that contains no learning at all. It is not a lie, it is just not information.

Machine learning — systems that learn from data

ML narrows it to systems that learn patterns from data rather than following rules a person wrote. You supply examples; the system derives the rule.

The classic split is by what the data looks like. Supervised learning trains on labelled examples — inputs paired with correct answers. Unsupervised learning finds structure in unlabelled data. Reinforcement learning learns through trial and feedback against a reward.

Most production ML in business today is unglamorous and supervised: churn prediction, demand forecasting, fraud scoring, document classification. It is also where the technology is most reliable, because the task is narrow and the success measure is obvious.

Deep learning — many-layered neural networks

Deep learning is ML using neural networks with many layers, which learn hierarchical representations: early layers pick up simple features, later ones combine them into complex ones.

It is what made progress on messy, high-dimensional inputs — images, audio, language — where hand-designed features had stalled. The cost is data and compute, and a model whose reasoning is much harder to inspect than a decision tree.

Generative AI — models that produce new content

Generative AI is the part of the field that creates new content — text, images, audio, video, code — rather than classifying or scoring existing content. Generative models are older than deep learning (Markov chains were generating text decades ago), but everything people mean by the term today is built on deep neural networks, which is why it sits inside that ring.

The distinction that matters in practice is the shape of the output. A classifier returns one of a fixed set of answers and you can measure whether it was right. A generative model returns open-ended content, and "right" is often a judgement call. That single difference drives most of the difficulty in evaluating, testing and governing these systems.

Large language models are the generative family most businesses encounter. They are first pre-trained on very large text corpora to predict the next token, which produces a model that can continue text but does not reliably follow instructions. The assistants people actually use have been through a second stage: tuning on examples of instructions and good answers, and on human (or AI) judgements of which responses are better. That second stage is what turns a text predictor into something usable for summarising, drafting, extracting, translating and writing code.

Why the nesting matters when you are buying

If a vendor says "AI", ask which ring. The failure modes are completely different: a rules engine fails predictably and visibly, a supervised model degrades quietly as the world drifts away from its training data, and a generative model produces fluent text that can be confidently wrong.

Ask how they know it worked. For a classifier there is a straight answer — accuracy, precision, recall against a held-out set. For a generative system the honest answer involves human review, defined evaluation criteria, or a downstream signal like whether the tests passed.

And be suspicious when a genuinely narrow problem gets a generative solution. A great deal of what is currently sold as generative AI is a classification problem that a much smaller, cheaper, more testable model would solve better.