Four rings, one inside the next

The word covers too much

When a shop sign, a job advert and a newspaper headline all say “AI”, they are usually pointing at four different things. The four sit inside each other like rings. Once you can name which ring something is in, most of the confusion goes.

Artificial intelligence is the outer ring. It is the oldest and vaguest of the four: any attempt to get a machine to do something that seems to need intelligence. A chess program from 1997 is in this ring. So is the routing that finds you a way round a traffic jam. Nothing about the ring says how the thing works.

Machine learning sits inside it. Here, the behaviour was not written down by a person. It was derived from examples. Somebody collected a pile of data, ran a procedure over it, and the result decides what the system does. Your bank’s fraud check is almost certainly in this ring.

Deep learning sits inside machine learning. It is the subset that uses neural networks with many layers: the technique that took over from about 2012 onwards. Face unlock on your phone is deep learning. So is the speech recognition in a voice note.

Generative AI sits inside deep learning. These are the models whose output is new content rather than a label or a number: text, images, audio, video. ChatGPT, Gemini, Midjourney, DeepSeek. This innermost ring is what most people now mean when they say AI, which is why the word has become so slippery: the newest and smallest ring has captured a name that belongs to the largest.

Large language models are a slice of that innermost ring. So a chatbot is a large language model, which is generative AI, which is deep learning, which is machine learning, which is artificial intelligence. Every one of those sentences is true, and each tells you something different.

Why the ring matters

Because it tells you what question to ask.

If something is in the outer ring only (rules a person wrote) then you can ask to see the rules. Somebody can print them out. When a housing benefit system refuses your claim, and the logic is written rules, the refusal can be explained line by line.

If it is machine learning, the rules are not printable. The right question becomes: what examples did you train it on, and how do you know it works? Those two questions are the whole of the difference. Nobody can show you the reasoning, so they have to show you the data and the testing instead.

If it is generative, add a third question: how do you know the output is true? A label can be checked against the right label. A paragraph invented on the spot has no right answer sitting next to it.

The thing beginners get backwards

Most people assume deep learning is better, and that anything serious must be a neural network by now. It is not so.

For data that lives in rows and columns (a spreadsheet of customers, a table of transactions, a sensor log) the method that usually wins is not deep learning at all. It is a family of models built from decision trees, with names like XGBoost, LightGBM and random forests. They train in minutes on an ordinary laptop, they need no graphics card, and on tabular data they routinely beat neural networks that cost a hundred times more to run. This has been tested repeatedly and it keeps coming out the same way.

This matters for you in a practical sense. If somebody tells you that predicting which customers will cancel requires a large language model, they are either selling something or have not looked. The free tools here are free and good: Python with scikit-learn and xgboost, or LibreOffice Calc for the first look at the data. The paid platforms (DataRobot, SageMaker, Vertex AI) are wrapping the same algorithms in a dashboard.

Where the rings are argued about

Be warned that the rings are conventions, not laws. Three honest disagreements:

• Reinforcement learning (learning by trial and reward, the method behind game-playing systems and part of how chatbots are tuned) does not sit neatly inside any of the four. Some diagrams place it beside machine learning, some inside it.
• A chess engine like Stockfish scored positions with a hand-written evaluation from its first release in 2008 until 2020, when it added a small neural network. Since 2023 its evaluation has been entirely neural, though its search is still hand-written. It moved rings without changing its name.
• ”AI” as a product label is applied to plain database queries and if statements every day, because the word raises valuations and prices. There is no authority stopping this and no penalty for it.

So the rings are a thinking tool, not a certificate. Their use is that they replace one useless question with three useful ones.

The habit

When you next meet a system described as AI, ask yourself which ring, out loud if you like. If you cannot tell, that is itself information: it means nobody has told you how the thing works, and you should treat every claim about it as unverified until they do.