Calling ChatGPT 'AI' is like calling your dog 'an animal'.

Google has a diagram that ends a hundred LinkedIn arguments: a big circle labeled AI, a smaller circle inside it labeled ML, and a bright dot deep inside that labeled Gen AI. Three circles, one inside the next.

Here's what it's quietly telling you: these words are not synonyms. They're nested dolls.

So when someone says "ChatGPT is AI," they're technically right and almost completely unhelpful — the same way "your dog is an animal" is right. Sure. It's also a mammal, a vertebrate, a Labrador, and Steve. You skipped about five levels of precision to land on the vaguest word available.

That vagueness isn't harmless. It's the exact gap marketers slip through to slap "AI" on a spreadsheet — and the gap regulators are now writing fines into. Let's un-blur the circles.

1. The words are nested, not interchangeable.

Read the diagram from the outside in. AI is the whole ambition — machines doing things that used to need a human. Machine learning is one way to get there: instead of hand-coding rules, you let a machine learn patterns from data. Deep learning is one kind of ML — neural networks with many layers. Generative AI is what you get when you point that at making new content instead of just classifying it. Foundation models are the giant, broadly-trained models underneath. And an LLM is one species of foundation model — the one tuned for language.

So the real lineage is: AI ⊃ ML ⊃ deep learning ⊃ foundation models ⊃ LLMs ⊃ the chatbot in your tab. Every arrow is "a more specific kind of." ChatGPT isn't "an AI." It's a great-great-grandchild of the word.

Here's that same nesting as actual nested shells. The outer shell is the broadest idea; the glowing gold core is the thing you actually open in a browser tab. Drag it around. Tap any shell to see what it means.

And it's young. "AI" was coined at the Dartmouth workshop in 1956 ↗. The transformer that makes modern chatbots possible — "Attention Is All You Need" ↗ — is from 2017. The term "foundation model" didn't exist until Stanford coined it in August 2021 ↗. ChatGPT launched November 2022. The thing everyone now means by "AI" is the thinnest, newest slice of a field older than your parents.

2. The boundary that actually matters: who labels the data.

Most people memorize the circles and still miss the line that actually separates old AI from the new wave. It's not "neural networks." It's labels.

Classic machine learning is supervised: a human labels thousands of examples — this email is spam, this one isn't — and the model learns to copy that judgment. The bottleneck is people. Someone has to sit and tag the data, by hand, forever.

Foundation models broke that. Stanford's definition is precise: a foundation model is trained on broad data "generally using self-supervision at scale." Translation: nobody labels anything. You feed it the raw internet and make it play fill-in-the-blank — predict the next word, over and over — until it has absorbed grammar, facts, and reasoning as a side effect. Unlabeled data is basically free and basically infinite. That's the unlock. Not the network. The fact that you fired the labelers.

// Supervised ML: humans tag the data
const trainingData = [
{ text: "win a free iphone now", label: "spam" },
{ text: "lunch at 1pm?",         label: "ham" },
{ text: "claim your prize!!!",   label: "spam" },
// ...repeat 50,000 times, by hand
];

model.train(trainingData);
// bottleneck = humans tagging examples

// Foundation model: self-supervised
const trainingData = theWholeInternet; // unlabeled

// task: just predict the next word
//   "lunch at 1___"  -> "pm"
//   "the capital of France is ___" -> "Paris"

model.pretrain(trainingData);
// it learns grammar, facts, reasoning
// as a SIDE EFFECT. no labelers needed.

The cost cliff nobody puts in the explainer.

There's a reason the company that trains the foundation model and the company that "has an AI feature" are almost never the same company. Training one of these from scratch fell off a financial cliff. Drag the slider and watch the bill.

3. "We use AI" is usually a marketing sentence, not an engineering one.

Once you know the hierarchy and the cost cliff, the pitch decks read differently. When a seed-stage startup with no research team says "powered by our proprietary AI," the honest translation is almost always: a prompt and an API call to someone else's foundation model, with a nice UI on top. Which is a fine business! It's just not "their AI."

This is common enough that regulators put a name and a price on it. On March 18, 2024, the SEC brought its first "AI washing" enforcement actions ↗, fining two investment advisers — Delphia and Global Predictions — a combined $400,000 for overstating their AI. The Enforcement Division's Gurbir Grewal put it bluntly: if you're rushing to claim you use AI to capitalize on the hype, "stop."

There's also the inverse, the "AI effect": the moment a technique actually works — spam filters, GPS routing, autocomplete — we stop calling it AI and call it "software." So "AI" stays stuck as the label for whatever doesn't quite work yet, which is a weird thing to build your whole brand on.

So what's the actual shape?

The circles aren't trivia. They're a precision tool. Use the smallest word that's actually true:

• AI — the 70-year-old ambition. Almost always too vague to be useful in a sentence.
• ML — learning patterns from data instead of hand-coded rules.
• Deep learning — ML with many-layered neural networks.
• Foundation model — one giant model, self-supervised on unlabeled data, adaptable to many tasks.
• LLM — a foundation model specialized for language. This is the thing you're actually using.

Next time someone says "we added AI," ask the only question that matters: which layer? The answer tells you whether they built something, fine-tuned something, or just typed a good prompt.

Your dog is an animal. It's also a Labrador named Steve. Both are true; only one tells me anything. ChatGPT is AI — and it's also a language model, which is a foundation model, which is deep learning, which is ML. Say the specific one. It's how people know you actually understand the thing.