Different Types of AI — and What Isn't AI, and Why That Doesn't Always Matter

When you hear people talk about AI, what actually comes to mind? For most people, it's some mix of a chatbot, a sci-fi robot, and a vague sense that computers are getting smarter. The word gets used for all of it — and that's exactly the problem. “AI” now covers things that have almost nothing in common with each other.

Here's a good example of just how far the word gets stretched.

There's a PC game called Alien: Isolation — a survival horror game where you're being hunted through a space station by a single alien. Watch anyone play it, or watch a streamer get chased through the vents, and you'll hear the same thing every time: the alien feels like it's actually hunting you. Like it's learning your habits. Like it knows you keep hiding in that one locker, so this time it checks there first.

People have called it “the smartest AI in gaming” for over a decade. And here's the twist — it's not learning anything at all. It's a fixed set of rules, written by very clever designers, that reacts to what you do without ever actually adapting or thinking. The alien can't see you directly — it only “senses” you, the same way a real predator would, through noise and movement. That's the trick. It feels intelligent because it's working with limited information, not because there's a mind behind it. Some of its more advanced moves are even sitting there unused from the start of the game, only switched on once you've triggered them enough times — which feels like learning, but is really just a very well-timed reveal.

So: no learning, no adapting, nothing close to what we mean by “AI” today. Just extremely clever, entirely fixed programming. And it's still one of the best horror games ever made — nobody needed to slap “AI” on the box for it to work.

Which raises the obvious question: if that's not AI, what is?

So what actually counts as AI?

Here's the thing — “AI” isn't one thing. It's a label sitting over a bunch of genuinely different technologies, and most people never see the difference spelled out. Most of what's below is real AI, just not always recognised as such. One entry, though, is on this list precisely because it's the opposite — it wears the label without earning it. Here's the full picture, plainly, with an example for each:

• Chat/Conversational AI — what most people picture first. You type a question, it answers. ChatGPT, Claude, Gemini — this is the one everyone's used by now.
• Generative AI — creates new content instead of just answering questions: images, audio, video, code. Chat tools are actually a subset of this. Ask ChatGPT to “draw a cartoon of me based on what you know about me” — that's generative AI at work, just with a picture instead of a paragraph.
• Agentic AI — doesn't just respond, it acts: takes several steps on its own, uses other tools, works toward a goal without you approving every move. Think JARVIS from Iron Man, quietly running checks and executing tasks in the background. Genuinely newer, and most people haven't actually used a real version of this yet, even if the word gets thrown around a lot.
• Predictive/Analytical AI — the oldest, least flashy category, and probably running more of your life than any of the above. Fraud alerts on your card, your credit score, “customers who bought this also bought.” No chatting, no pictures — just a number or a category, predicted from data.
• Computer Vision — AI that reads images or video. Your phone unlocking when it sees your face, a doctor's scan getting flagged for a closer look.
• Recommendation Systems — a cousin of predictive AI, and genuinely AI in the technical sense — but worth naming separately, since it's the AI you actually experience the most (every app feed, every “you might also like”) while being the one people least often think to call AI at all.
• RAG (Retrieval-Augmented Generation) — not really its own type of AI, more of a technique. Picture a company that's built a tool where you can “chat” with all of its internal documents — policies, product info, whatever's sitting in its files. That's RAG. The AI isn't specially trained on that company's data; it's just been given a fast way to look things up first, then answer using what it found. This is what's actually behind most claims like “our AI knows your account” or “chat with your documents.”
• Rule-based automation, wearing an AI costume — and here's the impostor. Unlike everything above, this genuinely isn't AI in any technical sense — no model, no learning, nothing under the hood but plain logic. A lot of what gets called “AI” in business is just old-fashioned if-this-then-that scripting with a new label stuck on it. “Our AI routes your support ticket” quite often just means a basic keyword-matching script that's been running since well before anyone called it AI.
• Narrow AI vs. AGI (Artificial General Intelligence) — every AI you've ever used, including the impressive stuff, is narrow: built to do a specific job or range of jobs well. AGI is the sci-fi version — think the Terminator, minus the killing. Set that part aside and what's left is a machine using adaptive intelligence the way a person would: reading a room, adjusting on the fly, working out how to outsmart whoever's in front of it — across any situation, not just the one job it was built for. That's the part that doesn't exist yet, no matter what a headline might suggest.

And the alien from earlier sits off to the side of all of this — it's what's usually called “game AI,” a much older use of the term built from fixed logic and pathfinding, going back decades before any of the above existed. It just never merged with what the word means everywhere else. Game bots are the same story, and probably a more familiar one — the “AI” opponent in a racing game or a shooter isn't thinking about you at all. It's running a formula and reacting to whatever you do. Which is exactly why, if you play against the same bot enough times, you start to find its pattern and beat it consistently. You're not out-thinking a mind. You're just learning a formula.

The word isn't the problem — what it's covering up sometimes is

Here's where it gets genuinely interesting, and it's not really about definitions anymore.

A while back, at a previous company, I watched a product get launched — a RAG tool, the “chat with your documents” kind from earlier in this post — with all the fanfare you'd expect. Big internal announcement, a lot of talk about how it was going to change how people worked.

I never actually used it. Not because it didn't work — because I always had a person I could just ask instead, and that was faster and more reliable than typing a question into a tool and hoping it found the right document. Looking back now, knowing a lot more about what AI can actually do than I did then, I don't think that tool was a bad build. I just think its real impact — financially, in time saved, in anything that actually mattered to the business — was probably fairly marginal, especially compared to what other AI approaches could have delivered with the same effort and budget. It was AI, genuinely. It just wasn't the right AI for the actual problem.

Compare that to the alien. Nobody marketed it as AI. It was sold purely on being terrifying, and it delivered — and people are still calling it “the best AI in gaming” a decade later, correctly sensing something clever was going on, even without the label.

Same word, very different outcomes. One thing was never marketed as AI at all, and turned out to be genuinely brilliant. The other was marketed as AI, was technically legitimate, and still barely moved the needle — because being real AI was never the actual problem it needed to solve.

That's really the whole point of pulling these categories apart — not to win an argument about what “really” counts as AI, but so the word stops doing the convincing for you. When a business says “we use AI,” the label alone tells you almost nothing — not whether it's genuine, and not whether it actually solves anything. What matters is what the thing does, and whether that was ever the actual problem worth solving.

The word was never the point. What the thing does with it — that always was.