A phishing email never reached your inbox this morning. Something binned it in roughly 40 milliseconds, and that something is artificial intelligence. So is the model that spotted a hurricane forming days before the official forecast caught up, and the one that suggested the song you skipped on the drive in.
Here’s the awkward part. The phrase “types of artificial intelligence” gets used for at least four different classification schemes. Some people split AI by ambition, some by capability, some by what the system actually does, and some by how it learns. All four are useful. Mixing them together is where most of the confusion comes from.
The three-tier split: narrow, general and super
This is the classification researchers reach for first, because it’s about how much an AI can do rather than how it works.
Narrow AI (artificial narrow intelligence)
Everything in commercial use today lives here. Spam filters, credit scoring models, Google Translate across 130-odd languages, the recommendation engine behind a streaming queue, AlphaFold and its 200 million predicted protein structures. Narrow AI is very good at one job and brittle the moment you step outside it. A chess engine that plays at superhuman level cannot book you a flight, recognise your face, or explain why it resigned.
Artificial general intelligence
AGI describes a system that can handle essentially any cognitive task a person can, including the ones it wasn’t trained for. Nobody has built one, and there’s no agreed test for deciding when someone has. The ARC-AGI benchmark, built around novel visual puzzles, still humbles frontier models. The open argument is whether scaling language models gets us there or whether something structural is missing: persistent memory, reliable planning, a working model of cause and effect. That gap is exactly why labs are chasing always-on agents that keep working in the background rather than waiting for a prompt.
Artificial superintelligence
The top rung. Nick Bostrom defined it in 2014 as an intellect that substantially outperforms humans across virtually every domain, including scientific research and social persuasion. Nothing close exists. Forecasters disagree wildly on whether it’s 30 years away or 300, and a minority argue the question is malformed. Treat the term as a placeholder for a debate, not a product category.
The four types you’ll find in textbooks
Computer scientist Arend Hintze sketched a second framework, and it stuck because it’s easy to picture. It describes stages of awareness rather than intelligence in the abstract.
Reactive machines
These respond to the current situation and store nothing afterwards. IBM’s Deep Blue evaluated around 200 million positions per second to beat Garry Kasparov in 1997, and it had no memory of the games they’d already played. Feed it the same position twice and it reacts identically. Most classic game-playing systems work this way.
Limited memory
Here the system uses recent history to inform the next move. Self-driving stacks keep a rolling picture of nearby vehicles and pedestrians; large language models carry a context window of 128,000 tokens or more. The memory is temporary and scoped. Ask a chatbot about last Tuesday’s conversation with no saved history and it has nothing.
Theory of mind
A system that models what other agents believe, want and might be wrong about. This is where sarcasm, negotiation and white lies live. No deployed AI has it. Chatbots can produce impressively empathetic text without any representation of your mental state, which makes the failure easy to miss until a customer is furious and the model keeps apologising cheerfully.
Self-aware
Machines with consciousness or genuine self-knowledge. Speculative philosophy, not engineering. Every claim you’ve read about a “self-aware” chatbot has been a language model generating plausible sentences about its own inner life.
Types by what the system actually does
Strip away the theory and most working AI sorts into a handful of applied families. This is the vocabulary you’ll meet in job ads and product specs.
- Computer vision — reading X-rays, spotting defective solder joints on a production line, counting cars at an intersection.
- Natural language processing — translation, summarisation, sentiment scoring, search ranking. Frontier reasoning models now sit here too; DeepMind’s AI co-mathematician proposes proof steps alongside researchers.
- Speech and audio — transcription, voice isolation, hearing-aid tuning that adapts to a noisy room.
- Recommender systems — the models deciding which product, video or article appears next.
- Predictive forecasting — demand planning, equipment failure, weather. Meteorological agencies now lean on systems like DeepMind’s hurricane prediction model, which flagged storm tracks ahead of conventional physics-based forecasts.
- Generative models — text, images, audio, proteins, molecules. The same techniques that design a candidate drug can be pointed at less friendly targets; a team used generative AI to design 16 new viruses, which tells you something about how thin the dual-use line has become.
- Robotics and control — warehouse pickers, surgical assistants, agricultural drones.
These categories overlap constantly. A modern voice assistant runs speech recognition, language modelling, retrieval and speech synthesis in sequence, and calling it “NLP” undersells what’s happening.
Types by how the model learns
The last useful cut is the training method, and it explains a lot about why systems fail the way they do.
Supervised learning trains on labelled examples: 50,000 mammograms marked benign or malignant. It’s the workhorse of industry and it inherits every bias in the labels. Unsupervised learning hunts for structure without answers attached, which is how customer segmentation and anomaly detection usually work. Self-supervised learning builds its own labels by hiding part of the input and predicting it, the trick behind modern language models. Reinforcement learning learns from reward signals through trial and error; AlphaGo’s move 37 in game two against Lee Sedol, a move commentators called a mistake before it won, came out of that process.
Few production systems pick just one. Instruction-tuned chatbots typically combine self-supervised pretraining with reinforcement learning from human feedback.
Risk categories are a different animal
Regulators classify AI by harm rather than capability, and the two taxonomies don’t line up neatly. The EU AI Act, in force since August 2024, sorts systems into four tiers: unacceptable risk (social scoring, most real-time biometric surveillance, banned outright), high risk (hiring, credit, medical devices, critical infrastructure, all facing audit and documentation duties), limited risk (chatbots and deepfakes, which must disclose what they are), and minimal risk (spam filters, video games, no obligations).
A narrow AI system that sorts CVs carries more legal weight than a general-purpose model used to write birthday poems. Capability and consequence are separate axes, and policy people care far more about the second.
The categories are already blurring
Foundation models broke the tidy labels. A single multimodal system can caption an image, write code, transcribe a meeting and call an API, which makes “narrow” feel inadequate even though none of it is general. Agentic tools that shop, book and negotiate on your behalf push further, and the commercial fights are already starting: Amazon shut Meta’s shopping agent out of its storefront because an AI acting on a customer’s behalf is a different kind of visitor than a person browsing.
So the taxonomy is a map, not the territory. When someone asks which type of AI a system is, the useful follow-up is narrower: what does it do well, where does it fail quietly, and who is accountable when it gets something wrong. A model that’s 97% accurate on a task performed a million times a day produces 30,000 errors, and no amount of labelling it “narrow” or “general” tells you whether those errors are tolerable.