Types of Artificial Intelligence: Three Ways to Sort AI, and Five Levels

There are three common ways to sort the types of artificial intelligence. By breadth: narrow AI does a clearly scoped task, general AI handles a wide range of tasks, and superintelligence would beat every human. By function: four types, from reactive machines to self-aware AI (Hintze, 2016). By performance: five levels, from Emerging to Superhuman (Morris and colleagues at Google DeepMind, ICML 2024). These are researchers' frameworks, not an official list.

Below: what each list says, the percentile behind every DeepMind level, our arithmetic on where those percentiles would sit on the IQ scale (labelled clearly as not an IQ), and where today's chatbots sit in the paper's own table.

What are the main types of artificial intelligence: narrow, general and super?

Narrow AI does one task or a clearly scoped set of tasks. General AI handles a wide range of non-physical tasks, including learning new skills. Superintelligence is general AI that outperforms every human.

One common answer to what are the three types of AI is narrow, general and super, and DeepMind's table in Morris and colleagues has all three. It uses narrow and general as its two columns and names Artificial Superintelligence (ASI) at the top of the general column. These 3 types of artificial intelligence differ by breadth, not by how a system is built. The paper's narrow examples include the smart speakers Siri and Alexa, IBM's Watson, Deep Blue, AlphaGo and AlphaFold. Its general examples are chatbots: ChatGPT, Bard, Llama 2 and Gemini.

What are the four functional types of AI?

Arend Hintze of Michigan State University sorted AI by what a system can remember and understand: reactive machines, limited memory, theory of mind and self-awareness. He set the list out in The Conversation on 14 November 2016.

• Type I, reactive machines: no memory of the past and no use of past experience. Deep Blue, the chess computer that beat Garry Kasparov, is his example
• Type II, limited memory: can look into the past. Self-driving cars watch other cars' speed and direction, but Hintze says that information is only transient and is not saved as experience
• Type III, theory of mind: machines that form representations of other agents, such as their thoughts and emotions
• Type IV, self-awareness: machines that form representations of themselves, which Hintze says would mean building machines that have consciousness

These are different types of artificial intelligence systems judged by what is going on inside them, which is why they do not line up neatly with narrow and general. A chess engine is narrow and Type I. Hintze treated the last two types as targets, not products. He called theory of mind the next class of machines and wrote that self-aware machines were probably far off.

He wrote before today's chatbots and does not classify them, so we do not assign them a type either.

How do researchers grade AI performance: the five DeepMind levels?

Morris and colleagues grade AI on two things, performance and generality, and use five performance levels from Emerging to Superhuman. Their paper, "Levels of AGI for Operationalizing Progress on the Path to AGI", appeared at the 2024 International Conference on Machine Learning and was revised on arXiv in September 2025.

If you were looking for the 5 main types of artificial intelligence, this is the closest published list of five. It is a ladder of performance, not five kinds of machine. A general system is rated by the minimum level it reaches over most tasks, and it can reach higher levels on some of them.

• Level 1, Emerging: equal to or somewhat better than an unskilled human. Narrow example: simple rule-based systems such as SHRDLU. General: ChatGPT, Bard, Llama 2, Gemini
• Level 2, Competent: at least the 50th percentile of skilled adults. Narrow examples: smart speakers, Watson. General: not yet achieved
• Level 3, Expert: at least the 90th percentile of skilled adults. Narrow examples: Grammarly, Imagen, DALL-E 2. General: not yet achieved
• Level 4, Exceptional: at least the 99th percentile of skilled adults. Narrow examples: Deep Blue, AlphaGo. General: not yet achieved
• Level 5, Superhuman: outperforms 100% of humans. Narrow examples: AlphaFold, AlphaZero, Stockfish. General (Artificial Superintelligence): not yet achieved

Level 4 was called Virtuoso until the September 2025 revision renamed it Exceptional. The percentiles for every level above Emerging are measured against adults who have the relevant skill. Competent English writing, for example, is measured against adults who are literate and fluent in English, not against all adults.

Can you read the five DeepMind levels as an IQ?

No. The levels are percentiles of skilled adults on tasks, and the paper never mentions IQ. Reading them on the IQ scale is our arithmetic, and it is a loose one.

On the IQ scale the mean is 100 and the standard deviation (SD), the usual spread of scores, is 15. The 50th percentile is 100. The 90th percentile is about 119, and the 99th is about 135. Superhuman, which beats 100% of humans, has no finite value. You can check any percentile with our IQ percentile calculator.

• Level 2, Competent (50th percentile): IQ-style 100
• Level 3, Expert (90th percentile): IQ-style 119.2
• Level 4, Exceptional (99th percentile): IQ-style 134.9

The reading breaks in two places. The reference group is skilled adults on one task, not the whole population that an IQ scale is built on. And an AI that beats the 90th percentile at one task has not shown the broad ability that an IQ test samples. Our guide to why an AI score on an IQ-style test is not an IQ explains the same gap.

Which type is today's AI, and has any system reached general AI?

By DeepMind's own table, frontier chatbots were Emerging general AI in September 2023, and no public system had reached Competent general AI. The authors wrote that these models were Competent on some tasks, such as short essay writing and simple coding, but still Emerging on most, such as mathematics and tasks involving facts.

So the authors counted them as Level 1 general AI until performance rises across a broader set of tasks. The table in the version revised in September 2025 still lists Competent general AI as not yet achieved. That is a 2023 judgement carried into a 2025 revision, not a verdict on 2026 models.

For the evidence on whether AI now matches people, see our guides to what AGI means and to whether AI is smarter than humans. Two news pieces cover the test results: ARC-AGI-3 and whether AI models have a g factor. A related test of a different kind is covered in our guide to the Turing test study.

The paper adds a separate scale for autonomy, which describes how much control the human keeps, with levels such as AI as a tool, AI as a consultant and AI as an expert. A system can rate high on performance and low on autonomy, so the scales are not the same ladder.

To measure a person instead of a machine, take the IQ Metrics test. The percentile arithmetic above works the same way for people.

Common questions