No standards body, regulator, or major textbook defines a fixed list of seven stages of artificial intelligence. The phrase usually refers to a popular list that joins four "types" of AI proposed by Arend Hintze in 2016 (reactive machines, limited memory, theory of mind, and self-awareness) with three capability tiers (narrow, general, and super intelligence). Only the first two of those seven items describe systems that exist, and the order of the list is a convention, not a measured progression.
Key takeaways
- The seven-stage list is a teaching convention built from two separate ideas, so its items are not steps that every system passes through.
- Hintze's 2016 article names reactive machines and limited memory as existing, and theory of mind and self-awareness as not yet built.
- A 2023 research paper from Google DeepMind authors proposes six performance levels, numbered 0 to 5, that rate depth and breadth separately.
- Tests of theory of mind in language models have produced disputed results, and a 2023 study showed that trivial rewording changed outcomes.
- A claim that a product sits at a given stage is a marketing statement unless it names a test and a result.
Where the seven-item list comes from
Hintze, a professor at Michigan State University, published "Understanding the four types of AI" on November 14, 2016. He defined the first type as reactive machines with no memory, using IBM's Deep Blue chess computer as the example. The second is limited memory, where a system looks briefly into the past, as self-driving cars do when tracking the speed and direction of nearby vehicles. The third would model the thoughts and emotions of other agents. The fourth would form representations of itself. Hintze wrote that machines of the last two kinds were not yet achieved.
The other three items come from a different vocabulary: artificial narrow intelligence for single-purpose systems, artificial general intelligence for human-level range, and superintelligence for systems beyond that. Course pages and blogs often stack the two vocabularies into one list of seven. Searching for a primary source that publishes the combined list as a standard turned up none, so it is best treated as a mnemonic.
The seven items and what the evidence supports
The table below is a reading aid. The status column describes the state of the evidence as of October 2026 and rests on the sources listed at the end.
| Popular label | Meaning in the list | Status |
|---|---|---|
| Reactive machines | Respond to the current input only, with no stored experience | Exist (Deep Blue is the standard example) |
| Limited memory | Use recent observations or training data to act | Exist (most deployed machine learning) |
| Theory of mind | Model beliefs and intentions of people | Research question, results disputed |
| Self-aware | Hold a representation of their own state | No accepted example |
| Narrow AI | Perform one task or a small set | Exist |
| General AI | Match human range across tasks | No agreed test; contested definitions |
| Super AI | Exceed humans across tasks | Hypothetical |
The list also mixes two different questions. The first four rows ask how a system uses information. The last three ask how broad and how strong its abilities are. A chess engine is both reactive-style and narrow, so the rows overlap instead of forming a ladder.
A staged framework that comes with definitions
For readers who want stages with testable definitions, a 2023 paper titled "Levels of AGI for Operationalizing Progress on the Path to AGI" by Meredith Ringel Morris and seven co-authors proposes a grid. One axis is performance, measured against skilled adults, and the other is generality. The performance levels run from Level 0 (no AI, such as a calculator) through Emerging (equal to or somewhat better than an unskilled human), Competent (at least the 50th percentile of skilled adults), Expert (90th), Exceptional (99th), and Superhuman (outperforms all humans).
The authors placed frontier language models of September 2023 at Emerging for general ability, with Competent performance on some tasks such as short essays and simple coding. Because it is a proposal from researchers at one company and others, it is a discussion framework, not a standard, but it shows the difference between a label and a defined, testable scale.
Why theory of mind is a poor stage marker
Some papers reported that language models pass psychology tests of false-belief reasoning. Tomer Ullman's February 2023 preprint, "Large Language Models Fail on Trivial Alterations to Theory-of-Mind Tasks", showed that small changes which keep the same logic reversed those results. He argued that skepticism should be the default and that outlying failures count for more than average success. The result does not prove that no model can ever do the task. It shows that passing a test item is weak evidence of a general capacity.
Questions to ask when a product claims a stage
- Which test was used, and was it published with the data?
- Was the same system tested on reworded or unfamiliar items?
- Does the claim describe the underlying model, or a product built around it?
- Is the stage a capability claim (what it can do) or an internal-design claim (how it stores information)?
FAQ
Are there officially seven stages of AI?
No. No standards body or regulator publishes such a list. The number seven comes from combining Hintze's four types with three capability tiers in teaching material.
Which stage is ChatGPT at?
The question depends on which list is used. Under the type list it is a limited-memory system, since it uses training data and a context window. Under the 2023 levels paper, general-purpose language models of that period were rated Emerging for generality.
Has any AI reached theory of mind or self-awareness?
No accepted evidence exists for either. Test results on theory of mind have been disputed, and self-awareness has no agreed test.
Is general AI the final stage?
Not in most lists. Super AI is usually placed after it, and the 2023 levels paper goes further by treating performance and generality as two separate scales.
Related reading
- How the parts of an expert system work together
- Conversational interfaces and how they developed
- A free learning path for beginners
- Where AI is used in scientific research
More in AI Basics.
Last verified: October 2026
Sources
- Understanding the four types of AI, from reactive robots to self-aware beings, Arend Hintze, The Conversation, November 14, 2016
- Levels of AGI for Operationalizing Progress on the Path to AGI, Morris et al., arXiv 2311.02462, 2023 (ICML 2024)
- Large Language Models Fail on Trivial Alterations to Theory-of-Mind Tasks, Tomer Ullman, arXiv 2302.08399, 2023