Types of Artificial Intelligence: The Complete Map

The three axes of the AI map

There is no single official list of AI types. The confusing labels – narrow AI, generative AI, machine learning, expert system – do not describe the same thing; they answer three independent questions. Separate these three axes and you can classify any system cleanly instead of blending buzzwords.

• Capability – How broad is the intelligence? From narrow specialisation through hypothetical generality to superintelligence.
• Mechanism – Where does the behaviour come from? From hand-coded rules or from patterns learned from data.
• Output – What does the system deliver? A prediction or classification (predictive) or new content such as text, images, code (generative).

Capability: narrow, general and superintelligence

This axis concerns the question Alan Turing raised in 1950: how general is the machine? The practical point: every system that exists today is narrow (weak) AI. General AI and superintelligence are terms of research and debate, not products you can buy.

• Narrow/weak AI (ANI): specialised on one task – translation, face recognition, chess. ChatGPT belongs here too, however versatile it seems.
• General/strong AI (AGI): a hypothetical system with human-like, cross-task intelligence. Does not exist; timing and feasibility remain open.
• Superintelligence (ASI): a purely speculative system that would surpass human intelligence in virtually every domain.
• Alternative lens (after Arend Hintze): reactive machines (e.g. Deep Blue), limited-memory systems (today's ML models), theory of mind and self-aware AI – the last two remain hypothetical.

Mechanism: rule-based versus learning

The second axis separates classic, symbolic AI from modern, learning AI. Rule-based systems (also called GOFAI or expert systems, influential since the 1970s) follow human-written if-then rules. They are transparent, deterministic and auditable – but rigid: whatever is not captured as a rule, they cannot do.

Learning systems (machine learning, ML) derive their behaviour from data rather than being explicitly programmed. Arthur Samuel coined the term back in 1959. Deep learning with neural networks is today's dominant variant. Such models are adaptive and strong at patterns, but data-hungry and often hard to interpret. Many real systems are hybrid (neuro-symbolic), combining both worlds.

Output: predictive versus generative AI

Since the rise of ChatGPT this is the most-discussed axis. Predictive AI answers 'what is this?' or 'what happens next?': it classifies, scores or forecasts. Generative AI answers 'create something new for me': it produces text, images, code, audio or video. Technically both often share the same ML foundations; generative models additionally learn to model the data distribution well enough to produce new, plausible examples. The Transformer architecture of 2017 made today's large language models possible.

• Predictive: spam filtering, credit scoring, demand forecasting, medical triage, image recognition, fraud detection.
• Generative: language models (ChatGPT, Claude, Gemini), image generators (diffusion models), code assistants, speech synthesis.

Learning methods in detail: supervised, unsupervised, reinforcement

Within learning AI, the main distinction is the kind of training. This distinction explains why some systems need labelled data while others learn from experience.

• Supervised learning: from examples with known answers (labels), e.g. emails as 'spam'/'not spam'. The basis of most predictive applications.
• Unsupervised learning: finds structure without labels, such as customer segments or anomalies (clustering).
• Reinforcement learning: learns by reward through trial and error; enabled AlphaGo (DeepMind, 2016) and robot control, for example.
• Self-supervised learning: creates the training signal from the data itself and is the basis on which large language models are pre-trained.

How the types combine – concrete examples

The three axes are independent, so every system can be described as a combination (capability • mechanism • output). All of today's examples are narrow AI:

• Spam filter: narrow • learning (supervised) • predictive.
• Tax or compliance rule engine: narrow • rule-based • deterministic/predictive.
• ChatGPT / Claude: narrow • learning (self-supervised) • generative.
• Online-shop product recommendation: narrow • learning • predictive.
• AlphaGo: narrow • learning (reinforcement) • decision-oriented.
• Image generator (diffusion model): narrow • learning • generative.

Swiss perspective: choosing and governing the right type of AI

The type of AI should follow the task, not the hype. For explainable, regulated decisions (credit, insurance, public administration), rule-based or predictive models with clear explainability are often the better choice. Generative AI excels at drafts, summaries and ideation, but carries its own risks: hallucinations, unclear sources and possible data leakage.

In Switzerland the revised Federal Act on Data Protection (nFADP) applies to the use of any type of AI; the supervisory authority is the FDPIC. Anyone feeding personal data into predictive or generative systems must clarify purpose, transparency and data-processing agreements – especially with cloud language models. Companies with an EU nexus also face the EU AI Act as an additional risk-based framework. Human oversight remains the guardrail across all types.