What types of artificial intelligence are there, and what can they do in practice?
Why not all AI is the same
The term “artificial intelligence” (AI) is ubiquitous, yet often misunderstood. AI is not a single uniform system, but rather an umbrella term for a wide range of technologies, methods, and applications. While some AI models can independently generate texts, others can only produce predefined responses. In order to sensibly assess opportunities and risks, it is important to know the different types of AI.
This distinction is particularly crucial in public administration: Not every AI is suitable for every task. Not every AI may legally be used. Being able to classify AI systems lays the foundation for informed decisions, regulatory compliance, and fit-for-purpose applications.
Systematic classification: the four types of AI
A common classification distinguishes four types of artificial intelligence based on their level of cognitive capability – i.e. how they learn, reason, and understand.
How does AI work?
In addition to the cognitive categorisation, AI can also be distinguished by its mode of operation or complexity.
Symbolic AI vs neural networks
Symbolic AI:
is based on fixed rules, logical inference and knowledge bases (classic)
Neural networks:
learn from large volumes of data, in an abstracted way similar to the human brain (modern)
Relevance in practice: Symbolic systems are suitable for the likes of rule-compliant processes (such as control logic, where fixed rules and decisions are predefined), while neural networks are suited to unstructured data (such as images and language).
Narrow AI vs. General AI
- Narrow AI (weak AI): is specialised in a clearly defined task (e.g. text recognition)
- General AI (strong AI): can solve a range of tasks flexibly, comparable to human intelligence
Note: All AI applications currently available, including powerful LLMs, are considered Narrow AI. General AI remains theoretical to date.
What is an LLM (Large Language Model)?
An LLM is an AI system trained on vast quantities of text and able to generate, analyse, and understand language, at least in statistical terms. Examples include GPT-4 and BERT (an AI model by Google). They belong to type 2 AI (“limited memory”) and are considered powerful, but not “conscious”, tools. In public administration, LLMs can, for example, help analyse documents, classify inputs, or automatically respond to citizen enquiries.
AI and its significance for public administration
For public authorities, it is crucial to understand what type of artificial intelligence they are dealing with. This is because not every AI technology is suitable for use in sensitive administrative processes, and not every one is legally permissible.
- At present, public administration almost exclusively uses AI systems from the “limited memory” (Type 2) category. This includes, for example, large language models (LLMs) such as ChatGPT. These systems are based on past training data and can use it to recognise patterns, make predictions, or generate text. They are powerful, but not conscious, not autonomous, and not “understanding”.
- The more advanced types 3 (“Theory of Mind”) and 4 (“self-aware AI”) currently exist only as theoretical concepts or long-term research ideas. They are currently not market-ready, not regulated, and therefore neither relevant nor usable for day-to-day administrative practice.
Especially in the context of regulatory issues, such as the EU AI Act, which classifies AI systems into different risk categories, assigning a system to a specific AI category is very important. It helps to
- assess risks realistically,
- define requirements for transparency and control, and
- establish appropriate testing and governance mechanisms.
Anyone making decisions about AI applications in public administration should understand the functional principle on which the system is based. Only then can the limits of use, risks and potential be assessed soundly and decisions made with confidence.
