Key takeaways
- An AI system is a machine-based system that infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that influence physical or virtual environments.
- The EU AI Act sets the binding definition in Article 3(1), and the European Commission’s 2025 guidelines break it into seven characteristics, with inference as the element that separates AI from ordinary software.
- AI systems are commonly grouped by capability (narrow, general, super) and by functionality (reactive machines, limited memory, theory of mind, self-aware); almost every system in use today is narrow, limited-memory AI.
- The same regulation classifies AI systems by risk: prohibited, high-risk, limited (transparency), and minimal. The risk tier, not the technology, decides the obligations.
- Governing AI systems starts with knowing which ones you run. A current inventory that records each system’s purpose, role, and risk tier is the foundation for EU AI Act, ISO 42001, and NIST AI RMF compliance.
What is an AI system?
An AI system is software that reasons from data. Instead of following a fixed script written line by line, it draws conclusions from the inputs it receives and produces an output that can act on the world around it. That output might be a credit decision, a medical triage suggestion, a generated image, or a route recommendation.
The most consequential definition is the legal one. The EU AI Act fixes it in Article 3(1), and every downstream obligation depends on whether a system meets it:
“a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments.”
Read that definition as a chain. A machine runs the system. It works with some independence from step-by-step human control. For a goal it has been given, whether stated openly or built in implicitly, it takes an input and infers an output. That output then influences an environment, digital or physical. This is a functional definition: it does not name a technology like neural networks or large language models, so it stays valid as techniques change.
That legal framing is what sets this guide apart from a dictionary entry. Knowing that a chatbot or a fraud model is an AI system is only useful once you can tell whether it falls inside a regulation and what that regulation then asks of you.
The seven characteristics of an AI system
In February 2025 the European Commission published guidelines on the AI system definition to help organisations apply Article 3(1). The guidelines unpack the definition into seven characteristics: machine-based, autonomy, adaptiveness, objectives, inference, outputs, and influence on the environment. Not all seven carry equal weight.
Machine-based and autonomous
Every AI system is machine-based, meaning it is developed with and runs on machines. The Commission reads this broadly, so it covers conventional computers and extends to emerging quantum computing hardware. Autonomy means the system operates with some degree of independence from human action. The bar is low: even limited autonomy counts. A model that scores a loan application without a person recalculating each variable is operating autonomously in this sense, even if a human signs off on the final decision.
Inference: the defining element
Inference is the characteristic that matters most. An AI system infers, from the input it receives, how to generate its outputs. This is the line between AI and ordinary software. Systems that only execute rules a human wrote out in full, simple heuristics, or basic data processing fall outside the definition, because they do not infer, they follow instructions. The Commission is explicit that rule-based systems entirely specified by people are not AI systems. Inference covers both machine-learning approaches, which learn patterns from data, and knowledge or logic-based approaches that derive conclusions from encoded rules and facts.
Adaptiveness, objectives, and outputs
Adaptiveness, the capacity to keep changing after deployment, is often assumed to be mandatory. It is not. Article 3(1) says a system “may” exhibit adaptiveness, so a model whose weights are frozen at release is still an AI system. Objectives can be explicit, such as a stated accuracy target, or implicit, embedded in how the system was trained. The outputs the definition lists are broad on purpose: predictions, content, recommendations, and decisions. Because those outputs can influence physical or virtual environments, a recommendation engine that reorders a feed qualifies just as much as a robot arm that moves.
Types of AI systems
Beyond the regulatory definition, AI systems are usually sorted along two axes that answer the questions searchers ask most: how capable is the system, and how does it handle information. Both are worth knowing, because they shape how much oversight a system needs.
By capability: narrow, general, and super AI
By capability, AI falls into three tiers. Narrow AI, also called weak AI, is built for a specific task: spam filtering, image classification, demand forecasting, or a language model answering prompts. Every AI system in production today is narrow AI. General AI, sometimes called artificial general intelligence, would match human ability across any intellectual task; it does not yet exist. Super AI, a hypothetical system that would exceed human intelligence across the board, is further still. For governance purposes, the practical universe is narrow AI, which is precisely what the EU AI Act regulates.
By functionality: reactive, limited memory, theory of mind, self-aware
By functionality, a common four-part taxonomy describes how a system uses information. Reactive machines respond only to the present input with no memory of the past; classic game-playing engines are the textbook case. Limited memory systems learn from historical data and use recent context, which describes nearly all modern machine learning, from recommendation engines to self-driving perception stacks and large language models. Theory of mind systems, which would model the beliefs and intentions of the people they interact with, remain a research goal. Self-aware systems, which would possess consciousness, are speculative. Mapping your own tools to these types is a quick sanity check: if a system learns from data and adapts to context, it is limited-memory AI and it needs governance.
How the EU AI Act classifies AI systems by risk
Here is where the governance lens changes the picture. The EU AI Act does not regulate all AI systems the same way. It sorts them into four risk tiers, and the tier, not the sophistication of the model, decides the rules.
Prohibited AI practices sit at the top. Article 5, in force since 2 February 2025, bans eight practices, including manipulative systems that cause significant harm, social scoring by public authorities, untargeted scraping of facial images, emotion recognition in workplaces and schools, and certain uses of real-time remote biometric identification in public spaces.
High-risk AI systems are the regulation’s centre of gravity. Under Articles 6 to 49, a system is high-risk either because it is a safety component of a product already regulated under EU law (such as medical devices or machinery) or because it operates in one of the eight areas listed in Annex III: biometrics, critical infrastructure, education, employment, access to essential private and public services (including creditworthiness and insurance), law enforcement, migration and border control, and the administration of justice. A hiring model that ranks candidates or a system that scores loan applications lands here.
Limited-risk systems carry transparency duties under Article 50: chatbots must disclose that a user is talking to a machine, and generative outputs and deepfakes must be labelled. Everything else is minimal risk and carries no specific AI Act obligations, though other laws such as the GDPR still apply. Classifying each system into the right tier is the pivot of any AI risk management programme.
Who is responsible: providers and deployers
The EU AI Act assigns duties by role, not by who happens to touch the system. Two roles carry most of the weight. A provider, defined in Article 3(3), is the party that develops an AI system, or has one developed, and places it on the market or puts it into service under its own name or trademark. A deployer, defined in Article 3(4), is a party that uses an AI system under its own authority, outside a purely personal activity.
The strictest obligations fall on providers. The catch is that a deployer can become a provider without building anything. Under Article 25, an organisation inherits provider obligations if it puts its own name or trademark on a high-risk system, substantially modifies one, or repurposes a lower-risk system for a high-risk use. A company that fine-tunes and rebrands a third-party model can cross that line, which is why role assignment belongs in your records, not in an assumption. For increasingly independent systems, this question grows sharper, as our analysis of autonomous AI agents explains.
Governance obligations for AI systems
Once a system is high-risk, a defined obligation set applies. Providers must run a risk-management system (Article 9), govern their training and testing data against bias (Article 10), keep technical documentation (Article 11) and automatic logs (Article 12), provide transparency and instructions for use (Article 13), design meaningful human oversight (Article 14), and meet accuracy, robustness, and cybersecurity requirements (Article 15). They wrap those controls in a quality-management system, run a conformity assessment, affix the CE marking, and register the system in an EU database before it goes live.
Deployers have their own duties under Article 26: use the system according to instructions, keep competent people in the oversight loop, monitor operation, retain logs, and inform affected people. Public authorities must also complete a fundamental-rights impact assessment.
General-purpose AI models carry a separate track. Their providers owe documentation and copyright-compliance duties, and a model is presumed to pose systemic risk once its training compute passes 10^25 FLOP, which triggers evaluation and incident-reporting obligations. The timeline is staged: prohibitions applied from February 2025, GPAI duties from August 2025, high-risk application rules from August 2026, and high-risk product rules from August 2027. Standards such as ISO/IEC 42001 and the NIST AI Risk Management Framework give organisations a repeatable way to operationalise these duties.
From definition to governance: building an AI system inventory
Every obligation above assumes one thing: that you know which AI systems your organisation runs. Most do not. Models arrive through procurement, embedded vendor features, and tools staff adopt on their own, the pattern known as shadow AI. You cannot classify or govern a system you have not recorded.
A practical AI system inventory captures, for each system, its purpose, the data it uses, its provider or deployer role, and its risk tier. That record becomes the backbone of an AI governance framework: it drives which controls apply, where evidence is needed, and which systems require a conformity assessment. This is the work AI Sigil is built for. The platform maintains a live registry of AI systems and their components, classifies each one against the frameworks you are subject to, and attaches the controls and evidence that turn a definition into a defensible compliance position. Treating AI system governance as a continuous inventory, rather than a one-off audit, is what keeps a programme current as models and rules change.
Frequently asked questions
What are the main types of AI systems? AI systems are grouped two ways. By capability they are narrow (task-specific, the only kind in real use), general (human-level, still theoretical), or super (beyond human, hypothetical). By functionality they are reactive machines, limited memory, theory of mind, or self-aware. Nearly all systems today are narrow, limited-memory AI.
What are examples of AI systems? Common examples include large language models and chatbots, recommendation engines on streaming and shopping sites, fraud-detection and credit-scoring models, medical image analysis, spam filters, voice assistants, and the perception systems in self-driving vehicles. Each infers an output from input data rather than following a fully hand-written script.
Is all software an AI system? No. Software that only executes rules a person wrote out in full, simple heuristics, or basic data processing is not an AI system under the EU AI Act, because it does not infer. The distinguishing feature is inference: an AI system works out how to generate its output rather than following predetermined instructions.
What makes an AI system high-risk under the EU AI Act? A system is high-risk if it is a safety component of a product already regulated under EU law, or if it operates in one of the eight Annex III areas, such as employment, credit and insurance, biometrics, critical infrastructure, education, law enforcement, migration, or justice. High-risk systems carry the fullest set of obligations.
Who has to comply with AI system rules? Obligations follow roles. Providers, who develop and place systems on the market, carry the strictest duties. Deployers, who use systems under their own authority, carry a lighter set. A deployer can become a provider under Article 25 by rebranding or substantially modifying a high-risk system.
How many types of AI are there? It depends on the axis. Capability gives three types (narrow, general, super). Functionality gives four (reactive, limited memory, theory of mind, self-aware). The EU AI Act adds a governance axis of four risk tiers (prohibited, high-risk, limited, minimal).
Conclusion
An AI system is more than a piece of clever software; under EU law it is a defined category with consequences. The Article 3(1) definition, the seven characteristics, and the risk tiers together decide whether a system is regulated and what its owners must do. The taxonomies of capability and functionality help you describe your tools, but the governance question, which tier and which role, is what carries obligations. Organisations that keep a live inventory of their AI systems, classified and controlled, turn a legal definition into an operating advantage. Start by mapping what you run, then bring each system under an AI governance programme that keeps pace with the rules.