Under Article 3(1), an AI system is a machine-based system designed to operate with varying levels of autonomy that, for stated or implicit goals, infers from its inputs how to generate outputs that can influence physical or virtual environments. The key word is infers. Plain rule-based software is not the target.
Reading the definition piece by piece
The text breaks into parts you can test against your own software:
- Machine-based: it runs on hardware and software.
- Varying levels of autonomy: it works with some independence from human involvement.
- May adapt after deployment: some systems keep learning, though this is not required.
- Infers from input how to produce outputs: predictions, content, recommendations or decisions.
- Outputs can affect physical or virtual environments.
Where the line usually falls
A fixed spreadsheet formula, a hard-coded if-then rule engine or a sort function does not infer anything. A trained classifier that scores loans, a recommender, a chatbot built on a language model or a vision model that flags defects all derive their outputs from data patterns, so they sit on the AI side.
Edge cases exist. A rules engine that was partly tuned by a learning step, or a statistical model that is simple but still learns parameters from data, needs a careful read. The Commission has published guidance on the definition, but we did not fetch it for this post, so check it for borderline systems.
Related definitions you will need
- Provider: whoever builds an AI system or general-purpose AI model, or has one built, and places it on the market under its own name or trademark, for payment or free.
- Deployer: an entity using an AI system under its authority, except for personal non-professional use.
- General-purpose AI model: a model showing significant generality that can competently perform a wide range of distinct tasks and be integrated into many downstream systems.
- General-purpose AI system: a system based on such a model that can serve a variety of purposes.
A model and a system are different things. A model provider has one set of duties. The company that wraps that model in a product may be the provider of the system and carry another.
Why the definition matters first
Every later question depends on it. If a tool is not an AI system, the risk tiers, literacy duty and transparency rules do not attach to it. If it is, you move on to classify use and role. Document the reasoning for borderline tools, including why you concluded they do or do not meet the definition.
If you are unsure about a borderline tool, write down the reasoning anyway and date it. A regulator or customer will care less about whether you were perfectly right than whether you looked carefully and recorded why.
Quick check
Is a keyword filter an AI system?
Usually not, because it applies fixed rules and does not infer outputs from data. A filter that uses a trained model to score content probably is.
Is a language model the same as an AI system?
No. The Act separates general-purpose AI models from AI systems built on them, and the duties differ.
Sources
- AI Act Article 3, definitions (an unofficial reader; confirm wording against the Official Journal)
Last reviewed: 19 October 2026.
