EU AI Act AI System Definition: The Official 7-Element Test

The EU AI Act only applies to systems that meet the legal definition of an “AI system” in Article 3(1) of Regulation (EU) 2024/1689. Before anyone can discuss risk categories, prohibited practices, or compliance obligations, one gateway question comes first: does the tool in front of you qualify. This guide explains the European Commission’s official 7-element test. The Commission published the test as Guidelines on the definition of an AI system, reference C(2025) 5053 final, dated 29 July 2025. This guide turns the test into a checklist you can run against your own tools.

Scope note: These Guidelines are non-binding Commission guidance. Only the Court of Justice of the European Union (CJEU) can give an authoritative interpretation of the AI Act. This guide reflects the Commission’s published position as of July 2025.

What you need before starting

• A specific tool, model, or process in mind. The test applies system by system, not to “AI” in the abstract.
• Basic knowledge of how the system produces its output (rule-based, statistical, or machine-learning-based)
• No legal background. The Commission designed the test for providers and deployers, not only for lawyers.

Why classification comes before compliance

Article 96(1)(f) of the AI Act required the Commission to issue guidance on the Article 3(1) definition. This definition decides whether the Act applies at all. The definition entered into application on 2 February 2025, alongside the prohibited-practices rules in Article 5. Skipping this step leads to two costly mistakes: treating a rule-based spreadsheet as if it needs AI Act compliance work, or assuming a machine-learning tool is exempt because it “just does statistics.”

The Commission states clearly that there is no shortcut here: “the definition of an AI system should not be applied mechanically; each system must be assessed based on its specific characteristics.” No exhaustive list of qualifying or non-qualifying systems exists. The seven elements below are the test.

The EU AI Act AI system definition

Article 3(1) AI Act defines an AI system as:

“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”

The Commission’s Guidelines break this definition into seven elements. Assess a system against all seven elements across its lifecycle: the “building” phase and the “use” phase.

Step 1: Is it machine-based?

This element sets the lowest bar. “Machine-based” covers both hardware (processing units, memory, storage, input/output interfaces) and software (code, instructions, programs, operating systems). The Guidelines confirm that this term also covers unconventional computing. Quantum computing systems and biological or organic systems both qualify, “so long as they provide computational capacity.”

In practice: almost every digital tool passes this element. It rarely disqualifies a system on its own.

Step 2: Does it operate with some degree of autonomy?

The Act requires the system to be “designed to operate with varying levels of autonomy.” This phrase means “some degree of independence of actions from human involvement.” The Guidelines set the bar deliberately low. A system that generates an output “without this output being manually controlled, or explicitly and exactly specified by a human” already has some degree of independence of action.

What fails this test: a system designed to operate only through full manual human control, with no capacity to generate an output on its own.

What passes: an expert system that takes human-provided input and produces a recommendation on its own, even when a human supervises or approves the output afterward.

Step 3: Might it adapt after deployment?

This element is optional. “Adaptiveness” means self-learning capability that changes the system’s behavior during use, so the same input can produce a different output over time. The Guidelines use the word “may” for a deliberate reason: a system does not need self-learning capability to qualify as an AI system. This element is facultative, not decisive. Do not rule out a system only because it stays static after deployment.

Step 4: Does it pursue explicit or implicit objectives?

AI systems operate toward one or more objectives. An objective can be explicit (directly encoded, for example an optimization target, a probability, or a reward function) or implicit (deduced from training data or from the system’s interaction with its environment, without a stated goal).

The Guidelines draw a useful distinction here: a system’s internal objectives are not the same as its intended purpose. Objectives describe what the system’s tasks are optimized to achieve. Intended purpose is the externally oriented context of deployment, as defined in Article 3(12) AI Act: the use the provider designed the system for.

Step 5: Does it infer how to generate outputs?

This element is the decisive one. Recital 12 AI Act states that “a key characteristic of AI systems is their capability to infer.” The recital distinguishes AI systems from “simpler traditional software systems or programming approaches” based on “rules defined solely by natural persons to automatically execute operations.”

The Guidelines identify two families of technique that enable this inference:

• Machine learning approaches: supervised learning (for example, spam detection trained on labeled emails), unsupervised learning (for example, clustering chemical compounds for drug discovery), self-supervised learning (for example, language models that predict the next token), and reinforcement learning (for example, a robot arm that learns to grasp objects through trial and error)
• Logic- and knowledge-based approaches: systems that reason from encoded expert knowledge, rules, or symbolic representations, using deductive or inductive engines, rather than learning from data

If a system does neither, and only executes rules defined solely by a human without deriving models or drawing inferences, it fails this element regardless of how sophisticated it looks.

Step 6: Does it produce predictions, content, recommendations, or decisions?

Article 3(1) lists four output categories. Each category involves a different degree of human involvement:

• Predictions: an estimate of an unknown value from known inputs. This category requires the least human involvement.
• Content: new material the system generates, such as text, images, video, or music (the output type most associated with generative AI, built on Generative Pre-trained Transformer-style technologies)
• Recommendations: suggestions for actions, products, or services. If a system applies a recommendation automatically without human review, the Guidelines note that it effectively becomes a decision.
• Decisions: conclusions the system makes with no human intervention. This category automates a process that human judgment traditionally handled.

Step 7: Can those outputs influence a physical or virtual environment?

The final element confirms that the system is not passive. Outputs must be able to influence a physical environment (for example, a robot arm) or a virtual environment (digital spaces, data flows, software ecosystems). A system that generates an output with no capacity to act on or within any environment does not complete the definition.

What falls outside the definition

The Guidelines are equally specific about what does not qualify, even when a system technically has some capacity to infer. Four categories are explicitly excluded.

Systems for improving mathematical optimization. These are tools that accelerate or approximate established optimization methods (for example, linear or logistic regression) without going beyond “basic data processing.” The Guidelines cite physics-based simulations that use machine learning to speed up computation, and satellite bandwidth allocation systems that use machine learning to match the performance of traditional optimization methods, as outside scope.

Basic data processing. These are systems that follow predefined, explicit human-programmed rules with no “learning, reasoning or modeling” at any stage: database queries, standard spreadsheet formulas, or software that calculates a population average from survey data.

Classical heuristics. This category covers rule-based, pattern-recognition, or trial-and-error problem-solving that does not involve data-driven learning. The Guidelines give the example of a chess program that uses a minimax algorithm with heuristic evaluation functions.

Simple prediction systems. These are machine-based systems “whose performance can be achieved via a basic statistical learning rule.” Examples include a stock-price forecast that always predicts the historical average, or a demand forecast based on a static mean. Even though these systems technically use a machine learning approach, the Guidelines exclude them “due to its performance.”

The common thread across all four exclusions: the system does not go beyond basic data processing, and its self-adjustments (if any) optimize computational efficiency rather than the decision-making model itself.

Decision checklist

Run your tool through these seven questions. Elements 2 and 5 are most likely to disqualify a system. Element 3 never will, since it is optional.

1. Machine-based? Does it run on hardware and software? (Almost always yes.)
2. Autonomous? Can it generate an output without a human who manually and exactly specifies that output?
3. Adaptive? Does its behavior change after deployment? (Optional. A “no” here does not disqualify the system.)
4. Objective-driven? Does it pursue an explicit or implicit goal, distinct from its deployment context?
5. Inference-capable? Does it use machine learning or logic/knowledge-based techniques, rather than purely human-defined rules?
6. Output type? Does it produce a prediction, content, a recommendation, or a decision?
7. Environmental influence? Can that output act on a physical or virtual environment?

If Step 5 fails, and the system runs purely on rules a human wrote and executes them mechanically, it is not an AI system under the Act, regardless of how the other six elements score.

Troubleshooting

“It is just a spreadsheet with formulas.” This is a correct call to exclude it. The Guidelines name standard spreadsheet software explicitly as basic data processing, provided it has no AI-enabled functionality layered on top.

“It is a forecasting tool, so it must be in scope.” Not necessarily. The Guidelines exclude a tool that always predicts the historical average or a static mean as a simple prediction system, even though it is technically a statistical model. The exclusion turns on performance, not on whether “machine learning” appears in the product description.

“It only recommends, a human always approves it.” This is still likely an AI system if it infers the recommendation using machine learning or logic-based reasoning. Human review afterward does not remove the system from scope, though it is relevant to how the system uses the recommendation downstream.

“It is a rules engine built by our own experts.” This depends entirely on Step 5. A system that reasons from encoded expert knowledge using deductive or inductive inference (a logic- and knowledge-based approach) can qualify as an AI system. A system that mechanically executes fixed rules a human wrote, with no inference step, does not qualify.

What you can do now

You can now run a specific tool, model, or internal process through the seven-element test. This lets you reach a defensible position on whether the AI Act’s definition applies to it at all. This is the necessary first step before you assess risk category, transparency duties, or any other obligation under the Act.

This checklist resolves three situations quickly:

• A lab or team wants to know if its internal forecasting tool needs AI Act review before a wider rollout
• A consultant needs to document the classification reasoning behind a client’s tool inventory
• A research officer is triaging which of the institute’s software tools need a full AI Act risk assessment

Related reading on The Science Talk

• EU AI Act Article 4: What AI Literacy Requirements Mean for European Research Institutes: covers the parallel Article 4 AI-literacy provision, which the July 2026 Digital Omnibus simplified at company level. Note that the piece describes the original, stricter wording.

Want more guides like this? Browse all AI Guides or get in touch →