AI GovernanceApril 15, 2025

What exactly is an AI system under the EU AI Act?

This is the starting point for anyone working on AI compliance in Europe. Why? Because it defines whether your software falls under the scope of the regulation.

Posted by

Ana Belen Barbero Castejon

What are the key elements?

As you might have noticed, the definition is intentionally broad. The reason is that Europe needs flexibility to keep up with rapid technological developments in the field. Thus, each system must be assessed based on its specific characteristics.

The definition includes seven key elements:

The 7 Elements of an AI System

EU AI Act Definition

1. Machine-based

It carry out computations

2. Autonomy

Independent operation capability

3. Adaptiveness

Learning from experience

4. Objectives

Explicit/implicit goal-driven

5. Inferencing

It can generate outputs from inputs

6. Outputs

Predictions, content, recommendations, or decisions

7. Impact

Influences environments

Let's shed some light on what they actually mean.

What makes a system machine-based in the context of AI?

An AI system is considered machine-based when it is developed with and operates on machines that include both hardware and/or software components. In other words, all core functions of the AI like model training, data processing, predictive modeling and automated decision-making at scale depend on computational processes.

The term machine-based doesn't just refer to traditional computers. It also includes new types of machines like quantum computers or even biological systems, as long as they can process information and perform calculations like a computer. What matters most is that the system can carry out computations, no matter how it's built.

What does mean that AI systems are designed to operate with varying levels of autonomy?

AI systems can work with different levels of independence from human control. This depends on three main factors:

1. Inference capability

Autonomy and inference are connected. AI systems must be able to make decisions, predictions, or recommendations on their own. This decision-making ability is key to their autonomy.

2. Human intervention or involvement

Human involvement and intervention can be direct (e.g., manual control) or indirect (e.g., monitoring automated processes). The EU AI Act excludes systems that are designed to operate with full manual human involvement and intervention.

3. Interaction with the environment

Autonomy also depends on how an AI system interacts with its environment. It's not just about the technology (e.g., machine learning) but how it works independently in real-world situations.

One example of autonomy could be the Netflix or Amazon recommendation engine. These systems require human input, like your viewing history. However, they autonomously generate recommendations based on this data, without needing any direct action from the user. The system analyzes your data and suggests content that you didn't specifically request.

What does it mean for an AI system to be adaptive?

Adaptiveness refers to an AI system's ability to learn and adjust its behavior over time. This means the system may change how it operates based on new data, and as a result, it could produce different outcomes for the same inputs after adapting.

Adaptiveness in action

Initial state

System starts with base knowledge and initial parameters

Accuracy: 85%

Learning phase

Processes new data and feedback to improve performance

Accuracy: 92%

Adapted state

Refined behavior based on learned patterns

Accuracy: 97%

However, it's important to note that not all AI systems need to be adaptive or self-learning after deployment. A system can still be considered an AI system even if it doesn't have the ability to learn or change its behavior automatically after it's been set up. The ability to learn or identify new patterns in data is optional and not required for something to qualify as an AI system.

What are AI system objectives, and how do they work?

AI systems are designed to achieve specific goals, known as objectives. These objectives can be either explicit or implicit:

Explicit objectives

These are clearly defined goals directly programmed by the developer, such as optimizing a cost function, probability, or reward. For example, an AI system might be tasked with minimizing errors or maximizing efficiency.

Implicit objectives

These goals are not directly stated but can be inferred from the system's behavior or the assumptions built into it. These objectives may emerge from how the AI interacts with data or its environment.

Important note: The AI Act (Recital 12) clarifies that the objectives of an AI system might not always align with its intended purpose. While the system's objectives are focused on internal tasks (e.g., accuracy, performance), its intended purpose relates to the external context—such as the specific role it's meant to play in a business setting.

Example: A corporate virtual assistant AI system might aim to answer questions accurately and with few errors (its objectives), but its intended purpose could be to support a specific department's work by following certain guidelines (e.g., document formatting) and integrating into broader workflows.

What does it mean that an AI system must be able to "infer how to generate outputs"?

A core feature of an AI system is its ability to infer, that is, to determine how to produce an output based on the input it receives. This is what sets AI apart from traditional software, which follows only fixed, human-defined rules.

According to Recital 12 of the AI Act, inference refers to the process by which an AI system produces predictions, content, recommendations, or decisions that can influence physical or virtual environments. It can also involve the system deriving models or algorithms from data. It cam happen during the building phase of the AI system or during its use.

To understand how an AI system "infers how to generate outputs," we must look at the techniques used during its building phase. These techniques define whether a system qualifies as an AI system under the AI Act:

Machine Learning (ML) approaches

Supervised Learning

Learns from labeled datasets (input-output pairs)

Email spam filters
Legal doc classification
Credit scoring

Unsupervised Learning

Discovers patterns in unlabeled data

Customer segmentation
Legal document clustering
Anomaly detection
Fraud rings detection

Self-Supervised Learning

Generates its own training labels from raw data

LLM predicting next word
Image models predicting next pixel
Autoencoders

Reinforcement Learning

Learns through trial, error and reward feedback

LLMs post-training alignment
Recomendation algorithms
Policy gradient for robotics

Deep Learning

Uses neural networks to learn directly from raw data

GenAI
Voice transcription
Facial recognition
Multimodal review

Logic & knowledge-based approaches

Applies encoded rules and expert logic — no learning involved.

Code linter
Expert systems in healthcare (e.g., clinical decision support)
Events and alert trigger rules

What falls outside the scope of an AI System?

Traditional Software

Systems based solely on human-defined rules for automated operations. No inference, learning, or modelling involved.

Mathematical Optimization

Systems using ML to accelerate traditional methods without autonomous decision-making.

ML-accelerated physics simulations
Satellite bandwidth management
Function approximation models

Basic Data Processing

Systems using explicit, fixed rules without AI techniques.

Spreadsheets & SQL queries
Simple statistics software
Opinion poll software

Classical Heuristic Systems

Rule-driven systems without dynamic learning capabilities.

Chess programs with minimax
Trial-and-error systems
Pattern recognition systems

Simple Prediction Systems

Basic statistical systems serving as baselines for ML.

Stock price averaging
Sales demand estimation
Support ticket resolution prediction

These systems enhance efficiency but lack the key characteristics of AI systems: they don't make autonomous decisions, learn patterns beyond narrow tasks, or exhibit the adaptability and generalization typical of modern AI

In what ways can AI-generated outputs impact physical or digital environments?

For a system to qualify as an AI system, it must infer how to generate outputs that influence environments.

Type of output	Description	Example
Predictions	Estimates of unknown values based on known inputs	Forecasting traffic in self-driving cars Predicting customer churn risk
Content	Creation of new material (text, images, video, etc.) by the system	Generative AI like GPT Image/video synthesis
Recommendations	Suggestions based on user data, behaviour, or context	Recommending products Suggesting actions Potential hire recommendations
Decisions	Autonomous conclusions or actions taken by the system	Credit scoring automation Autonomous driving decisions

How do AI system outputs influence environments?

Outputs of AI system can influence both physical and virtual environments. This means that AI systems are not passive; they actively impact the environments where they are used.

In physical environments, AI systems can directly interact with and manipulate real-world objects. For instance, AI-powered robotic arms in manufacturing plants can precisely assemble products or handle materials.

In virtual environments, AI systems can transform digital spaces by processing data streams, automating software operations, and modifying virtual systems. This includes everything from content recommendation algorithms to automated trading systems.

Ready to explore AI safety and governance in depth? Join the Contrasto AI Club, a community of professionals dedicated to advancing responsible AI development and implementation.