The European Union has introduced the EU Artificial Intelligence Act, a comprehensive law that regulates the use of AI systems in the EU. In this series, ‘Decoding AI: The European Union’s Take on Artificial Intelligence’, we break down everything you need to know about the law for you.

Before we dive in, make sure to get your free copy of Palqee’s EU AI Act Framework. The first of its kind with a useful guide on AI compliance. You can download it here!

While the final text is still pending confirmation and the definition of an AI system is expected to be refined, it is important to note that the current understanding provides a useful overview of the systems covered by the EU AI Act.

In general, any technology that extensively processes data to make decisions or draw conclusions is likely to fall within the scope of the regulation. This broad definition ensures that a wide range of data-driven technologies can be effectively addressed by the AI Act.

According to the EU AI Act, an AI system is a software developed using the following techniques:

A. Machine learning: It involves training software to learn from data and make predictions or decisions without explicit programming. The Act recognises three main types of machine learning:

• Supervised learning: The AI system learns from labelled examples to make predictions or decisions based on new data, e.g., train an AI system to recognise fruits by providing labelled images for it to learn from. It can then predict the fruit type of new images based on its previous training.

• Unsupervised learning: The AI system learns patterns and structures from unlabelled data without predefined outcomes, e.g., analyse a dataset of customer purchase history without labels to discover patterns and groupings, revealing valuable insights without specific guidance.

• Reinforcement learning: The AI system learns through interactions with an environment, receiving feedback or rewards for its actions, e.g., teach an AI system to play a video game through interactions with the game environment, receiving rewards for successful actions and learning to improve gameplay over time.

• Deep learning: involves complex neural networks capable of processing vast amounts of data, e.g., in image recognition, a deep learning model can learn to identify objects by training on labelled images, enabling tasks like facial recognition and object detection.

B. Logic- and knowledge-based approaches: it relies on representing and reasoning with knowledge to make decisions. The Act highlights several techniques, including:

• Knowledge representation: AI stores and organises knowledge to facilitate reasoning and decision-making., e.g., an AI system stores medical knowledge and uses it to assist in diagnosing medical conditions, providing recommendations based on patient symptoms and medical history.

• Inductive (logic) programming: AI systems derive general rules or patterns from specific examples, e.g., by analysing customer preferences and purchase history, an AI system can identify patterns and make predictions, recommending complementary products to customers based on their previous purchases.

• Inference and deductive engines: AI systems use logical reasoning to draw conclusions from available information, e.g., an AI system assists with legal research by using logical reasoning and inference to draw conclusions from legal information, helping lawyers analyse cases and build stronger arguments.

• Symbolic reasoning and expert systems: AI systems emulate human expertise by employing symbolic representations and rule-based reasoning, e.g., an AI system provides personalised financial advice by incorporating the knowledge of financial experts into a rule-based system, using symbolic representations and rule-based reasoning to assess a person's financial situation and recommend investment options.

C. Statistical approaches: it involves analysing and interpreting data to make predictions or decisions. The Act mentions specific methods, such as:

• Bayesian estimation: AI systems use Bayesian statistics to update beliefs or probabilities based on new evidence, e.g., an AI system predicts customer purchase probability by updating its belief using Bayesian statistics, incorporating new data to refine predictions.

• Search and optimisation methods: AI systems employ algorithms to find the best solutions or optimise parameters, e.g., an AI system optimises delivery routes using algorithms like genetic algorithms or simulated annealing to find the most efficient solutions that minimise time, distance, or cost.

The EU AI Act requires compliance from all AI systems. The regulation includes comprehensive and future-proof regulations, such as principle-based requirements for an AI system to comply with. Simultaneously, it introduces a clear and risk-based regulatory framework specifically tailored for what it deems as "high-risk AI systems".

These high-risk AI systems must meet mandatory requirements for trustworthy AI and undergo conformity assessments before entering the Union market. Providers and users of these systems have clear and proportionate obligations to ensure safety and compliance with existing laws protecting fundamental rights throughout the AI systems' lifespan. These restrictions are proportionate and limited to preventing serious safety risks and infringements of fundamental rights.

For AI systems that are not classified as high-risk, the proposal suggests, for example, minimum transparency obligations, particularly for AI systems like chatbots or 'deep fakes'.