To understand artificial intelligence (AI), it is necessary to distinguish its three key foundational layers: how it learns, what it actually produces, and who bears responsibility for its decisions.

The challenge of artificial intelligence today is not only how it works, but also how we understand it. Discussions about AI are often conducted between two extremes. On the one hand, there is technological enthusiasm that presents AI as a system that “understands,” “reasons,” and “makes decisions.” On the other hand, there is skepticism that reduces such systems to ordinary computational tools. In reality, understanding artificial intelligence requires a clearer distinction between the levels at which it emerges and operates.

One useful conceptual framework is based on distinguishing three layers of AI: learning, the model, and the regulatory framework. These three layers do not represent a complete technical description of artificial intelligence, but rather serve as an analytical model that helps separate its technological origin, its mode of operation, and its societal control.

Learning: where AI “comes from”

AI emerges through a process of learning (training) on vast amounts of data, including texts, images, records, and signals. During training, the system repeatedly predicts outcomes, measures its errors, and adjusts its internal parameters to reduce those errors over time. The essence of this layer is simple: the quality of AI depends on the quality of what it learns from. If the data are incomplete, biased, or poorly selected, the model will carry those biases and inaccuracies forward. For this reason, “learning” is both the technical foundation of AI and the most common source of later problems in practice: not because the system is “malicious,” but because it learned from poor or unbalanced examples.

Model/representation: what AI actually produces

The second layer is the most important for properly understanding AI: AI does not deliver truth, but rather a representation (a model-based estimate) that resembles it. For this reason, describing this process as a “representation” is appropriate, since AI does not “understand reality” in the human sense; instead, it forms a mathematical model that captures patterns and uses them to generate a response, estimate, or classification. Although highly convincing, it remains an approximation: AI often sounds confident even when it is wrong, because it lacks an internal mechanism of “understanding”, unlike human beings. AI mimics intelligence—it behaves as if it understands, whereas in reality, it produces model-generated outputs.

Framework/protection: control, limitations, and responsibility

The third layer concerns rules, limitations, and responsibility. Since AI produces representations rather than final truths, without understanding the consequences of its outputs, society must establish protective mechanisms, including legal norms, ethical standards, internal organisational policies, human oversight, verification procedures, and clear lines of accountability. The key point at this level is that AI itself cannot bear responsibility. Responsibility remains with the people and organisations that develop, deploy, and use these systems. Contemporary regulatory approaches are built precisely on this premise: the European AI Act introduces obligations related to risk management, transparency, and human oversight over AI systems, while the GDPR establishes principles of lawful data processing, data minimisation, and accountability when personal data is processed by AI systems.

Without this layer, AI can cause harm, and excessive control may stifle innovation. For this reason, the question of the regulatory framework is always a question of balance between benefits and risks.

Viewed through these three layers — learning, model, and regulatory framework — artificial intelligence ceases to be an obscure technological phenomenon and becomes a system we are able to understand, with risks we can assess, and a use we can regulate and control.

Understanding how AI models function is becoming particularly important at a time when such systems are being introduced into everyday business processes, including areas related to intellectual property rights. Artificial intelligence is already used in document analysis, prior-art research, data processing, and the preparation of various professional materials, highlighting the importance of understanding how these systems arrive at their results.

Although Serbia has not yet adopted an AI law, the European regulatory framework – primarily the AI Act and existing data protection rules – already influences how these systems are developed and used. For companies and professionals working with innovation, patents, copyright, or data, it is essential to understand the basic principles of how AI systems operate – not only their capabilities, but also their limitations. Such understanding supports responsible use of technology, more accurate assessment of legal risks, and effective integration of AI tools into professional work.