I. BACKGROUND
On August 1, 2024, the European Council and the Council of the European Union officially brought the EU AI Act (“AI Act”) into force. As of February 2, 2025, certain general provisions under Articles 1 to 4 of the AI Act (including the AI literacy requirement under Article 4) and the prohibitions on certain AI activities under Article 5 began to apply in practice.
Pursuant to Article 96(1) of the AI Act, the European Commission (the “Commission”) is required to issue guidelines for implementing the Act. Accordingly, on February 4 and 6, 2025, respectively, the Commission published: (i) the Guidelines on Prohibited AI Practices under the AI Act, which address harmful manipulation and deception, exploitation of vulnerabilities, and social scoring; and (ii) the Guidelines on the Definition of AI System.
In addition, the Commission released a living repository of AI literacy practices, an online resource platform that collects and shares submission from AI system providers as well as companies and organizations that use AI. The platform aims to enhance AI users’ understanding and competence, facilitate the exchange of experiences and best practices across organizations, and support more responsible and effective use of AI technology.
While these guidelines are not legally binding, may be subject to future revisions, and could ultimately be interpreted differently by the Court of Justice of the European Union (CJEU), they are expected to serve as an important interpretive reference for key provisions of the AI Act and to enhance legal clarity and predictability in practice.
II. GUIDELINES ON PROHIBITED AI PRACTICES UNDER THE AI ACT
The Guidelines, published on February 4, 2025, provide interpretive guidance on the application of eight prohibited AI practices defined in Article 5 of the AI Act, each of which is deemed to pose an inherent risk of infringing upon EU Fundamental Rights. The prohibited practices include: (i) harmful manipulation and deception; (ii) exploitation of vulnerabilities; (iii) social scoring; (iv) risk assessment and prediction of criminal offenses by individuals; (v) untargeted scraping of data to build facial recognition databases; (vi) emotion recognition; (vii) biometric categorization; and (viii) real-time remote biometric identification.
With the exception of real-time remote biometric identification, the prohibitions apply to AI systems at the stage of being placed on the market, put into service, or used within the EU. Real-time remote biometric identification, in particular, is prohibited in principle for public surveillance purposes and permitted only in narrowly defined circumstances (such as countering terrorist threats or investigation certain serious crimes) subject to prior judicial authorization and strict proportionality requirements.
The Guidelines also clarify several areas that fall outside the scope of the AI Act. These include activities related to national security, defense and military purposes, judicial and law enforcement cooperation with third countries, research and development, and personal, non-professional use.
The prohibitions on harmful manipulation and exploitation of vulnerabilities are addressed in particular detail. For the prohibition to apply, all three of the following conditions must be satisfied: (i) the AI system must employ subliminal techniques; (ii) the technique must have the purpose or effect of materially distorting the behavior of an individual or group; and (iii) the distorted behavior must cause, or be likely to cause, significant harm to the subject or a third party.
The remaining six prohibitions (social scoring, crime prediction, indiscriminate data collection, emotion recognition, biometric-based classification, and real-time remote biometric identification) do not require a step-by-step assessment of individual elements. Instead, these practices are prohibited outright once an AI system is used for such purposes or performs such functions. In this regard, the Guidelines interpret the wording of Article 5(1)(c) of the AI Act broadly, emphasizing that the prohibited conduct need not result in a formal assessment or decision, may occur across various contexts (including public, private, and industrial settings), and may involve personal data that is inferred or predictive in nature, as well as data reflecting explicit characteristics.
The Guidelines represent the Commission’s first official interpretive document on the AI Act and provide practical guidance on its implementation and interpretation. However, they also make clear that compliance with the AI Act does not exempt organizations from obligations under other applicable legal frameworks, including the General Data Protection Regulation (GDPR) and consumer protection laws.
III. LIVING REPOSITORY OF AI LITERACY PRACTICES
The Commission has published a list of AI literacy-related practices (the “Repository”) to promote learning and information exchange among providers and users (deployers) of AI systems. The Repository is continuously updated, with the latest version public available as of April 16, 2025.
The Repository is organized into three tiers reflecting levels of adoption: “fully implemented,” “partially rolled-out,” and “planned.” Each tier includes real-world examples and detailed information about the approaches adopted by participating organizations. The Repository also includes a disclaimer clarifying that implementing any initiative listed therein does not, by itself, guarantee compliance with the requirements set forth in Chapter 4 of the AI Act.
For instance, AI & Partners B.V., a user (distributor) of AI systems, is cited as an example of “full adoption” of essential AI literacy practices. The company established an internal framework featuring flexible training methods, such as self-paced modules, instructor-led workshops, and scenario-based exercises, that are accessible across roles and backgrounds throughout the organization. The Repository emphasizes that sector-specific adaptation, along with organization-wide collaboration and integration, are critical to successful AI literacy implementation.
For organizations classified as having only “partially rolled-out” AI literacy initiatives, the report identifies several key challenges, including limited understanding of AI governance and responsible implementation, knowledge silos within the organization, and the absence of a unified execution framework.
IV. GUIDELINES ON THE DEFINITION OF AI SYSTEMS
Pursuant to Article 3(1) of the AI Act, the Act applies only to systems that meet the definition of an “artificial intelligence systems” (AI system). On February 6, 2025, the European Commission published guidance to further clarify the scope of this definition, introducing a step-based framework for determining whether a particular system falls within the AI Act’s scope. The guidance also clarifies that the seven elements comprising the definition of an AI system need not be present simultaneously throughout the system’s lifecycle.
(1) Machine-based:
All AI systems must be machine-based, with “machine” broadly defined to include both software (e.g., computer code) and physical hardware (e.g., processing units). Emerging computational systems, such as quantum computing, are also included if they provide mechanical computational power.
(2) Autonomy:
AI systems must be capable of operating with varying degrees of independence. The key criterion is the system’s ability to generate predictions, decisions, recommendations, or other inferences that can affect the physical or virtual environment based on input. Fully manual systems are excluded from the definition. The Guidelines emphasize that the lower the degree of human intervention, the higher the associated risk, potentially warranting increased oversight.
(3) Adaptiveness:
Adaptability refers to a system’s capacity to learn or modify its behavior after deployment. However, the Guidelines clarify that the use of the term “may” in the AI Act indicates that such capabilities are optional and not mandatory for inclusion within the AI system definition.
(4) Objectives:
An AI system must function to achieve one or more explicit or implicit objectives. Explicit objectives are those deliberately programmed by developers (e.g., minimizing a cost function or maximizing a reward), while implicit objectives are inferred from the system’s behavior and underlying assumptions. The Guidelines distinguish this from the “intended purpose,” which relates to the task the system is designed to perform.
(5) Inferencing:
AI systems must be capable of inferring how to produce outputs from inputs, distinguishing them from traditional software in which outputs are generated strictly through predefined, human-coded rules. Inferring encompasses not only the generation of outputs from inputs but also the use of AI techniques during development to enable that functionality.
(6) Outputs:
AI systems must be able to produce outputs based on inputs, including predictions, content generation, recommendations, decisions, and similar functions.
(7) Influencing Environments:
The outputs of an AI system must materially affect the physical or virtual environment. This element underscores that AI systems are not passive processing tools but active systems that interact with and influence their external context.
V. IMPLICATIONS
As noted above, the Guidelines and the Repository serve as valuable resources, offering practical guidance on several concepts that were previously ambiguous or subject to differing interpretations. While it remains to be seen whether national supervisory authorities will adopt the Commission’s interpretation of these key concepts, it is notable that the Commission has articulated certain exceptions from the scope of the AI Act, signaling an effort to strike a balanced regulatory approach that safeguards fundamental rights without unduly stifling technological innovation.
Bae, Kim & Lee LLC’s AI Practice continues to closely monitor both domestic and international AI regulatory developments and will remain engaged in reviewing and analyzing relevant guidelines and interpretations to help minimize legal and institutional risks associated with AI development and deployment, while actively supporting sustainable and compliant business operations.
[Korean version]
