Nova pravila za umjetnu inteligenciju: vodeni znakovi za umjetnu inteligenciju

New rules for artificial intelligence have come into effect in Europe, requiring AI-generated content to be marked with watermarks so they can be easily identified. These regulations, part of the European Union’s AI Act, were initially set to take effect earlier but have been delayed as part of broader economic reforms. The new measures aim to increase transparency and hold providers accountable for their AI systems. The AI Act was approved two years ago, but its implementation has proceeded gradually. On August 2, several key provisions came into force, including requirements for AI-generated content to be clearly identifiable, either through visual markers such as watermarks or other technical means. Users must also be informed when interacting with AI, such as through chatbots or customer service hotlines. These changes mark a step toward greater oversight of AI technologies within the EU. Among the delayed provisions are stricter rules for high-risk AI systems, which will take effect later this year. The European Commission postponed these rules as part of a larger package aimed at supporting businesses while maintaining regulatory standards. This shift reflects ongoing efforts to balance innovation with safety and ethical considerations. The new rules empower the European Commission to impose fines and take action against companies that fail to comply with AI regulations. This includes “general-purpose AI models,” such as widely used chatbots like ChatGPT, Claude, and Gemini. These models are subject to specific obligations under the AI Act, and non-compliance could lead to penalties. Professor Philipp Hacker, who specializes in law and ethics of digital society, emphasized the significance of these developments. He noted that many powerful AI models currently perform poorly in security tests, raising concerns about their reliability and potential risks. Hacker described the Commission’s ability to investigate, request documents, conduct assessments, and enforce penalties as a “sharp sword.” However, he warned that such actions would require extensive legal procedures before being implemented. Researchers have also raised concerns about the AI Act leaving certain issues unresolved. For example, questions around liability and intellectual property rights remain unclear. Legal proceedings are already underway, with copyright holders suing companies that use protected works to train AI models. These cases highlight the complexity of applying existing laws to emerging technology. Dr. Kristian Kersting, head of the Machine Learning department at the Technical University of Darmstadt, stressed the need for clearer guidelines, particularly regarding intellectual property. He argued that small businesses, lacking dedicated legal teams, struggle to navigate the current framework. “What the market now needs are clear legal guidelines on intellectual property and practical recommendations that even small companies can understand and implement,” Kersting explained. In Germany, a new law took effect on Wednesday, establishing a framework for overseeing the implementation of the AI regulation. Under this law, the Federal Network Agency will coordinate oversight and handle complaints from consumers. It will also support small and medium-sized enterprises by providing resources such as a “AI Real Lab,” where companies can develop, train, and test AI systems. The next major deadline for implementing further provisions from the AI Act is August 2, 2027, according to the Federal Network Agency. This continued rollout underscores the EU's commitment to shaping a comprehensive regulatory environment for artificial intelligence.