Background: KB’s AI mandate and the plikttexemplar framework
The Swedish National Library (KB) has been tasked by the government with developing artificial intelligence language models for public use. A key part of the plan is to train these models on plikttexemplar, the legally mandated copies of published works that publishers must provide to KB. This framework is intended to preserve Sweden’s literary and cultural output, but it also raises complex questions about copyright and compensation when those same works are used to train AI systems. The debate sits at the intersection of public interest, authors’ rights, and advancing AI technology.
The Authors’ Association’s position
Sweden’s Författarförbundet (Authors’ Association) has publicly criticized the proposal. In a press release, the association argues that the legal status surrounding training AI on authors’ and translators’ works remains unclear. They contend that KB does not have the right to train models on these works, except for components that fall under research as defined by law. The association says its mandate is to defend members’ rights in mass utilizations — scenarios where individual authors cannot easily monitor rights violations due to the scale of use. This position puts the association in opposition to the government-backed plan and highlights the potential for compensation or licensing disputes if training requires broader rights clearance.
The KB position
KB has offered a different view. In a previous coverage, KB’s chief legal counsel, Karin Lodin, described a balancing act within copyright law designed to harmonize the author’s rights with broader societal interests. She pointed to explicit restrictions that apply to research purposes and asserted that KB’s approach aligns with these statutory provisions. According to KB, the research exemption provides a safe harbor for certain uses, including training AI under specified conditions, and the library feels confident in this interpretation. The conflict between the two positions underscores a broader policy question: how should public institutions leverage copyrighted works to advance AI while respecting creators’ rights?
Legal context: the research exemption and mass utilization
The Swedish copyright framework contains provisions intended to enable research and education, but the scope of these exemptions is hotly debated when applied to modern AI training. The language used in the law aims to balance creators’ incentives with the public interest in scientific progress. Critics argue that training AI on authors’ and translators’ works, especially where remuneration or licensing is not arranged, could erode the value of creative labor. Supporters counter that research-oriented use is essential for national AI capabilities, and that exemptions serve to protect cultural advancement and innovation. The ongoing legal interpretation will likely shape how AI models in Sweden are trained in the coming years.
Implications for AI development in Sweden
Should the Authors’ Association prevail on tighter controls or expanded licensing requirements, AI developers could face greater barriers to using plikttexemplar data. This might prompt more licensing deals, a push toward synthetic or non-copyrighted data, or stricter guidelines for what counts as permissible research training. Conversely, if KB’s interpretation gains ground, research-focused training could proceed with fewer transactional hurdles, potentially accelerating AI development at the expense of certain rights holders’ expectations. The debate also signals potential implications for compensation models and future policy work, as stakeholders seek clearer rules for mass utilization of copyrighted works in AI training.
Next steps and what to watch
The discussion is evolving, with legal clarification likely on the horizon. Both sides have signaled a strong interest in finding a balanced approach that protects authors while enabling public institutions to contribute to AI innovation. The ongoing dialogue, along with any forthcoming court interpretations or regulatory updates, will determine how Sweden navigates AI training on copyrighted works in the future. Readers should watch for further statements from KB and the Författarförbundet, as well as coverage of any legislative or judicial developments that define the permissible scope of AI training data.
