인공지능의 법적 규율 Ⅰ: 범용모델과 생성모델

Transformer-based decoders offer generative capabilities without requiring fine-tuning, blurring the distinction between general-purpose and generative models. Given their disparate risk profiles, however, tailored legal oversight should be devised for each category. First, safety and security control over high-impact general-purpose models must be crafted to ensure alignment with international standards and national security objectives. Second, to broaden the scope of lawful data access, nebulous dataset protection legislation should be refined, with developers and providers tasked with mitigation measures based on infringement risk assessment. Particularly, developers of models intended for use in generative tasks should be mandated, (i) for retrieval-based models: to anonymize publicly unavailable personal data and compensate rightsholders through collective rights management (CRM) or private copying levies (PCV); and (ii) for generation-based models: to adopt reasonable mitigation measures to qualify for a safe harbor. Third, the misuse of synthetic media should be addressed by not only reinforced enforcement against misusers but also clear allocation of responsibilities across the value chain. Fourth, machine-assisted creation should be granted weaker legal protections. Finally, accuracy, fairness, and transparency concerns in general-purpose/generative models should be addressed by developing and standardizing measurement and mitigation technologies instead of immediate legal interventions.