In the On-Device AI Era, the Boundary of Copyright Protection is Shifting to End Devices

The core of generative AI is migrating from cloud servers to end-user devices such as smartphones, tablets, laptops, vehicles, wearables, and robots. This shift offers benefits like reduced network latency and cloud inference costs, while keeping personal data contained within devices. However, it raises critical questions about copyright protection: as the boundary for controlling copyright infringement moves from cloud servers to end devices, how can we effectively safeguard intellectual property? Cloud-based AI allowed centralized control through terms of service, prompt filtering, output filtering, logging, abuse detection, and post-hoc blocking. On-device AI, however, performs most processing internally—often offline—making it difficult for servers to track user inputs, model outputs, or deployment locations. This creates new vulnerabilities for copyright protection. AI models themselves are now at risk of theft, with components like training data, model structures, optimization techniques, and hardware acceleration technologies becoming valuable intellectual property. For example, if a lightweight model fine-tuned for financial document analysis is deployed as a mobile app, an attacker could extract weights or adapters to launch a similar service, compromising domain knowledge and security policies. OWASP's LLM security risk classifications highlight model theft as a persistent threat, evolving from standalone risks to complex vulnerabilities tied to supply chain weaknesses and on-device deployment architectures.