Abstract
The transformative potential of generative artificial intelligence (AI) in creative society demands a nuanced approach to address its potential impacts on creators. Our research aims to understand the preferences of different creators regarding attribution and commercial use, while investigating the technical robustness of "machine unlearning" as a tool for copyright compliance.
Introduction
As generative AI models like Stable Diffusion and GPT-4 become ubiquitous, a significant tension has emerged between AI developers and the creative community. This paper explores two critical dimensions of this friction: the ethical gap in licensing alignment and the technical fragility of removing copyrighted content from trained models.
The Licensing Disconnect: Revealed Preferences
To quantify creator intent, we analyzed the distribution of licenses for over 700 million images on Openverse and millions of code repositories on GitHub. This analysis reveals a stark "attribution mandate." Over 90% of Creative Commons (CC) licensed images require attribution, a requirement largely ignored by current generative models.
As shown in Figure 1, there is a high prevalence of "Non-Commercial" (NC) and "Share-Alike" (SA) tags in the visual arts community. While the software community (GitHub) often permits commercial use, visual creators frequently signal a desire to restrict commercial exploitation—a preference currently overridden by indiscriminate web scraping for AI training.
Machine Unlearning and the "Relearning" Trap
When models are trained on unauthorized data, "machine unlearning" is often proposed as a surgical alternative to costly retraining. We tested Mass Concept Erasure (MACE), which uses cross-attention refinement to map specific tokens—such as "Adam Driver"—to generic categories like "person."
However, our "forget-then-finetune" pipeline reveals that these erasures are dangerously fragile. By finetuning a "cleansed" model on a small, unrelated dataset of different celebrities, we found that previously erased concepts, like the one shown in Figure 3, often resurface. This suggests that unlearning doesn't actually delete information; it merely shifts learned embeddings along a specific-to-generic axis. Subsequent finetuning on any specific concept can inadvertently push those generic embeddings back toward their original, specific identities.
Conclusion
Current unlearning methods act more as superficial masks than true deletions. The interconnected nature of the model's internal "map" means that erasing one concept often degrades similar ones, and the information remains latent enough to be recovered. Until these techniques become robust against downstream finetuning, they remain an unreliable solution for legal and privacy compliance in the creative economy.
Summary updated