Art as an Algorithmic Virus: Unifying Generative Collapse and AI Value Alignment via Cognitive Affordances

This article proposes the metaphor of art as an algorithmic virus, unifying the two phenomena of generative collapse and AI value convergence through the concept of cognitive affordances, providing an interdisciplinary new path for AI safety research. The core view is that art can serve as a key framework for understanding AI value alignment issues, revealing the mechanisms of value conflict and convergence in large model training.

AI value alignment is a core issue in current AI safety research, which requires ensuring that model behaviors align with human values. Generative collapse refers to value-level conflicts caused by the tension between the model's immediate requests and safety constraints (e.g., fluctuating outputs when rejecting harmful requests); AI value convergence means that diverse values tend to converge to specific patterns during training, but methods like RLHF have limitations such as uncertain intent inference, scale constraints, and execution failures.

Cognitive affordances borrow concepts from the field of human-computer interaction and are extended to the possibilities that artworks provide for human cognition, such as interpretation and emotional responses. Large models not only learn language rules during training but also grasp the cultural context, social norms, and values behind language. As a condensed expression of human culture, art provides rich value signals for models.

The metaphor of art as an algorithmic virus includes three layers: 1. Contagiousness—art and AI content can spread quickly and resonate; 2. Host: human mind (cognitive architecture, emotions, cultural background); 3. Selection pressure: the training mechanism based on human feedback is similar to the selection process in cultural evolution.

Generative collapse is the friction when there is insufficient cognitive affordance (the model cannot coordinate conflicting goals); value convergence is the result of capturing cognitive affordance (the model identifies and uses value patterns that meet multiple goals). Art provides an ideal training signal for this kind of learning, and the paper formalizes the framework through six hypotheses, providing a roadmap for empirical research.

Implications of the research for AI safety: 1. When designing value alignment solutions, focus on cognitive affordance design rather than just listing rules; 2. Use artistic narratives as a medium for value education (similar to how humans learn morality through stories); 3. Treat generative collapse as a signal of the model's internal value conflicts, not just a bug to be eliminated.

Theoretical limitations include insufficient empirical basis and unproven cross-cultural applicability; future directions: developing cognitive affordance measurement methods, exploring the impact of different art types on model values, researching human-AI collaborative value negotiation, and extending to multimodal AI systems.