A critical and emerging challenge dubbed "AI model collapse" is now a significant concern as generative AI models are increasingly, and often inadvertently, trained on synthetic, AI-generated data. This phenomenon creates a deleterious "self-consuming feedback loop" that risks a gradual but significant drift of the model's distribution away from reality, exacerbating existing biases, amplifying artifacts, and reducing the overall quality and diversity of subsequent AI generations. Experts warn that Large Language Models (LLMs) used in critical applications like search could face degradation in factual accuracy and a loss of creativity, as they become progressively detached from authentic human-generated data and real-world distributions. The NeurIPS 2025 workshop will convene leading researchers to address this urgent issue, focusing on theoretical and empirical studies of model deterioration, potential mitigation strategies (such as novel data curation techniques and hybrid training paradigms), and the profound ethical implications of this novel training paradigm, which could fundamentally compromise the trustworthiness and utility of future AI systems.