Massive web‑scale data has propelled generative AI (GenAI) to remarkable achievements across diverse sectors. Practitioners now train next‑generation models with AI‑synthesized data, easing the growing shortage of real‑world datasets. However, this creates a self‑consuming loop between model and data that can trigger model collapse (MC), undermining the trustworthiness of GenAI. Recent studies have identified MC manifestations such as quality degradation, distribution drift, and heightened adversarial vulnerability. Proposed countermeasures include diversifying synthetic data, applying model regularization, conducting adversarial training, and mixing real with synthetic data during training. Experiments across various scenarios show these methods can delay or mitigate collapse, yet challenges remain: lack of unified evaluation metrics, instability in cross‑modal transfer, and high costs for large‑scale deployment. This paper consolidates existing research, categorizes technical approaches, and highlights open problems in theoretical analysis, benchmark design, and interpretability. Blogger's Review: Model collapse poses a fundamental barrier to scaling generative AI, and only coordinated advances in data governance and model architecture can ensure sustainable intelligent systems.