Large Language Models Suffer From Their Own Output: An Analysis of the Self-Consuming Training Loop