The rise of generative artificial intelligence has made it possible for everyday people to access programs that can produce text, computer code, images, and music. This AI-generated content is rapidly taking over the internet, finding its way onto hundreds of websites, including popular ones like CNET and Gizmodo. However, as AI developers continue to scrape the internet for data, there is a growing concern that AI-generated content will start entering the training sets used to teach new models to respond like humans. This unintentionally introduces errors that accumulate with each subsequent generation of models. Evidence suggests that even a small amount of AI-generated text in a training diet can eventually become detrimental to the model being trained. This phenomenon, known as "model collapse, " leads to the model becoming practically meaningless. Computer scientists and researchers are already witnessing this with various types of AI models, such as language models, image generators, and probability distribution separators. This trend raises concerns about the future implications of model collapse, especially in more complex models that may exacerbate biases and lack the diversity found in human data. While larger models may offer more resistance to model collapse, there is little reason to believe they will be immune.

Research indicates that models suffer the most at the less frequently represented "tails" of their data, where a collapse can cause a loss of diversity in the AI's output. This poses a risk of amplifying existing biases, particularly against marginalized groups. As AI-generated content begins to infiltrate the realms relied upon for training data, such as language models used by news outlets and even Wikipedia, there is a need to address this growing saturation. Machine learning engineers are already exploring ways to protect the humanity of crowdsourced data by discouraging the use of language models and creating experiments that encourage more human-centered data. In the face of model collapse, one potential solution is to utilize standardized data sets for images that are carefully curated by humans, ensuring they consist only of human creations. However, distinguishing between human-generated data and synthetic content is a challenging task, even if the technology to do so existed. The pervasiveness of generative AI in tools like Adobe Photoshop further blurs the line between AI-generated and human-created content. In summary, the rapid increase in AI-generated content poses a threat of model collapse, leading to models that lose their meaning and perpetuate biases. We need to find innovative approaches to protect the integrity of human data and prevent the saturation of training sets with AI-generated content.