A couple months ago, Oxford and Cambridge researchers illustrated the risk of homogeneity in a study of AI Generated content in Nature magazine. The risk increases as AI gets trained not only on human-created content, but on other AI-generated content.

As an example, the researchers studied an AI model trained on images of different breeds of dogs. The source material included a naturally wide variety of dogs (French Bulldogs, Dalmatians, Corgis, Golden Retrievers, etc.). But when asked to generate an image of a dog, the AI model typically returned the more common dog breeds (Golden Retrievers) and less frequently the rarer breeds (French Bulldogs).

Over time, the cycle reinforces and compounds when future generations of AI models are trained on these outputs. It starts to forget the more obscure dog breeds entirely. Soon it only creates images of Golden Retrievers.

Eventually, the researches found, there’s “Model Collapse”, where the LLM is trained so much on AI-generated Golden Retriever images that the results turn nonsensical and stop looking like dogs at all.

Now, substitute dog breeds for whatever you’re trying to create — new products, packaging, advertising, communication, and the risk is that all outputs devolve to look the same.

tomfishburne • AI-Generated Homogeneity