How generative models will pollute their successors

Generative models like ChatGPT and Stable Diffusion are making it much easier for everyone to create high-quality content. But what happens when the internet becomes flooded with AI-generated content?

According to a study by researchers at several universities, machine learning models trained on content generated by generative AI will suffer from irreversible defects that gradually exacerbate across generations.

Key findings:

• Generative models approximate the distribution of the data used to train them

• When generative models are used to generate training data for their successors, the distribution of that data will be slightly different from the original

• When the process is repeated across multiple generations of models, the model’s behavior will change in erroneous ways

• Eventually, this will cause “model collapse,” a degenerative process whereby, over time, models forget the true underlying data distribution

• Model collapse can happen across all kinds of ML models, including LLMs, diffusion, VAEs, etc.

• The only way to prevent model collapse is to make sure to preserve access to the original data over time

• As the web becomes flooded with AI-generated content, access to human-generate data will become a new front for competition between tech companies

Read the full article on TechTalks.

For more on AI research:

• Why you should be careful about LLMs that imitate ChatGPT

• Are the emergent abilities of LLMs like GPT-4 a mirage?

• What you need to know about multimodal language models

• Can you trust ChatGPT and other LLMs in math?