AI Models Risk 'Model Collapse' from Training on Their Own Gibberish

AI models like ChatGPT generate synthetic content that fills the internet, which can then corrupt their own training data. This "model collapse" results in nonsensical outputs.

Recent studies demonstrate how recursive training on AI-generated data leads to blurry, unrecognizable images and text obsessed with random topics like jackrabbits.

Filtering out synthetic training data is becoming vital to prevent issues like exacerbated biases and gibberish outputs.

High quality human-generated data still outperforms synthetic data sets. AI could help de-bias human data.

For now, engineers must manually filter synthetic data to stop AI models from training on their own poor outputs. Human oversight remains essential.