Here’s why turning to AI to train future AIs may be a bad idea

Researchers caution against training AI language models on their own outputs, which could lead to "model collapse." This occurs when AI-generated text with subtle errors is used to train future models, potentially compounding errors over time and resulting in biased or nonsensical outputs. A study showed that after nine generations of such training, a model produced gibberish. To mitigate this, experts suggest using a mix of human and AI-generated data for training and monitoring for data drift.