AI Models Risk 'Collapse' When Trained on AI-Generated Data

New research warns that recursive use of AI outputs for training could degrade performance and lead to nonsensical results.

Studies show AI models trained on AI-generated data degrade quickly, losing original context and producing gibberish.
Researchers call this phenomenon 'model collapse,' which can occur within just a few generations of recursive training.
The issue arises as AI-generated content proliferates online, reducing the availability of original human-generated data.
Proposed solutions include watermarking AI-generated content and stricter vetting of training datasets.
Tech firms may need to rely more on human-generated content to maintain model accuracy and diversity.