Experts Warn Of AI "Model Collapse" As Generative Systems Risk Becoming Dumber

Experts and commentators are forecasting the end of the generative AI hype, with warnings of an impending catastrophic “model collapse.”

"Regurgitive training," is the term used for when AI systems trained on AI-generated data become progressively less effective.

This term refers to a hypothetical scenario where future AI systems become progressively less intelligent due to the increasing presence of AI-generated data on the internet, Aaron J. Snoswell of Queensland University of Technology wrote for The Conversation.

Modern AI systems are built using machine learning, where the underlying mathematical structure is set up by programmers, but the actual "intelligence" comes from training the system to mimic patterns in data.

High-quality data is crucial for this process, and big tech companies like OpenAI, Google, Meta, and Nvidia have been scouring the internet for content to feed their AI models.

However, since the widespread availability of generative AI systems in 2022, more and more online content is partially or wholly AI-generated. Human text data for AI training is expected to be depleted by 2026.

In 2023, researchers began experimenting with using AI-generated data for training instead of human-generated data. While AI-generated content is cheaper and easier to source, it lacks the quality and diversity that human-generated data provides.

This has led to a phenomenon dubbed "regurgitive training," where AI systems trained on AI-generated data become progressively less effective.

This decrease in quality is akin to a digital version of inbreeding, leading to a reduction in the helpfulness, harmlessness, and honesty of AI models. In short, the overuse of AI systems could be undermining the very data sources needed to keep them useful.