Ross Anderson Discusses AI Model Collapse as a Growing Problem in Online Content

Now, Anderson raises the alarm about a global threat to humanity—the collapse of large language models (LLMs). Until recently, most of the text on the internet was generated by humans. LLMs have now become the primary tool for editing and creating new texts, replacing human-generated content.

This shift raises important questions: Where will this trend lead, and what will happen when LLMs dominate the internet? The implications go beyond text alone. For instance, if a musical model is trained with compositions by Mozart, subsequent generations may lack the brilliance of the original and produce inferior results, comparable to a musical “Salieri.” With each subsequent generation, the risk of declining quality and intelligence increases.

This concept may remind you of the movie “Multiplicity,” starring Michael Keaton, where cloning leads to a decline in intelligence and an increase in the stupidity of each subsequent clone.

The same phenomenon can occur with LLMs. Training a model with model-generated content leads to irreversible defects and the degradation of text quality over generations. The original distribution of content becomes distorted, resulting in an influx of nonsensical information. Gaussian distributions converge, and in extreme cases, text may become meaningless. Subsequent generations may even misperceive reality based on the mistakes made by their predecessors. This phenomenon is known as “model collapse.”

The consequences of model collapse are significant:

✔️ The internet becomes increasingly filled with nonsensical content.

✔️ Individuals consuming this content may unwittingly become less informed and lose intellectual capabilities.

Fortunately, there is hope. Anderson suggests that model collapse can be prevented, offering a glimmer of optimism amid the concerns. To learn more about potential solutions and how to avoid the decline of online information quality, we encourage you to explore further.

While generative AI holds promise, it is essential to remain vigilant and address the challenges it presents. By understanding and mitigating the risks associated with model collapse, we can work towards harnessing the benefits of this technology while preserving the integrity of information on the internet.

Read more about AI: