Researchers find that AI-generated web content could make large language models less accurate

A newly published research paper suggests that the proliferation of algorithmically generated web content could make large language models less useful.

The paper appeared today within the scientific journal Nature. It’s based on a recently concluded research initiative led by Ilia Shumailov, a pc scientist on the University of Oxford. Shumailov carried out the project in partnership with colleagues from the University of Cambridge, the University of Toronto and other academic institutions.

AI models produce a growing portion of the content available online. In keeping with the researchers, the goal of their study was to guage what would occur in a hypothetical future where LLMs generate a lot of the text on the internet. They determined that such a scenario would increase the likelihood of so-called model collapses, or situations where newly created AI models can’t generate useful output.

The problem stems from the proven fact that developers typically train their LLMs on webpages. In a future where a lot of the web comprises AI-generated content, such content would account for the majority of LLM training datasets. AI-generated data tends to be less accurate than information produced by humans, which suggests using it to construct LLMs can negatively decrease the standard of those models’ output.

The potential impact is just not limited to LLMs. In keeping with the paper’s authors, the difficulty also affects two other sorts of neural networks generally known as variational autoencoders and Gaussian mixture models.

Variational autoencoders, or VAEs, are used to show raw AI training data right into a form that lends itself higher to constructing neural networks. VAEs can, for instance, reduce the dimensions of coaching datasets to lower storage infrastructure requirements. Gaussian mixture models, that are also impacted by the synthetic data issue flagged in today’s research paper, are used for tasks reminiscent of grouping documents by category.

The researchers determined that the difficulty not only affects multiple sorts of AI models but can also be “inevitable.” They determined that that’s the case even in situations where developers create “almost ideal conditions for long-term learning” as a part of an AI development project.

At the identical time, the researchers identified that there are methods to mitigate the negative impact of AI-generated training datasets on neural networks’ accuracy. They demonstrated one such method in a test that involved OPT-125m, an open-source language model released by Meta Platforms Inc. in 2022.

The researchers created several different versions of OPT-125m as a part of the project. Some were trained entirely on AI-generated content, while others were developed with a dataset wherein 10% of the knowledge was generated by humans. The researchers determined that adding human-generated information significantly reduced the extent to which the standard of OPT-125m’s output declined.

The paper draws the conclusion that steps can have to be taken to make sure high-quality content stays available for AI development projects. “To sustain learning over a protracted time frame, we’d like to make sure that that access to the unique data source is preserved and that further data not generated by LLMs remain available over time,” the researchers wrote. “Otherwise, it might turn into increasingly difficult to coach newer versions of LLMs without access to data that were crawled from the Web before the mass adoption of the technology.”

Image: Unsplash