Experts at Oxford University have warned that the rise in artificial intelligence (AI)-generated content could lead to the collapse of machine learning models if the industry fails to mitigate the risks. reports Edition of The Register.
A team of researchers found that using AI-generated datasets to train future models can lead to a failure to understand meaning, known as model collapse. As an example, they cite a model that started with a text about medieval European architecture and ended it with meaningless discussions about rabbits.
In a paper published in the journal Nature, the authors, led by Ilya Shumailov, a Google DeepMind fellow and graduate student at the University of Oxford, note that AI may miss less common lines of text in training datasets. This means that subsequent models trained on the output will not be able to account for these nuances, leading to an iterative loop.
Language modelers have already faced long-term attacks such as click farms, content farms, and trolls that mislead social media and search algorithms, but with the advent of large-scale language models (LLMS), the scope of potential content poisoning will increase significantly.
Emily Wenger, an assistant professor of electrical and computer engineering at Duke University, demonstrated model breakdown using a system that generates images of dogs. If a model overrepresents certain dog breeds, such as Golden Retrievers, subsequent models trained on the AI-generated dataset may forget that less common breeds exist. As a result, the model will crash and fail to produce meaningful content.
One way to solve this problem is to add watermarks to AI-generated content.
Previously, Elon Musk’s neural network was used without user consent I started teaching In posts on X.
What are you thinking?
Source: Gazeta
Jackson Ruhl is a tech and sci-fi expert, who writes for “Social Bites”. He brings his readers the latest news and developments from the world of technology and science fiction.