It seems that producing intelligent content is beyond our human ability, as the saying goes, shit in, shit out. It seems that maybe what saves millions of jobs is what AI robots could do. Well, at least as present internet users, that would include me and others like me.
AI relies on copying content; it steals material without reimbursing the original content writers.
As you may know, top AI labs like Anthropic, OpenAI, Google, and more are heavily dependent on content uploaded and otherwise shared by humans on the internet to train their LLMs (large language models). Last year, a report suggested that these companies had hit a wall due to a lack of high-quality content for training, preventing them from developing advanced AI models.
And as it now seems, the same issues continue to haunt advances and development in the AI landscape. According to a new study by Cornell University, LLMs can get “brain rot” due to prolonged exposure to low-quality online data. It further elaborated that this heavily contributes to a decline in their cognitive capabilities.
For context, this kind of internet “brain rot” refers to prolonged exposure and consumption of low-quality and trivial online content. Studies show that this negatively impacts human cognitive capabilities, reasoning, and focus. The same can also be said about AI-powered models.
The researchers used two measures to assess and identify internet junk content. The first test was centered on engagement with short and viral posts, with a lot of engagement, while the latter focused on semantic quality with a bias on posts considered as low-quality and rife with a clickbait writing style.
Consequently, the researchers used the measures to construct datasets containing varying proportions of junk and high-quality content. They used the datasets to determine the impact of low-quality content on LLMs like Llama 3 and Qwen 2.5.
The goal behind the study was to determine the impact on AI systems when they continuously depend on low-quality content uploaded to the web, which is seemingly flooded with short, viral, or machine-generated content.
Perhaps of more concern, the study revealed that the accuracy of AI models purely using junk content fell from 74.9% to 57.2%. Their long-context comprehension capabilities were also negatively impacted, dropping from 84.4% to 52.3%. The researchers further revealed that the AI models’ cognitive and comprehension capabilities would only worsen with prolonged exposure to low-quality content for training, a phenomenon they referred to as a dose-response effect.
I suppose the bots are very similar to the saying that stupid people get dumber as they age, while smart people get smarter. Thank God that the Musks of the world create stupid robots.
Click on the image below. The article is entertaining
