As the research shows, when trained on low-quality data, neural networks begin to demonstrate a decline in their cognitive functions, reminiscent of the attention degradation observed in users who actively consume superficial content from the internet. For analysis, the team of scientists identified two key aspects: engagement (short viral posts) and semantic quality (presence of clickbait, errors, and clichéd phrases). Based on these criteria, datasets with varying proportions of "information waste" were collected.
The Llama 3 and Qwen 2.5 models, trained on such data, showed a sharp decrease in accuracy—from 74.9% to 57.2%. The reduction in the ability to understand long contexts was even more significant—dropping from 84.4% to 52.3%. The scientists termed this effect "dose-dependent": the more low-quality data, the faster the model's intelligence deteriorates.
Moreover, the research revealed a phenomenon known as "personality drift"—the models lost ethical boundaries and began generating more errors. Their logical reasoning became superficial, and their responses resembled disjointed fragments of text.
These results raise an important question about the future of artificial intelligence, as virtually all major laboratories, such as Anthropic, OpenAI, and Google, rely on human-generated content to train their models. However, as mentioned in other studies, companies are experiencing a shortage of quality data to train new models.
Alexis Ohanian, co-founder of Reddit, expressed the opinion that a significant part of the internet is "already dead" due to the abundance of bots and content created using quasi-artificial intelligence. He noted that "proof of life," meaning real human presence, is becoming increasingly rare.
Sam Altman, head of OpenAI, also shares these concerns, stating that most accounts on X are managed by LLM-based bots. He characterized the situation as a manifestation of the "dead internet theory."
According to data from Amazon Web Services, about 57% of content on the internet is created or translated using artificial intelligence, negatively impacting the quality of search results.
Jack Dorsey, former head of Twitter, warned that in the next 5–10 years, it will become increasingly difficult to distinguish real images and videos from generated ones, due to the spread of deepfake technologies. He stated that users will have to rely not on content but on their own experience to recognize reality.
Cornell scientists warn that if the internet continues to be filled with machine noise, artificial intelligence will start learning from its own distortions, losing the ability to think logically. Ultimately, "smart" models may turn into a reflection of the digital chaos they themselves have created.
