According to a study published on the preprint server arXiv, cited in the journal Nature, scientists from the University of Texas at Austin analyzed the effect of “useless” data — such as short, superficial posts and sensationalist material — on the behavior of artificial intelligence. The study focused on multiple aspects, including logic and reasoning, extracting information from long texts, ethics, and even the personal traits of the models.

The results showed that the higher the percentage of bad data in the training process, the errors the language models had and the less logical they were, including in multiple-choice tests.

Lead researcher Zhangyang Wang reiterated the classic principle in AI science: “Garbage in inputs, garbage in outputs.”

The new analysis emphasized the importance of carefully selecting data when training models. The researchers used a million posts from a popular social media platform to retrain the open models Llama 3 and Qwen — the former is known for following instructions, while the latter is classified as a reasoning model.

The analysis showed that Llama’s model changed its behavior after training on low-quality data, as “positive” traits decreased and “negative” traits such as narcissism and psychopathy appeared.

Attempts to correct the imbalance — such as retraining on high-quality data or modifying instructions — only partially improved performance, while problems with reasoning and skipping analytical steps persisted.

This issue is of particular importance as social media platforms tend to expand the use of user data to train artificial intelligence systems. For example, starting this November, LinkedIn plans to use European user data in its generative systems.

Source: Naukatv.ru