In the quest to develop advanced artificial intelligence (AI) models, companies like OpenAI and Google are turning to unconventional sources of data. One such source is YouTube videos, which provide a wealth of text-based information that can be used to train AI models. However, a recent study conducted by digital media researchers at the University of Massachusetts Amherst has shed light on the privacy risks and potential concerns associated with using YouTube videos for this purpose.
The researchers collected and analyzed random samples of YouTube videos to gain a deeper understanding of the platform’s archive. What they discovered was a trove of videos that were not intended for wide dissemination. Many of these videos were meant for personal use or were created by children under the age of 13. Additionally, a significant portion of the videos had low view counts but high engagement, indicating that they were targeted at a small but highly engaged audience, such as friends and family.
This side of YouTube, which accounts for the vast majority of the estimated 14.8 billion videos on the platform, remains poorly understood. Big tech companies have become increasingly resistant to researchers, making it difficult to illuminate this aspect of social media. However, it is crucial to have a comprehensive understanding of the content that is being ingested by AI models, particularly when it involves user-generated videos.
The New York Times recently published an exposé on how OpenAI and Google have been mining YouTube for data to train their large language models. The archive of YouTube transcripts provides an extraordinary dataset for text-based models. However, there are concerns about YouTube’s terms of service, copyright issues, and the sheer scale of the archive. With over 14 billion videos uploaded by people worldwide, it becomes challenging to determine the content it contains.
One surprising finding from the study was the prevalence of videos featuring children or created by them. YouTube requires users to be at least 13 years old to upload videos, but the researchers observed numerous videos featuring children who appeared to be younger than the age requirement. While age validation on the internet is notoriously difficult, it raises questions about what content is being consumed by AI models developed by tech giants.
Contrary to popular belief, AI companies like OpenAI do not necessarily rely solely on highly produced influencer videos or TV newscasts for training their models. Research on large language model training data has shown that even virtually unwatched conversations between friends can provide valuable linguistic training data. This raises concerns about the lack of transparency regarding the training materials used by AI companies, as biases and privacy issues can arise without proper oversight.
The sheer volume and complexity of YouTube make it impossible to fully review its content. Without strong policies in place, AI companies that ingest a significant fraction of the YouTube archive may unintentionally include content that violates privacy regulations, such as the Children’s Online Privacy Protection Rule. This rule prohibits the collection of data from children under the age of 13 without proper notice.
As individuals, we may unknowingly contribute to the training of AI models like ChatGPT and Gemini through our seemingly innocuous YouTube uploads. To AI, every video carries potential value, regardless of its view count or the uploader’s intentions. This highlights the need for comprehensive privacy legislation and stronger legal protections for user data in the United States.
The study conducted by the University of Massachusetts Amherst researchers serves as a reminder that AI companies must navigate the complex landscape of user-generated content carefully. It is crucial to strike a balance between harnessing the potential of vast data sources like YouTube while respecting privacy rights and complying with regulations. As AI continues to evolve, it is essential to remain vigilant in addressing the privacy risks and concerns associated with its development.
Use the share button below if you liked it.