Fuel for the AI machine
It was with this broader understanding that we read The New York Times exposé on in a race to find new troves of data to train their large language models. An archive of YouTube transcripts makes an extraordinary dataset for text-based models.
There is also speculation, by an from OpenAI’s chief technology officer Mira Murati, that the videos themselves could be used to train AI text-to-video models such as OpenAI’s
The New York Times story raised concerns about YouTube’s terms of service and, of course, the that pervade much of the debate about AI. But there’s another problem: How could anyone know what an archive of more than 14 billion videos, uploaded by people all over the world, actually contains? It’s not entirely clear that Google knows or even could know if it wanted to.
Kids as content creators
We were surprised to find an unsettling number of videos featuring kids or apparently created by them. YouTube requires uploaders, but we frequently saw children who appeared to be much younger than that, typically dancing, singing or playing video games.
In our preliminary research, our coders determined nearly a fifth of random videos with at least one person’s face visible likely included someone under 13. We didn’t take into account videos that were clearly shot with the consent of a parent or guardian.
Our current sample size of 250 is relatively small – we are working on coding a much larger sample – but the findings thus far are consistent with what we’ve seen in the past. We’re not aiming to scold Google. Age validation on the internet is infamously difficult and fraught, and we have no way of determining whether these videos were uploaded with the consent of a parent or guardian. But we want to underscore what is being ingested by these large companies’ AI models.
Small reach, big influence
It’s tempting to assume OpenAI is using highly produced influencer videos or TV newscasts posted to the platform to train its models, but previous research on large language model training data shows that the most popular content is not always the most influential in training AI models. A virtually unwatched conversation between three friends could have much more linguistic value in training a chatbot language model than a music video with millions of views.
Unfortunately, OpenAI and other AI companies are quite opaque about their training materials: They don’t specify what goes in and what doesn’t. Most of the time, researchers can infer problems with training data through biases in AI systems’ output. But when we do get a glimpse at training data, there’s often cause for concern. For example, Human Rights Watch released a report on June 10, 2024, that showed that a popular training dataset includes many photos of identifiable kids.
The history of big tech self-regulation is filled with moving goal posts. OpenAI in particular is notorious for asking for forgiveness rather than permission and has faced increasing criticism for putting profit over safety.
Concerns over the use of user-generated content for training AI models typically center on intellectual property, but there are also privacy issues. YouTube is a vast, unwieldy archive, impossible to fully review.
Models trained on a subset of professionally produced videos could conceivably be an AI company’s first training corpus. But without strong policies in place, any company that ingests more than the popular tip of the iceberg is likely including content that violates the Federal Trade Commission’s Children’s Online Privacy Protection Rule, which prevents companies from collecting data from children under 13 without notice.
With last year’s executive order on AI and at least one promising proposal on the table for comprehensive privacy legislation, there are signs that legal protections for user data in the U.S. might become more robust.
