OpenAI’s recent interview with personal tech columnist Joanna Stern from the Wall Street Journal shed light on the importance of data details in their latest project, Sora. The interview took an unexpected turn when Stern inquired about the specific data used to train Sora, to which OpenAI CTO Mira Murati responded vaguely, stating that publicly available and licensed data were used. However, she was unable to provide concrete details regarding whether YouTube, Facebook, or Instagram videos were part of the training data.
While the interview may not have been a PR masterpiece, it highlighted the ongoing copyright battles surrounding generative AI models. Various stakeholders, including authors, photographers, artists, lawyers, politicians, regulators, and enterprise companies, are interested in understanding the data that trained Sora and other models. They want to ensure that the data was properly sourced and legally used.
The issue of training data extends beyond copyright concerns; it also raises questions of trust and transparency. If OpenAI did use publicly available videos without explicitly informing the public, it brings into question whether individuals were aware of their content being utilized. Similarly, other tech giants like Google and Meta, who own YouTube and Facebook/Instagram respectively, are likely employing publicly shared videos and images for training their models. While this may be legally permissible, changes in Terms of Service agreements often go unnoticed by the public.
The issue of training data goes beyond individual companies; it is a fundamental concern for generative AI. The datasets that enable large models like Sora to process vast amounts of data have implications for those whose creative work is included in these datasets. Until recently, the broader implications of using such datasets had not been deeply considered outside the AI community.
Data collection has a long history primarily focused on marketing and advertising. In theory, individuals provide their data in exchange for personalized advertising or a better customer experience. However, generative AI training data for massive models does not involve a direct exchange and may be seen as a violation of individuals’ work or privacy. The use of publicly available data for research purposes is generally accepted, but when it comes to commercial models, concerns arise.
The public’s acceptance of the fact that their social media posts and videos have been used to train commercial AI models remains uncertain. Will knowing that Sora was trained on SpongeBob videos and publicly available birthday party clips diminish the magic of the model? Perhaps over time, these concerns will fade away, especially if OpenAI and other companies prioritize developers and enterprise customers over consumer opinion. It’s possible that consumers have already resigned themselves to issues of data privacy.
Nonetheless, the devil lies in the details of the data. While companies like OpenAI, Google, and Meta may currently have an advantage, the long-term consequences of AI training data practices remain uncertain. The ongoing battles surrounding data details in AI training highlight the need for transparency and ethical considerations in the industry.
In conclusion, OpenAI’s interview revealed the significance of data details in their Sora project. The concerns raised about the sourcing and usage of training data extend beyond copyright to encompass trust, transparency, and public awareness. The repercussions of using publicly available data for commercial AI models are yet to be fully understood. As AI continues to advance, it is crucial for companies to address these concerns and ensure ethical practices surrounding data usage in order to build trust with stakeholders.
