Article by Melissa Heikkilä and Stephanie Arnett: “AI is all about data. Reams and reams of data are needed to train algorithms to do what we want, and what goes into the AI models determines what comes out. But here’s the problem: AI developers and researchers don’t really know much about the sources of the data they are using. AI’s data collection practices are immature compared with the sophistication of AI model development. Massive data sets often lack clear information about what is in them and where it came from.
The Data Provenance Initiative, a group of over 50 researchers from both academia and industry, wanted to fix that. They wanted to know, very simply: Where does the data to build AI come from? They audited nearly 4,000 public data sets spanning over 600 languages, 67 countries, and three decades. The data came from 800 unique sources and nearly 700 organizations.
Their findings, shared exclusively with MIT Technology Review, show a worrying trend: AI’s data practices risk concentrating power overwhelmingly in the hands of a few dominant technology companies.
In the early 2010s, data sets came from a variety of sources, says Shayne Longpre, a researcher at MIT who is part of the project.
It came not just from encyclopedias and the web, but also from sources such as parliamentary transcripts, earning calls, and weather reports. Back then, AI data sets were specifically curated and collected from different sources to suit individual tasks, Longpre says.
Then transformers, the architecture underpinning language models, were invented in 2017, and the AI sector started seeing performance get better the bigger the models and data sets were. Today, most AI data sets are built by indiscriminately hoovering material from the internet. Since 2018, the web has been the dominant source for data sets used in all media, such as audio, images, and video, and a gap between scraped data and more curated data sets has emerged and widened.
“In foundation model development, nothing seems to matter more for the capabilities than the scale and heterogeneity of the data and the web,” says Longpre. The need for scale has also boosted the use of synthetic data massively.
The past few years have also seen the rise of multimodal generative AI models, which can generate videos and images. Like large language models, they need as much data as possible, and the best source for that has become YouTube.
For video models, as you can see in this chart, over 70% of data for both speech and image data sets comes from one source.
This could be a boon for Alphabet, Google’s parent company, which owns YouTube. Whereas text is distributed across the web and controlled by many different websites and platforms, video data is extremely concentrated in one platform.
“It gives a huge concentration of power over a lot of the most important data on the web to one company,” says Longpre…(More)”.