- Meta uses public photos and texts from Instagram and Facebook to train its AI text-to-image generator.
- Meta executive Chris Cox told Bloomberg’s Tech Summit that the company doesn’t “train on private things.”
- The chief product officer’s comments come as big tech companies rush to obtain data to train AI models.
Big tech companies are fighting for AI training data, and Meta appears to have a big advantage over its rivals: using photos from Instagram and Facebook.
Meta chief product officer Chris Cox told Bloomberg’s Tech Summit on Thursday that it uses publicly available photos and text on the platforms to train its text-to-image generator model called Emu.
“We don’t train on private things, we don’t train on things that people share with their friends, we train on things that are public,” he said.
Meta’s text-to-image model can produce “really amazing quality images” because Instagram has so many photos of “art, fashion, culture, and also just images of people and us,” Cox added.
Users can create images in Meta AI by typing a prompt starting with the word “imagine,” and this will generate four images, according to its website.
AI models need to be fed and trained with data to be effective. It has been a controversial issue as there has been almost there’s no way to avoid it copyrighted content is scraped from the Internet and used to create an LLM.
However, the US Copyright Office has been trying to address this issue since early last year and is considering updating its laws to address it.
One way companies try to obtain data is by joining forces with other companies. OpenAI, for example, has partnered with several media outlets to license its content and develop its models.
Meta even considered acquiring publisher Simon & Schuster in a bid to get more data to train its models, The New York Times reported last month.
In addition to raw data sets, companies use “feedback loops” – data collected from past interactions and results that are analyzed to improve future performance – to train their models. It includes algorithms that tell AI models when there is an error so they can learn from it.
Meta CEO Mark Zuckerberg last month told The Verge that feedback loops will be “more valuable” than any “initial corpus.”
Meta did not immediately respond to a request for comment from Business Insider, made outside of normal business hours.
