Apple paid Shutterstock to train its AI

Tech giant Apple paid Shutterstock to use their image and video libraries to train their AI, it has been revealed. Apple allegedly paid between $25 million and $50 million to gain access to millions of photographs and videos.

Apple wasn’t the only tech firm to strike a deal with Shutterstock. In 2022, Google, Meta and Amazon all made a deal to use the photo and video libraries for their data.

According to a report by Reuters, OpenAI was also an early client of Shutterstock, paying between 2 and 4 cents per image. They also struck deals with Associated Press and Reuters to use their image and text to train the large language models (LLMs).

Apple paid between 5 cents and $1 per photo and $1 per video to use for data scraping from PhotoBucket. PhotoBucket was a popular photo-sharing app in the early 2000s. Although it has since become less popular, it still has around 2 million active users.

Who is winning here?

Of course, paying the stock sites gets around the unethical aspects of data scraping copyrighted material. But does it really? How well were the photographers compensated for those photographs, and did they consent to their work being used in this way when they signed up to stock sites?

According to Reuters, the authors of the photos and videos were paid for their usage—anything between $1 and $300 for videos lasting over an hour. The creators got roughly 20 to 30% of that. But is that enough?

It’s one thing to post a photo or video on a stock site and allow that image to be licensed for commercial or editorial work. It is quite another, at least in my opinion, for that image to be used to train generative AI tools.

Adobe trained Firefly off the back of its stock library. They claimed that this was ethical, as they had the copyright and licensing agreements to do this. Stock photographers were inevitably irked by this arrangement, at the time claiming that it wasn’t in the original terms and conditions when they signed up to the site.

They might be correct here. I think it is high time that giving permission to train AI data sets should be a separate agreement to other usage rights. Some people will be okay with that, and that’s fine.

However, by giving the rights to an image for editorial use, you are giving permission for that one image for x amount of time under x conditions. Training an AI model is essentially putting yourself out of work, at least as a commercial or stock photographer, in the long term. It makes very little sense to be on board with this.

Maybe people are just trying to make a quick buck while they still can.

[via 80lv]

---