Summary
- Twitch has allowed Amazon to utilize streamer content for training generative AI models, with this option enabled by default.
- Mike Minton, the chief product officer, explained that it was made opt-out because "if it was opt-in, nobody would opt in."
- Disabling this option does not prevent Twitch from using channel content for functions like captions, recommendations, and moderation.
Twitch has officially started permitting Amazon to train generative AI models using the content created by its streamers, and this setting is automatically activated.
The Amazon-owned service announced this update on Wednesday, informing users about the new option that allows them to opt out of having their channel content used for generative AI training across Amazon. This announcement also revealed that such training was occurring, despite streamers having previously reported in late July that sources had indicated this change was forthcoming.
All content on a channel, including streams, videos on demand, clips, highlights, chat messages, text, and images, can be utilized to train models aimed at generating or synthesizing text, audio, images, or video. According to Twitch's FAQ, this could involve using a streamer's audio to enhance speech-to-text models, thereby improving captions both on Twitch and across Amazon.
The toggle for this feature, labeled "Training for Generative AI," is located under Security and Privacy at the bottom of the account settings page.
Image: Twitch'No One Would Choose In'
In response to a viewer's inquiry about whether their videos had already been utilized for training, Minton admitted he was unsure, as he lacked information on what Amazon had done regarding model training. Kish mentioned she had opted out personally, describing the control as a response to community concerns about their content being used for generative models. Minton expressed that he would feel comfortable with the training and would not opt out if he were a streamer.
When pressed about what happens to users who opt out but appear in streams of those who do not, Minton could only respond with, "That's a good question," according to Kotaku. However, the published FAQ clarifies that a streamer's preference will dictate whether a viewer's chat can be used. He confirmed that there are currently no plans to allow users to see which channels have opted in.
Opting out may not be as comprehensive as it appears, since Twitch can still operate AutoMod and generate captions, which the company states do not store content for training purposes. Some of these functionalities are designed to protect the platform overall, and therefore, the option to disable them for individual users is not available.
This development comes as AI companies are increasingly in need of training data and are turning to the platforms they own. TechCrunch highlighted that Twitch's vast repository of a streamer's recorded voice and visuals makes it a particularly valuable resource, and many content creators may unknowingly consent to this use.