Twitch will train Amazon’s AI on your streams unless you find the setting

Twitch has added a setting letting streamers stop Amazon using their streams, clips and chat to train generative AI, buried in security settings and off by default. Chief product officer Mike Minton explained the default plainly, saying that if it were opt-in, nobody would opt in.

Twitch has added a setting that stops Amazon training generative AI models on your content. It is off by default, which means the training is on. Chief product officer Mike Minton was asked why it works that way and answered directly: “If it was opt-in, nobody would opt in.”

Minton said it on a livestream this week alongside head of community Mary Kish, after streamer Zach Bussey found the toggle and the screenshot spread. He added that he is comfortable with AI training and would not opt out himself if he were a streamer.

What is covered is close to everything. Streams, past broadcasts, clips, chat messages, and the images and text on a channel page. Twitch’s own example is that a streamer’s voice could train speech-to-text models, improving captions on Twitch and Amazon’s AI products elsewhere.

Chat is where the setting stops working. Opting out protects what happens on your channel, but messages you send in someone else’s chat can still be collected if that streamer has not opted out. Gaps like that have pushed creators elsewhere before, to services built for human artists only.

The framing that Twitch has just started doing this is not quite right. The toggle is new, not the practice. Asked whether their videos had already been used, Minton said he did not know what Amazon had done for model training, and the platform is reported to have been feeding the pipeline for a couple of years.

Opting out also only stops training. AI keeps running on Twitch for chat moderation, ad selection and recommendations, and nothing in this change touches that.

The setting sits at twitch.tv/settings/security. Asked on the same stream why Twitch had not simply emailed people about it, Kish said not everyone reads their emails.

A default chosen because users would refuse it is the textbook definition of a dark pattern. What is unusual here is hearing an executive describe the reasoning without dressing it up.

None of the coverage says how the default works in Europe. That matters, because opt-out consent for AI training has been challenged there before, and Meta had to pause an identical approach in 2024 before regulators let it resume.