Twitch has handed creators a way to keep future channel material out of Amazon’s generative-AI training. The catch is that the door was installed after years of uncertainty — and it opens only for users who know to find it.

Amazon bought Twitch in 2014, and by 2024 a Twitch executive had publicly acknowledged that the parent company used platform content in AI work. “Yeah, for sure,” then-chief monetization officer Mike Minton said when asked whether Amazon used Twitch to train models, while adding that the work had to stay within privacy rules and user trust.¹

This week, Twitch formalized an opt-out setting covering streams, VODs, clips, chat, and channel text and images for future Amazon models that generate or synthesize text, audio, images or video.² The company’s practical case is that training can improve products such as speech-to-text captions on Twitch and across Amazon.

But the rollout landed as a confirmation, not a reassurance. During Twitch’s livestream explaining the change, Head of Community Mary Kish conceded, “We don't expect you to be happy or excited about this.”³ Chat filled with anti-AI messages, reflecting creators’ fears that their voices and long hours of video could help build tools that compete with them — without payment or explicit consent.

Minton’s response to demands for opt-in made the divide starker: “If this was opt-in, nobody would opt in. That’s honestly the answer.”⁴ Twitch argues it is not alone: Minton said public content is widely used for model training, permitted or otherwise. Yet he also said he could not say whether Amazon had already trained on Twitch livestreams because he was not responsible for Amazon’s training efforts.³

For Twitch, the toggle is a privacy choice. For many streamers, it is belated control over work Amazon may already have treated as raw material.