Twitch's AI opt-out shipped after the training: consent as a retrofit

Twitch has added an account setting that lets streamers stop Amazon from training generative AI models on their channel content, and the setting sits off by default, which means the training is on ^[1][2]. The toggle is new, the practice is not, and that ordering is the part operators should read closely ^[14]. The scope is broad: streams, past broadcasts, clips, chat messages, and the text and images on a channel page ^[4]. According to a Twitch support page cited by Engadget, opting out keeps those out of Amazon's models "in the future," with the models defined as ones whose purpose is "to generate or synthesize text, audio, images or video" ^[5]. Twitch's own worked example is a streamer's voice training speech-to-text, improving captions on Twitch and on Amazon AI products elsewhere ^[6]. The control stops at the channel boundary: messages you post in another streamer's chat are governed by that streamer's setting, not yours ^[7]. And it only stops training. AI keeps running for chat moderation, ad selection and recommendations ^[8], and Twitch says some safety functions are not individual settings because "turning those off for one person would weaken safety and security for everyone else on the platform" ^[9]. The reason for the default was stated out loud. On a livestream with head of community Mary Kish, chief product officer Mike Minton answered the question the chat was spamming: "If it was opt-in, nobody would opt in." ^[10] TechCrunch put the live audience at nearly 3,000, many of them posting anti-AI messages ^[11]. Minton added that he is comfortable with AI training and would not opt out himself ^[12]. TNW's read is that a default chosen because users would refuse it is the textbook definition of a dark pattern, and that the unusual thing here is an executive describing the reasoning without dressing it up ^[24]. Nobody at Twitch would say how long this has been running. Asked whether a user's videos had already been used, Minton said, "I don't actually know the answer to that question because I don't know what Amazon [...] has done in terms of model training and what they've used and not used." ^[13] Wired reports it remains unclear when the practice began ^[15]; TNW reports the platform has been feeding the pipeline for a couple of years ^[14]. Twitch's terms of service, in effect since March 2024, already granted Twitch and its sublicensees the right to use, reproduce, modify, adapt and create derivative works from user content, but did not explicitly cover generative AI training ^[17]. Because the opt-out only applies going forward and the company's own product chief cannot say what has already been ingested, the setting offers no remedy for anything already used ^[25]. The distribution of this news matters as much as its content. Twitch framed the change as adding a setting rather than announcing that training would begin ^[18]; the practice surfaced when streamer Zach Bussey found the toggle and a screenshot spread ^[27]; Wired counted more than 16,000 creators registering opposition in a forum ^[16]. Asked why Twitch had not emailed streamers, Kish said not everyone reads their emails ^[21]. Kish also framed the opt-out itself as a response to community sentiment ^[20], and Minton argued the broader practice is beyond Twitch's control, saying it is "quite reasonable to assume that almost any publicly available content is used to train models in one way or another, with or without permission" ^[19]. Three things to watch. Whether Twitch ever publishes a start date and a list of which Amazon models consumed which content, since the current answer is that its own executives do not know ^[13]. Whether the March 2024 terms get amended to name generative AI explicitly, which would tell you whether the toggle is a policy change or a UI change ^[17].