Twitch Says It Will Use Your Videos to Train Amazon's AI — Unless You Turn It Off
Twitch will use creators' videos to help train artificial intelligence models for its parent company Amazon, and streamers will be included automatically unless they manually turn it off. Chief Product Officer Mike Minton and Head of Community Mary Kish announced the policy on August 12, 2026, to a live audience of nearly 3,000 users on the official Twitch channel. TechCrunch
Twitch described the change as adding a setting that lets users opt out of having their channel content used to train AI models across Amazon. To turn it off, users go to their channel settings, find the security and privacy tab, and toggle off the option labeled "training for generative AI." TechCrunch
Generative AI refers to systems that create new content, like text, images, or video, by learning patterns from large amounts of existing material. Training is the process of feeding that material into the system so it can learn those patterns.
The choice to make this opt-out rather than opt-in was deliberate. Minton told the live audience that if AI training were opt-in, nobody would opt in. TechCrunch
When asked whether streamers' content had already been used for Amazon's model training, Minton said he did not know. TechCrunch
Kish sought to contextualize the policy by noting that Meta uses public content from its platforms to train its own AI models. TechCrunch
Minton's thinking on this subject predates the announcement. While serving as Twitch's Chief Monetization Officer, a role he held before becoming CPO, Minton said Google's licensing of Reddit data shows that social content can be useful for training AI. The Information On his X account, Minton posted that criticizing AI is the "cool thing to do" that gets viewership, clicks, and hits on social platforms, and referenced grifters ripping off others. X
The distinction between Google's approach with Reddit and Twitch's approach with its own creators is worth noting. Google's deal involved a licensing arrangement, negotiated between two companies with defined commercial terms. Twitch's policy, by contrast, is a platform-level default applied to its user base, with the burden of action placed on individual creators. The Reddit precedent Minton cited involved payment flowing to a platform for data access. The Twitch policy offers no comparable compensation mechanism in the announced details.
Minton's public framing positions the policy as a rational step that happens to align with Amazon's broader AI ambitions, and treats AI criticism as driven by social engagement metrics rather than substantive concern. Whether that framing adequately addresses creator concerns about consent, compensation, and content ownership is a separate question, and one the live audience of nearly 3,000 may well have been weighing.
The admission that Minton did not know whether streamer content had already been used for training raises a practical question. If content has already been fed into Amazon's training pipelines, the new opt-out toggle addresses future use but may not address what has already occurred. For a platform built on live broadcasts that disappear after they end, tracking down and removing past training data presents challenges different from those faced by text-based platforms.
Research on opt-out defaults in data policy has shown that participation rates are dramatically higher than under opt-in systems. This is why Minton's candid acknowledgment that nobody would opt in reads less as an observation about user preferences and more as an acknowledgment that the default is doing the work. The framing of the toggle as a new "setting" inverts the typical product narrative; usually a new setting adds capability rather than removing a default commitment.
For Twitch's creator base, the practical reality is straightforward. Streamers who want to prevent their content from contributing to Amazon's AI models must locate and toggle a setting that many will not know exists. Those who do discover it will find it in a privacy tab nested inside channel settings, not shown during streaming or on the main dashboard.
The broader pattern here is one we have watched unfold across the social platform landscape. Meta, as Kish noted, has moved in the same direction. Platform operators are increasingly treating user-generated content as a training-data asset by default, relying on opt-out mechanisms and the gap between policy existence and user awareness to maximize the amount of content available. Twitch's move brings that pattern to live video, which raises its own distinct questions around performance, likeness, and real-time creative work.
What remains unclear is the scope of what Amazon intends to do with this content. Whether the models in question are aimed at content moderation, recommendation systems, generating new video, or some combination is not specified in the announced policy. The impact on creators will differ substantially depending on which of those applications materializes. A moderation model trained on stream content has very different implications than a model capable of producing video in a streamer's style.
Twitch has, at minimum, been transparent about the default and the mechanism to change it. The gap between transparency and informed consent, though, is where the real tension lives.
