Twitch Rewrote Its Content License the Day It Opened Amazon AI Training

Twitch amended its Terms of Service on August 12, 2026, the same day it added a default-on setting that lets Amazon train generative AI on channel content. The purpose sentence of the user-content license had read, through the December 8, 2025 archive, “Twitch can exercise these rights in connection with monetizing the Twitch Services.” The live page now says “in connection with Twitch and/or its affiliates’ business.” Warren Pandiscia, a Connecticut streamer, filed a proposed nationwide class action against Twitch Interactive and Amazon on August 20, first reported by Courthouse News, alleging the companies used streams, VODs, clips, chat and channel pages without consent or a license fee.

This is not a copyright case. The complaint says Twitch videos are not registered with the Copyright Office, and it sues on implied contract, express contract, unjust enrichment and California’s Unfair Competition Law. It is also not a payout-split story: the license Amazon is relying on is, in Twitch’s own current terms, “royalty-free.”

Twitch’s Help page is more specific than the support post that launched the setting. Opting out means streams, VODs, clips, stream chats and pictures and text on the channel “will not be used in future training” of an Amazon model whose purpose is to generate or synthesize text, audio, images or video. Future is the word. The same FAQ leaves AutoMod, captions, recommendations and “real-time sponsorship campaign assistance” on. Chat is channel-scoped: if a viewer who has opted out types in someone else’s stream, “their opt-out preferences govern.” The Verge’s Jay Peters found the toggle already on when he opened settings on August 12.

Mike Minton, Twitch’s chief product officer, was asked on the August 12 Patch Notes stream why the control was not opt-in. “If it was opt-in, nobody would opt in,” he said, per Tubefilter. “That’s honestly the answer.” He also said “Twitch has not been training models” and that he could not “speak to what Amazon is training or not training as it relates to any specific usage of content.” The complaint treats that gap as the point: Amazon’s actual training set is not in the Help article.

The same-day paperwork is what the filing turns into evidence. The August 12 Terms add a recital with no counterpart in the December 2025 archive: by going live, a user is “directing Twitch to also cache and/or store” the work “and every individual piece of it.” The Privacy Notice, also last modified August 12, now says Twitch uses information “to develop or deploy our generative AI models and services,” and still contains a non-retroactivity sentence — Twitch “will not materially change our policies and practices to make them less protective of personal information collected in the past… without the consent of affected individuals.” The complaint alleges that covenant was broken. That is alleged, not found.

Pandiscia’s pleading puts him at more than 900 followers and over 1,000 hours streamed. He says he opted out once he learned the setting existed. The class is pleaded as every creator whose Twitch content was used to train Amazon AI without consent; numerosity is alleged at “millions.” The $5,000,000 figure in the complaint is the Class Action Fairness Act jurisdictional floor, not a damages schedule. Courthouse News, at filing, had no comment from either defendant; Engadget, writing Sunday, still did not.

Twitch has not published a training-license rate. Amazon has not said which models ingested Twitch video, how much of the catalog was used, or whether any of it was used before August 12. Those are the terms.