Twitch adds opt-out for amazon ai training but default remains on

Twitch updated account settings and added a toggle that lets streamers request that Amazon not use their content for training generative models, but the option is off by default. The move was discovered only after the settings update went live and prompted immediate backlash: more than 16 000 creators signed a dedicated forum petition against automatic use of their material.

The opt-out path runs through the avatar icon in settings, then security and privacy, then the generative AI training option. Twitch clarifies alongside the toggle that turning it off does not prevent use of content for other AI purposes listed in the privacy notice, such as the AutoMod community-safety tool, viewing-recommendation algorithms for new audience discovery, and real-time assistance for sponsorship campaigns. In other words, the opt-out is limited to model training and does not affect internal feature deployment.

The change was presented live by Marty Kish, head of community at Twitch, who admitted in advance that the audience would not like it. Mike Minton, head of product, described his answer as “honest” when he said keeping the toggle on by default was essential because otherwise “no one would have participated” in the process. Minton added that the mechanism is not unique to Twitch and estimated that all publicly available content is probably already being used for training in one way or another, with or without permission, “much of it beyond our direct control.”

The terms of service effective March 2024 grant Twitch and its sublicensees a broad right to use, copy, modify, adapt, distribute and create derivative works from the content. Until now the clause did not explicitly mention generative model training, and the current update exposes the gap between the general wording and the specific usage in practice.

The episode illustrates the widening shortage of high-quality data for training. While OpenAI is signing agreements with publishers such as Condé Nast, the parent of WIRED, and Meta is using posts and images from Facebook and Instagram, the pool of available sources is shrinking relative to demand. The resurfacing of alternative approaches again raises questions about transparency and the ethics of training processes.