Twitch’s AI training opt-out became the centre of a user revolt this week after the platform confirmed that Amazon has been harvesting streamer and audience data by default to train its generative AI models, and that the setting to stop it is buried deep in a dashboard most users never visit.
The announcement came via a Twitch livestream on Wednesday. Chief product officer Mike Minton defended the arrangement by saying the platform was ‘respecting’ users by offering the ability to opt out at all. He then undercut that framing entirely: ‘If it’s opt-in, nobody would opt-in. That’s the honest answer.’
That admission tells you everything about the power dynamic here. The decision to default to collection was not a technical necessity; it was a deliberate choice premised on user inertia. Twitch knows most people never audit their privacy settings. Amazon benefits from that inertia. The opt-out exists largely so the company can say it exists.
How the Twitch AI Training Opt-Out Actually Works
To disable data collection, users must open their Streamer Dashboard, navigate to the Security and Privacy tab, scroll to the bottom of the options list, and toggle off ‘training for Generative AI’. It is not a prominent setting. Several users reported that after toggling it off and returning to the page, the feature appeared switched back on.
Even for those who successfully opt out, the scope of what remains is worth reading carefully. WIRED noted that Twitch’s own FAQ states the opt-out ‘does not opt you out of Twitch and Amazon using your channel content for other purposes described in the Twitch Privacy Notice,’ including real-time sponsorship assistance, viewer recommendations, and community safety tools such as AutoMod.
Gizmodo reported that the opt-out covers only future training of generative AI models. Platform features such as AutoMod and automatic caption generation are explicitly excluded, though Twitch says content processed for those features will not be retained for model training.
The eligible content pool is broad. According to Insider Gaming, it includes live streams and their associated chat, VODs, Clips, Highlights, and any text or images on a channel page. If you type in someone else’s chat, that channel owner’s opt-out preferences govern whether your words can be used for training, not your own.
Minton said during the stream that he did not know whether users’ data had already been scraped, adding that he was unsure what Amazon ‘has done in terms of model training and what they’ve used and not used.’ That is a troubling admission from a chief product officer. He announced a policy whose historical scope he cannot account for.
Amazon’s Long Game With Twitch
Amazon paid $970 million in cash for Twitch in 2014, with retention agreements pushing the total deal value above $1 billion, according to Reuters. GeekWire reported the deal formally closed on 25 September 2014.
At the time of the acquisition announcement, Twitch had more than 55 million monthly unique visitors, over 1 million broadcasters, and users averaged 106 minutes per day on the platform, with total monthly views in July 2014 topping 15 billion minutes, according to CNBC. Amazon did not buy Twitch for the gaming audience alone; it bought a live, engaged, content-producing community that would only grow.
That community is now the raw material. Minton said during the stream that data collected will not be re-sold to other companies. He noted it may be used to refine speech-to-text models, which would improve automatic subtitles on both Twitch and Amazon video products. The beneficiary chain runs one way: from creator to Amazon’s product suite.
Head of community Mary Kish was at least honest about the reception: ‘We don’t expect you to be happy or excited about this. I don’t expect anyone to react to this favourably.’ The stream’s chat column filled with hundreds of objections. One viewer wrote that ‘nobody would opt in because nobody wants to feed AI with our creativity and content.’ Another argued that Twitch had an opportunity to set an industry standard by pushing back against AI data collection, rather than joining the queue.
Kish called data collection for AI training ‘an industry standard.’ That framing is doing a lot of work. Industry standard is not the same as user consent, and the gap between the two is precisely what made this announcement land so badly.
The episode raises a question game developers have also started asking: given that streamers broadcast copyrighted titles for hours every day, what does it mean for a rights-holder when a platform trains commercial AI models on that gameplay footage without a separate licensing arrangement?
Minton cannot answer what has already been used. That uncertainty is the real story, and it will not resolve itself in a settings toggle.
