Tech and AI

Twitch turns AI training on by default and says why

By Staff Writer  |  13 August 2026

Streamers can now stop their channels being used to train Amazon's models. The switch arrives already enabled, and the product chief gave the reason in one sentence.

Twitch has added a control called Training for Generative AI to its channel settings, which lets a streamer refuse permission for their material to be used in training models across Amazon. The control is switched on when a streamer finds it. Everyone on the platform is therefore contributing unless and until they go and look.

With the setting enabled, the platform says channel content may be used to improve Amazon's generative models, and it defines that content broadly: livestreams, videos on demand, clips, highlights, text, images and chat messages. The resulting models are not confined to the platform either. The company gives the example of stream audio being used to refine speech to text, which then improves captions on Twitch and elsewhere across Amazon.

The setting belongs to the channel owner, so opting out of training on your own channel does nothing for what you type in somebody else's chat if that streamer has left training on.

The reason given was unusually direct

Twitch's chief product officer, Mike Minton, was asked during a livestream why the control had not been made opt-in. Nearly 3,000 users were watching, and a large share of the chat was asking the same question.

If it was opt-in, nobody would opt in. That's honestly the answer. So it's going to be on by default.

Mike Minton, chief product officer, Twitch

That is a candid statement of the commercial position, and it is also an admission about consent. A default that exists because the alternative would collect no data is not really a choice architecture; it is a collection mechanism with an exit door. Streamers on the platform record many hours of their own voice and image every week, which is precisely what makes the archive valuable and precisely why the default matters.

What the switch does not do

The wording is forward looking. Opting out means content will not be used for future model improvements. Nothing in the announcement suggests that flicking the switch removes anything already absorbed into a trained model, and on current technique nothing could.

Nor is this the beginning of the practice. Minton has previously confirmed in public that Amazon was already using Twitch data to train models, which was reported in 2024. Asked during the livestream whether a particular user's videos had already been used, he said he did not know what had been used and what had not. The change is that there is now a documented way to say no to the next round.

Twitch's head of community, Mary Kish, told the same audience that the option existed because the community had made its opposition to generative AI training plain, and noted that other platforms take public content without offering an equivalent control at all.

Where to find it

The control sits in channel settings rather than the creator dashboard, under the security and privacy tab, near the bottom, labelled training for generative AI. It is a toggle, and it needs turning off.

For anyone whose livelihood depends on their own recorded voice and likeness, the practical lesson is old and dull: read the settings page when a platform changes it, because the default was chosen by someone whose interests are not yours.