Twitch put creator streams into Amazon's AI training by default

Twitch began using creators' channel content to train Amazon's generative AI models this week, with streamers opted in by default and required to find a buried settings toggle to opt out. Chief Product Officer Mike Minton defended the choice on the official Twitch channel to a live audience of nearly 3,000 angry creators with the most candid possible answer: "If this was opt-in, nobody would opt in."

Minton's quip is the design philosophy in one sentence, and it is the story. Twitch, owned by Amazon, has stopped asking its creators whether their streams should become training data and started assuming the answer is yes. The opt-out path lives in channel settings under "security and privacy", inside a toggle labeled "training for generative AI." It exists, and it works, and it is the kind of place almost no one will look.

The change was announced on the official @TwitchSupport account as a user-control improvement: "We've added a setting that lets you opt out of having your channel content used to train generative AI content models across Amazon." Read forward, that is a new privacy option. Read backward, it is a confirmation that Amazon has begun training on creator streams and is letting streamers reverse the decision, one by one, if they can find the right menu.

Minton sat beside Head of Community Mary Kish for a roughly hour-long stream that drew sustained criticism, including pointed posts from creators like @Legundo urging streamers to flip the setting off and push back on the default. Pressed on why the platform had not asked creators first, Minton returned to the participation argument: an opt-in policy would, in his view, have produced an empty opt-in list.

That is the part of the episode that travels beyond Twitch. The same week, Twitch publicly compared its approach to Meta's, which trains AI models on public Facebook and Instagram posts. Meta does not ask its users to opt in either; it asks them to opt out. The two companies are now operating from the same playbook, in which consent is presumed and the burden of refusal is shifted onto the person whose data is being used. Twitch did not invent this pattern, but the platform's near-monopoly on long, unbroken live voice and video sessions makes Twitch an unusually rich source for the kind of data AI training runs consume at scale.

The open question is whether the opt-out applies going forward or whether the horse has already left the barn. When a creator asked Minton directly whether streams already posted had been used to train models, his answer was: "I don't actually know the answer to that question because I don't know what Amazon has done in terms of model training and what they've used and not used." A platform executive cannot, on the record, tell creators whether their existing content is already in a training set. That non-answer is now part of the policy.

For now, the toggle is the only lever a streamer can pull. The setting is off by default for new uploads of recorded streams from today, and creators can disable it retroactively for prior content through the same channel-settings menu. Twitch has not said whether retroactive opt-outs will trigger removal of already-ingested data from active training runs. The unresolved part of the policy is whether "opt out" means "stop using my streams" or only "don't use the streams I make from this point on."