Skip to content

Twitch turned on AI training for every streamer, then explained why it wasn't opt-in

On August 12, Twitch surfaced a new account setting called "Training for Generative AI." Streamers found it already switched on.

The setting governs whether Amazon can use your channel to train its generative AI models. Not just the live stream. Twitch's own description covers streams, on-demand recordings, clips, chat messages, and the images and text you post to your channel, all of it feeding models that generate or synthesize text, audio, images, and video. If you want out, it's in channel settings under security and privacy, not in the creator dashboard where you'd look for it.

What the default switch collectsTwitch's own description of the content covered by Training for Generative AI.LivestreamsVODson demandClipshighlightsChatmessagesChannelart & textAmazon's generative AI models: text, audio, images, video

What made this a story rather than a policy footnote was the explanation.

The quiet part

Twitch ran a Patch Notes livestream that week with chief product officer Mike Minton and head of community Mary Kish taking questions from an audience of roughly 3,000 creators. Chat filled with the obvious one: why is this opt-out instead of opt-in?

Minton answered it straight. "If it was opt in, nobody would opt in," he said, adding that this was honestly the answer.

He's almost certainly right, and that's the problem. The reason a default is worth this much to a platform is precisely that it collects consent from people who would have refused if asked. Every argument about how much friction an opt-in adds is really an argument about how many people you expect to say no.

Asked whether Amazon had already trained on creator content before the setting existed, Minton said he didn't know, because he didn't know what Amazon had done. He has separately acknowledged that training on Twitch content has been running at prototype scale since at least 2024. The new toggle applies going forward, so if the answer to that question is yes, the switch does nothing about it.

Read the exclusions before you relax

Turning it off is narrower than it sounds. Twitch says captions and AutoMod keep running on your channel regardless, on the reasoning that safety tooling protects everyone and can't be disabled per-user. Twitch's position is that these don't retain data for model training, which is a meaningful distinction and also one you have to take on faith.

The part most people missed is about chat. Your setting governs your channel. When you type in someone else's chat, the channel owner's preference decides whether your messages get used. You can opt yourself out completely and still be in the training set of every stream you watch.

What the toggle actually changesRead the right-hand column before you assume you're out.OPTING OUT STOPSFuture genAI training on yourstreams, VODs, clips, chat,and channel artApplies going forward onlyIT DOESN'T STOPAutoMod and auto-captionsAnything already used beforethe setting existedTwitch says these don't retain training dataIn someone else's chat, their setting decides, not yoursYou can be fully opted out and still be in the training set of every stream you watch

Why this lands on us

It would be easy to file this under streaming drama, but it's the same question we've been tracking all year from three other directions. In July we covered Godot banning AI-generated contributions outright and the Steam census showing one in three new games now disclosing AI. Twitch is the third face of it: not what AI produces and not what you must declare, but what gets taken from you to build the thing in the first place.

The scraping fights of the last few years were about the open web, where nobody had a contract and everybody had a lawyer's opinion. This is different. These are creators in a formal relationship with a platform, on terms that platform writes, and the terms changed underneath them with a toggle they had to be told about. Roblox, Discord, itch.io, YouTube, and every UGC platform that hosts what people make now has a worked example of how to do this and how it goes over.

For anyone building on a creation platform, the practical advice is boring and worth repeating. Check the defaults on every service that holds your work, check them again after product updates, and treat "we added a setting" as a signal to go read what the setting actually excludes. The gap between the announcement and the exclusions is where the interesting terms live.

We build a platform where people make things, so the standard we'd want applied to us is the one worth stating plainly. A default that only works because people don't notice it isn't consent, it's collection. If a choice is genuinely good for the person making it, asking them costs you nothing.

References