Twitch turned on AI training for everyone. Here's where the off switch is.
Twitch added a setting on Tuesday called "Training for Generative AI". It lives in Security and Privacy, near the bottom of the page, and it was already switched on by the time you found out it existed.
That's the news. It takes about thirty seconds to change, and I'll get to exactly where it is in a moment.
The bit that made me sit up was the answer Twitch's chief product officer gave on stream when someone asked why a thing like this wasn't opt-in.
"If this was opt-in, nobody would opt in. That's honestly the answer."
Mike Minton said that out loud, on camera, to the people it applies to. Same stream, he also said of the toggle's location: "it is a bit hidden".
I want to be fair, because there's something almost admirable in saying it that plainly. Most companies would have given you three paragraphs about improving the creator experience. But sit with the sentence for a second and it's a company telling you it knows what its creators would pick, and has picked the other thing for them.
Where the toggle is
On desktop:
- Click your avatar, top right
- Settings
- Security and Privacy
- Scroll down to "Training for Generative AI"
- Toggle it off
Or skip all of that and go straight to twitch.tv/settings/security.
On the phone app: Profile, then Settings, then Security and Privacy, same toggle.
No confirmation step, no cooling-off period, no email. It's off when you flip it.
What the toggle actually covers
Left on, Amazon can use your channel's content to train generative models, meaning the kind that produce text, audio, images or video. That's your live streams, your VODs, your clips and highlights, your chat, and the text and images on your channel page.
Chat is the one I'd point at. That isn't only your words. It's your community's.
What it doesn't cover
This is where I'd slow down, because the toggle is narrower than the headlines make it sound.
Turning it off does not opt you out of AI at Twitch generally. AutoMod still reads every message in your chat and you cannot turn that off. Captions carry on. Recommendations carry on. So does the sponsorship matching that decides which campaigns you get shown. Twitch's own wording is that it doesn't opt you out of "other purposes described in the Twitch Privacy Notice", which is a fairly large door to leave open.
It's a consent setting for generative model training specifically. It is not a privacy switch.
One thing Twitch has been clear about, for what it's worth: Minton said opting out has "zero impact on your ability to be discovered on Twitch". I've no particular reason to doubt that, and it's the fear I'd expect most people to have.
The question nobody has an answer to
Someone asked whether content had already been used for training before the toggle turned up. The answer:
"I don't actually know the answer to that question because I don't know what Amazon has done."
I don't think that's evasion. I think it's genuinely how this works inside a company that size, and that's sort of the point. The setting is forward-looking. Whatever has already gone in has already gone in, and the person whose job title includes "product" for the platform can't tell you what's in there.
While you're at it, YouTube
If you stream to more than one place, YouTube has a version of this and it works the opposite way round.
YouTube Studio, then Settings, then Channel, then Advanced, then "Third-party training". It's off by default. You opt in, and you opt in per company from a list of about eighteen, or to all of them at once.
Two caveats. It's third-party only, so Google training on your videos is not what that setting governs. And it's limited to channels with Content Manager admin access, which a lot of smaller channels won't have.
Kick hasn't published an equivalent toggle that I can find. If you know otherwise, tell me and I'll correct this.
What I actually think
The toggle itself is fine. Thirty seconds, no downside, go and flip it if you want it flipped.
What bothers me is the default, and specifically the reasoning given for the default. "Nobody would opt in" isn't a justification, it's a finding. If you run that thought experiment and the result is that your users would decline, you have learned what your users want. Shipping it on anyway isn't a product decision that happens to be unpopular. It's a decision made in full knowledge of being unpopular, which is a different thing.
And the disclosure, since this is a blog run by people who build an AI chatbot: we're an AI company writing about AI consent, so weigh everything above accordingly. Our bot reads your chat, because it can't answer your chat otherwise. That's the deal, and it's the entire reason you'd install it in the first place. Whether that's the same thing as your stream quietly becoming training data for a model you'll never see is a distinction I think matters, but then I would think that.
What I can do is stop leaving our own position implied. Our privacy policy said nothing about AI training until this week, which I didn't love once I noticed it, so it now has a section that does: we don't currently train models on your chat, and if that ever changes it would be aggregated and wouldn't carry usernames. Writing it down is the part that makes it worth anything.
Go and look at the toggle. Then decide for yourself which way you want it.
Dan