The Twitch AI Training Switch Nobody Told Streamers About

Twitch AI training is now switched on by default for every channel on the platform, and most streamers heard about it from each other rather than from Twitch. On 12 August the company added a toggle called "Training for Generative AI" to the Security and Privacy tab in channel settings. It arrived already enabled. Leave it that way and Amazon can feed your streams, saved broadcasts, clips, chat logs and images into models that generate new content.

There was no email. No pop-up on login. No blog post flagging it in advance. Twitch's support account simply announced that a new control existed, framing the whole thing as a privacy option rather than what it actually is — a change in what happens to your work. A comment on Twitch's own support forum demanding the setting be opt-in instead of opt-out gathered close to 14,000 upvotes within hours.

If you stream from Brisbane or Perth, this lands on you exactly as it lands on someone in Los Angeles. Twitch runs one global rulebook. And that's worth paying attention to here, because Australia spent much of last year arguing over this precise question and landed somewhere very different from the answer now sitting in your settings menu.

What Amazon can actually take from your channel

The scope is broad. With the toggle left on, Amazon may use a channel's live streams, video on demand, clips, chat conversations, still images and text for future training of models that produce text, audio, images and video. That's not a narrow slice of metadata. It's the performance itself — voice, face, catchphrases, the bits that make a channel worth watching.

Asked during a livestream whether Amazon had already been using Twitch content before this control existed, chief product officer Mike Minton said he didn't know, because he couldn't speak to what Amazon had done. Twitch has been Amazon-owned since 2014. That answer did not calm anybody down.

Why Twitch AI training is on by default

Twitch's leadership took questions on the company's own channel in front of roughly 3,000 viewers, and the explanation was unusually blunt. Minton was asked why the setting wasn't opt-in, the way most people would expect a consent choice to work. According to TechCrunch, he answered:

If this was opt-in, nobody would opt in. That's honestly the answer.

It's a rare moment of corporate candour, and it gives the game away. Head of community Mary Kish described the new control as a response to the community making clear it didn't want to train generative models — which raises the obvious follow-up about why that response arrived as a switch you have to hunt for.

How to turn it off in about thirty seconds

The good news is that opting out is quick, free and doesn't require contacting anyone. The catch is that nothing prompts you to do it, so a creator who hasn't read gaming news this week is still opted in right now.

Where the switch actually lives

• Open your Twitch channel settings, not your general account preferences
• Go to the Security and Privacy tab
• Scroll to the bottom — the toggle sits near the end, labelled "Training for Generative AI"
• Switch it off

Turning it off stops future training use. It does not undo anything already done, and Twitch hasn't published a figure for how much content was ingested before the control appeared.

The gap the toggle doesn't close

Your setting only governs your channel. If you chat in someone else's stream and that streamer hasn't opted out, your messages can still be swept up under their preference. So a viewer who never streams has no real control at all — their words are governed by whichever channels they happen to hang out in.

What keeps running regardless

Twitch says other AI features are unaffected: AutoMod chat moderation, recommendations, automatic captions and real-time sponsorship tools all continue. The company's position is that these don't retain data to train content-generating models. That distinction matters, and it's the one thing in this rollout Twitch has explained clearly.

Australia already said no to this once

In August 2025 the Productivity Commission floated a text and data mining exception that would have let AI developers train on copyrighted Australian work without permission or payment. Writers, musicians and industry bodies pushed back hard. On 26 October, Attorney-General Michelle Rowland ruled it out, saying the government would not entertain such an exception, and pointed the conversation towards licensing instead.

Here's the uncomfortable part. That decision protects Australian creators from having their work scraped without consent — but a ticked box in a settings menu is consent, legally speaking. Copyright reform doesn't help when the permission was collected by default. For context on the stakes, one study cited during the debate put average Australian author earnings from writing at roughly $18,500 a year.

What to watch from here

Platforms rarely reverse these defaults on their own, but they do adjust under pressure — several have paused or region-limited similar rollouts after regulator interest. Watch for whether Australia's Copyright and AI Reference Group takes an interest in consent-by-terms-of-service, and whether Twitch adds a bulk control for viewers rather than only streamers.

In the meantime, don't wait for a notification that isn't coming. Open Security and Privacy today, check that toggle, and if you run a Discord or a stream team, tell your people to do the same — most of them still have no idea the switch exists.