Twitch streamers woke up this week to a new setting in their channel preferences: a toggle labelled “training for generative AI.” It defaults to on. Amazon, Twitch’s parent company, is now using streamer content — livestreams, clips, VODs, chat logs — to train its generative AI models. Streamers who do nothing are contributing whether they like it or not.
The setting was announced not as “we’re training on your content” but as “we’re adding a setting that lets you opt out.” That framing tells you most of what you need to know about how Amazon expects this to land.
The CPO Said the Quiet Part Out Loud
In a livestream on the official Twitch channel, Chief Product Officer Mike Minton and Head of Community Mary Kish addressed an audience of nearly 3,000 aggrieved users. The chat was full of anti-AI messages. One user asked the obvious question: why is this opt-out and not opt-in?
Minton’s answer was as honest as it was uncomfortable: according to TechCrunch’s reporting, he said “If this was opt-in, nobody would opt in. That’s honestly the answer.”
You don’t often hear a platform executive admit that the consent mechanism exists because the alternative would produce zero consent. It’s the kind of transparency that usually requires a congressional hearing.
What the Toggle Actually Does (and Doesn’t)
The opt-out, once enabled, applies only to data collected after the setting is changed. Previously ingested content remains in Amazon’s training datasets. There is no retroactive deletion. If your streams from the last two years have already been used to train a model, toggling the switch today doesn’t pull them back out.
The scope is also narrower than it appears. Twitch carved out AI-supported platform features — AutoMod, content recommendations — from the opt-out. Those systems continue processing user content regardless of the toggle. And a streamer’s chat activity on someone else’s channel is governed by that channel owner’s setting, not the individual user’s preference.
So the consent model is: default-on, prospective-only, with carve-outs for platform features and multi-party content. It’s opt-out in name, but the practical coverage is narrower than the toggle suggests.
This Is Not Just a Twitch Problem
The AI Governance Institute notes that Twitch’s design illustrates a consent architecture regulators in the EU and several US states are increasingly scrutinising. The EU AI Act’s transparency obligations took effect August 2, and California’s AB 2013 is pushing toward more prescriptive disclosure standards for training datasets.
But Twitch is not the outlier — it’s the pattern. Meta uses public Facebook and Instagram content to train its AI models, with no opt-out mechanism for most users outside the UK. X’s policy allows AI training on public posts. YouTube, owned by Google, feeds video into model development under terms of service that most users never read. The difference is that Twitch at least provides a toggle, even if it’s buried and backward-looking.
What makes the Twitch case notable is the admission. Minton saying “nobody would opt in” is the unspoken premise of every default-on training program. Twitch just said it on a livestream.
The Data Lineage Problem
For enterprises that source training data from third-party platforms, the Twitch episode highlights a gap that’s easy to miss. A user’s current opt-out status may not reflect the consent posture at the time their data was ingested. Content collected before the toggle was set — potentially years of streams — sits in training datasets with no mechanism for withdrawal.
This is the data lineage question that compliance teams keep running into: can you prove when consent was obtained, and can you honour a withdrawal that comes after the model has already been trained? Under GDPR’s right to erasure, the answer is increasingly “you have to try.” Under the EU AI Act’s transparency provisions, the provenance chain matters for general-purpose AI models. Neither framework has a clean answer for “the data was already used to train a model that’s now in production.”
NZ Angle: The Privacy Act Already Covers This
New Zealand’s Privacy Act 2020 doesn’t have AI-specific provisions, but it does require organisations collecting personal information to state the purpose — and training a third party’s AI models is a purpose that should be disclosed. The Office of the Privacy Commissioner has signalled that biometric data and AI-related collection are priority areas.
For NZ-based Twitch streamers, the practical question is whether Amazon’s default-on collection complies with the Privacy Act’s collection notice requirements. The toggle itself could be read as a form of notice, but defaulting to opted-in — especially when the CPO admits nobody would choose it — sits awkwardly alongside the principle that collection should be transparent and proportionate.
The broader issue is one we’ve covered before: as AI training becomes the default business model for platforms with user-generated content, the consent architecture determines whether users have a real choice or a fig leaf. Twitch’s toggle is at least a fig leaf with a name on it. Meta’s approach doesn’t even offer that.
❓ FAQ
Does opting out of Twitch’s AI training delete my past content from Amazon’s models?
No. The opt-out applies only to data collected after you toggle it off. Previously ingested content remains in training datasets with no retroactive withdrawal mechanism.
Can I opt out of my chat messages being used on other people’s channels?
Not individually. Your chat activity on another streamer’s channel is governed by that channel owner’s AI training setting, not your own preference.
Is Twitch the only platform training AI on user content by default?
No. Meta trains on public Facebook and Instagram content. X trains on public posts. YouTube feeds video into Google’s AI development. Twitch is notable for providing a toggle and admitting why it’s opt-out — not for being the only platform doing it.
What does New Zealand’s Privacy Act say about this?
The Privacy Act 2020 requires organisations to disclose the purpose of collecting personal information. Training a parent company’s AI models is a purpose that should be stated. The Act doesn’t have AI-specific rules, but the Privacy Commissioner has flagged biometrics and AI collection as priority areas.
📰 Sources
- TechCrunch — Amazon will train on Twitch streamers’ content by default, unless they opt out
- AI Governance Institute — Twitch’s Default Opt-In for AI Training Exposes Consent Design Risks
- Ars Technica — Twitch content has trained Amazon AI for years, but users can opt out now
- The Verge — Twitch streamers can now opt out from training Amazon’s AI
