The Switch That Was Already On

That is the category the switch governs. It is a much narrower category than "AI", and the narrowness is deliberate rather than sneaky: AutoMod, recommendations and sponsorship tools all sit outside it, and all keep running whichever way you flick the switch.

Captions are the interesting edge, and Twitch draws that line itself rather than leaving you to guess. The caption feature keeps working either way, and the FAQ says Twitch does not retain your stream or your clips to train a model that can generate new content. But the speech-to-text model behind captions is Twitch's own worked example of what allowing training actually feeds:

The feature is outside. The model underneath it is inside. That distinction is the whole switch in miniature.

Training is not the same as using. Using your content means a machine looks at it, does a job, and moves on. Training means your content becomes part of what the machine is, permanently, mixed into a set of numbers from which nothing can be individually extracted again. The first is a service performed on your stream. The second is your stream becoming a small part of a product. The switch only speaks to the second one.

And "opt out" is a phrase that implies a prior choice. There was not one. The mechanism is the same one you meet in a hotel, where the towels are already on the rail and the card explains how to decline them. The word for the switch is not consent. It is withdrawal.

Everything the switch does, it does forwards.

The FAQ says your content "may be used for future Gen AI model improvements". Nothing in the article says that anything already absorbed comes back out, and nothing in the article says which way the switch shipped either. I searched the rendered text of that help page for the word default and found it zero times. Not once, in a document explaining a default.

So the two facts a person would actually want are the two the documentation declines to supply: which way it was set when it arrived, and whether flicking it undoes anything.

On the first, Twitch's chief product officer answered on a company livestream later that day, the Patch Notes show the second post of the thread had pointed at. Two outlets transcribed him slightly differently, which is what happens to spoken words, and both versions contain the phrase "nobody would opt in" and the sentence "That's honestly the answer." The reasoning, as reported, is that opt-in would have produced almost no participation, so the switch went out on.

On the second, he was asked directly whether a streamer's existing videos had already been used, and the reported answer is that he did not know, because he did not know what Amazon had done in terms of model training.

I want to be fair about that, because it is the line most likely to be misread. "I do not know" from a product executive about what a parent company's model teams have already ingested is not evasion, it is org chart. It is also, if you are a streamer with six years of archive, precisely the answer that does not help.

One more thing, and it is the one I keep turning over.

The setting says that turning it off does not stop Twitch and Amazon using your channel content "for other purposes described in the Twitch Privacy Notice". Fine. Let us go and read the Privacy Notice.

The relevant passage lists a long set of uses, and ends like this: "This includes using data to develop or deploy our generative AI models and services. See Amazon's Generative AI Development Disclosure for more information."

Develop or deploy.

I searched the whole notice, all twenty-nine thousand characters of it, for the word train. It does not appear. Not once.

So the switch governs a verb the Privacy Notice never uses, and the Privacy Notice describes verbs the switch never mentions, and the line where developing a model stops and training it starts is drawn in neither document. That is not a gotcha and I am not claiming the notice quietly permits what the switch forbids. I have no evidence for that and it would be a cheap thing to imply. What I am claiming is narrower and duller and more useful: the boundary is where every question you actually have lives, and neither of those two documents draws it.

There is a third document, though. The notice points at it by name, an Amazon Generative AI Development Disclosure, and I pulled the link out of the notice rather than guessing at it. It refused to load twice and then loaded fine, which is worth saying out loud, because a page that throws an error at you twice is not the same thing as a page that is hiding, and I nearly wrote the wrong sentence about it.

It does not draw the boundary. It dissolves it. The disclosure opens by saying it describes the data Amazon uses "to develop or deploy our generative AI models and services". Its second section is headed "Responsible Training Data Practices" and starts: "For the generative AI services we make available to our customers, we train and test on a range of data intended to enhance our services' capabilities." Develop or deploy is the umbrella. Training is one of the things that happens under it, and Amazon says so in its own document, one section apart.

So the two vocabularies were never describing different activities. They were describing the same activity at two altitudes, and the switch operates at the lower one.

What the disclosure never does, in any of its four sections, is mention Twitch. Not once. So I am still not telling you that the notice permits what the switch forbids, because nothing here says that. I am telling you the broader word was chosen on purpose, and that the document explaining the broader word is a document about Amazon, not about you.

You may never have streamed anything. Three things still land in your lap.

The pattern is not about Twitch. Consent that belongs to the room rather than the speaker is now everywhere. The group chat where one person's setting decides what happens to everybody's messages. The shared document owned by whoever created it. The photo you are in and did not post. The meeting somebody else recorded. Once you have seen the rule stated out loud in one FAQ, you start noticing that most platforms have the same rule and simply have not written it down.

You get a question that works on any AI setting, anywhere. Is it retroactive? That splits into two, and you need both. Does this stop future use, and does it undo past use. Almost every switch of this kind answers the first clearly and goes quiet on the second, and the silence is not accidental, because the honest answer is usually that nothing can be pulled back out of a trained model. Ask it about your phone, your work chat tool, your photo library.

And read the FAQ, not the announcement. The announcement is written to be quoted and it told nobody anything. The FAQ was published by the same company, about the same feature, probably on the same day, and it contains the line about the stranger's chat. When a company looks like it is hiding something, the information is very often sitting in public in a duller document, because the duller document has to be accurate and the announcement only has to be defensible.