Twitch Switched On AI Training for Every Channel, Then Rewrote Its Contract the Same Day. A Lawsuit Says the Scraping Started in 2024.

On August 12, 2026, a new toggle appeared in the settings page of every Twitch channel on Earth. It is called Training for Generative AI, it covers your streams, your archived videos, your clips, your chat logs and your channel artwork, and it hands all of that to Amazon to train generative AI models.

It was switched on. Nobody switched it on. It arrived that way.

Asked on Twitch’s own Patch Notes livestream why the company had not simply asked, Chief Product Officer Mike Minton gave an answer of unusual candour.

Eight days later, on August 20, a Connecticut streamer named Warren Pandiscia filed a proposed class action against Twitch Interactive and Amazon.com in the U.S. District Court for the Northern District of California, docketed as 3:26-cv-08721. Pandiscia is not a celebrity. He has streamed for about ten years, has just over 900 followers and more than 1,000 hours of broadcast behind him — which is roughly the profile of the several million people who stream on Twitch in any given month.

His complaint does not really argue about the toggle. It argues about the two years before it.

The Edit Nobody Was Meant to Read

The most damaging paragraph in the complaint is not about AI at all. It is about a sentence.

Twitch’s Terms of Service have, since October 2023, granted the company a licence to use creator content “in connection with monetizing the Twitch Services.” On August 12, 2026 — the same day as the toggle — that clause, Section 8(a)(i), was quietly rewritten. The new wording covers use “in connection with Twitch and/or its affiliates’ business.”

The affiliate in question is Amazon. In the same edit, Twitch added language stating that users are “directing Twitch to also cache and/or store your User Content,” and the Privacy Notice gained a new permission to use member data to “develop or deploy generative AI models.”

Pandiscia’s lawyers turn that housekeeping into the centre of the case with one line: the amendments “are themselves evidence” that the earlier terms “did not convey the rights Defendants exercised.”

It is a genuinely awkward argument to answer. If the old contract already permitted training Amazon’s models on Twitch streams, no rewrite was necessary. If a rewrite was necessary, then the training that happened under the old contract happened without a licence — and a company cannot grant itself permission in 2026 for something it did in 2024.

“Yeah, For Sure”

The 2024 quote is doing enormous work here, and it came from inside the building. Asked directly whether Amazon was using Twitch material for AI, Minton — then Chief Monetization Officer, now Chief Product Officer — said “Yeah, for sure,” framing it as prototyping.

Two years later, on the August 2026 stream, he was asked whether Amazon had already trained on creator content before the disclosure. His answer: “I don’t actually know the answer to that question because I don’t know what Amazon has done.”

Take those two statements together and you have the shape of the problem. The platform confirmed in 2024 that the content was being used. In 2026 the executive responsible for the product could not say what had been done with it. Creators, in both years, were told nothing at all.

What Everyone Else Pays For This

The complaint’s unjust-enrichment count rests on a comparison that has become the standard yardstick for training data: what the same material costs when somebody has to buy it.

Google pays Reddit roughly $60 million a year for training rights to its forums. Reddit has disclosed more than $200 million in total AI data-licensing revenue across its deals. That is text — comment threads, no video, no audio, no faces.

Twitch is video, audio, screen capture, live human speech and synchronised chat, at a scale of 240 million-plus monthly users and millions of active creators. It is arguably the richest multimodal corpus of unscripted human behaviour that any single company controls. The complaint’s argument is that Amazon acquired it for nothing while competitors were writing eight-figure cheques for less.

Your Chat Messages Are Not Yours to Withdraw

There is a design detail in the toggle that has received less attention than it deserves, and it will matter to far more people than the lawsuit will.

The setting is channel-scoped. It belongs to the broadcaster. Chat messages are governed by whichever channel they were typed in.

So a viewer who has opted their own channel out — or who does not stream at all and has no channel to opt out — still contributes every message they send in somebody else’s chat, if that broadcaster left the default alone. Twitch chat is tens of millions of people, most of whom have never opened a stream key in their lives, and none of whom were offered a switch. The opt-out was built for creators. The conversation it captures is not only theirs.

The Case Is Not a Slam Dunk

The complaint pleads four counts: breach of implied contract and the implied covenant of good faith, unjust enrichment in the alternative, breach of express contract, and a violation of California’s Unfair Competition Law (Business and Professions Code §17200). California’s Invasion of Privacy Act appears too — but only as an unlawful predicate feeding the §17200 count, on the theory that Amazon is not a participant in the exchange between streamer and viewer and is therefore intercepting it. It is not pleaded as a standalone wiretap claim, which is a signal about how confident counsel is in it.

Note what is absent: there is no copyright count. That is a deliberate choice — contract and unfair-competition theories avoid the preemption fight and the fair-use battlefield where AI training cases have mostly gone to die.

The Other Default Is Right There

The obvious retort to Minton is that nobody would opt in. It is probably true. It is also an argument that the value being taken is real and that creators would price it above zero if asked, which is closer to the plaintiff’s case than to the defence.

And the counterexample is a direct competitor. YouTube’s third-party AI training control, shipped in December 2024, is off by default: creators must actively grant access, and can name which companies get it. YouTube still trains its own models on its own platform — the toggle governs outside developers, not Google — so it is not a purity contest. But on the specific question of whether you assume consent or ask for it, the two biggest video platforms in the world made opposite calls, eighteen months apart, and only one of them is being sued this week.

What To Watch

• Whether Amazon says what Nova was trained on. The complaint cannot prove Twitch data is in any specific model. Discovery is the only path to that answer, and it is the reason this case is dangerous to Amazon regardless of how the contract question resolves.
• The August 12 terms edit in court. If a judge reads a same-day contract rewrite as an admission that the old licence did not cover training, that reasoning travels — every platform that quietly broadened a licence clause in the last three years is exposed to the same inference.
• Class certification. Millions of creators with wildly different follower counts, revenue and content types is a commonality problem. This is where the case most plausibly dies.
• Chat, not streams. The channel-scoped design leaves viewers with no control over their own messages. It is the cleanest privacy claim in the whole affair, and nobody has brought it yet.
• Whether any platform reverses a default. Backlash has moved settings pages before. A default that flips back to off would be the first real evidence that creator pressure can price this at something other than zero.