Twitch and Amazon face a class action over livestream AI training

A Twitch streamer sued Twitch and Amazon on 20 August over the use of livestreams to train Amazon’s generative AI. The complaint runs to 37 pages. Counsel filed it in the Northern District of California, San Francisco Division. The case number is 3:26-cv-08721.

Its central documentary claim concerns a contract. Twitch announced an opt-out setting on 12 August. That same day it rewrote the clause of its Terms of Service that governs what it may do with a creator’s content.

The clause that changed on 12 August

Section 8(a)(i) had granted Twitch a licence tied to a stated purpose. The complaint quotes the old wording. It read: “Twitch can exercise these rights in connection with monetizing the Twitch Services.”

That phrase now reads “in connection with Twitch and/or its affiliates’ business”, according to the complaint. Amazon is the affiliate.

The same amendment added a new recital. It says that by streaming, a user is “directing Twitch to also cache and/or store your User Content and every individual piece of it”.

No comparable language appeared in any prior version, the complaint says. It argues the amendments “are themselves evidence” that the earlier terms “did not convey the rights Defendants exercised”.

What the old terms said about deletion

The complaint quotes section 8(a) on what happens when a creator removes something. The licence “terminate[s] once you delete such User Content from the Twitch Services, or generally by closing your account”. AI training does not appear among the uses that survive deletion, the complaint says.

It points at two more of Twitch’s own rules. Section 7 bars “use of any data mining, robots, or similar data gathering or extraction methods”. Section 10 bars access “by means of any robot, spider, scraper, crawler, or other automated means”.

What the privacy notice said before

The privacy notice last modified on 15 January 2026 listed the uses Twitch made of user information. None of them disclosed AI training, the complaint says.

On 12 August Twitch added new wording, according to the complaint. It says the company uses information “to test and improve technology, including machine learning and artificial intelligence”. A further line covers “using data to develop or deploy our generative AI models and services”.

Two promises appear in both versions. Twitch promised not to make its policies “less protective of personal information collected in the past” without consent. It also promised users notice before sharing their information with third parties.

Who the plaintiff is

Warren Pandiscia lives in Fairfield County, Connecticut. He has streamed on Twitch for close to ten years, the complaint says. It puts his channel at more than 900 followers and his output at over 1,000 hours.

He has spent tens of thousands of dollars on equipment, props and subscriptions, according to the filing. It describes his content as building LEGO sets and playing video games. He opted out after learning the setting existed. Almeida Law Group and Sterlington represent him.

What Mike Minton said

The complaint quotes Twitch’s chief product officer twice. He defended enrolling everyone by default in a live session on 12 August. “If it was opt-in, nobody would opt in. That’s honestly the answer.”

Users asked whether their content had already gone into training. He answered: “I don’t actually know the answer to that question because I don’t know what Amazon has done in terms of model training and what they’ve used and not used.”

The complaint also reaches back to 2024, when Minton was chief monetization officer. At an event hosted by The Information he was asked whether Amazon uses Twitch to train AI models. He answered “Yeah, for sure”. He put the use “in a prototyping, not in any kind of production scale, capacity”.

How the setting works

The setting covers a channel, not a person. The complaint quotes Twitch’s help pages: “If you chat on someone else’s stream, their opt-out preferences govern if that chat can be used for training.”

Twitch gave no advance notice, the complaint says. It sent no email and displayed no pop-up. A reporter, Zach Bussey, found the setting. Users have reported the toggle reverting to enabled after they switched it off. The desk covered the opt-out setting on 13 August.

The four claims

Pandiscia pleads breach of implied contract, unjust enrichment, breach of express contract, and violation of California’s Unfair Competition Law at Business and Professions Code section 17200. The implied contract count names Twitch alone. The other three name both companies.

The filing pleads no copyright count, and it says why. Twitch streams and videos “are not registered with the US Copyright Office”. That lack of registration “does not render them valueless or leave them unprotected”, it says.

The California Invasion of Privacy Act appears only as a predicate for the unfair competition claim. It is not a count of its own.

The proposed class

The complaint defines the class as “All content creators who used the Twitch platform at twitch.tv and whose content was used to train Amazon’s AI products without consent”. It sets no start date and no geographic limit.

The class “likely consists of millions of individuals”, the filing says. It puts the amount in controversy above $5mn. That figure is the threshold for federal jurisdiction under the Class Action Fairness Act, not a sum Pandiscia claims.

What the complaint asks for

The prayer for relief seeks class certification, declaratory and injunctive relief, damages, restitution and disgorgement of profits, plus interest and costs. It names no dollar figure.

A more specific demand sits inside the unfair competition count. It asks the court to order both companies to “identify, segregate, and delete content, communications, and data unlawfully obtained”. It also asks them to stop using training corpora and models “to the extent derived from such material”.

What it says about Amazon’s other training data

The complaint names one Amazon model. Nova Reel is a text-to-video system sold through Amazon Bedrock. The filing cites a March 2024 Amazon research paper, VidLA, in which Amazon researchers described using 20 million YouTube videos to build a corpus of roughly 800 million clips.

It does not allege that any Twitch stream went into Nova Reel. On that point it pleads only that Twitch content sits “among the so-called ‘proprietary data’”, on information and belief. It makes no claim that any AI output reproduced Pandiscia’s work. Scraped training data surfaced in a hack at Suno in July.

What it says the content is worth

The complaint reaches for licensing comparators rather than a valuation of its own. Google negotiated a fee of $60mn a year to train on Reddit posts, it says. It puts Reddit’s AI data-licensing revenue above $200mn.

Amazon bought Twitch in 2014 for about $970mn. The filing puts Twitch above 240 million monthly active users in 2026. Between 3.2 million and 6.9 million creators go live each month.

The pattern the filing sits in

Default-on data settings have drawn several rounds of this. HubSpot reversed a customer-data feature after a four-day revolt in July. Cursor shipped a code platform last week without publishing its data terms.

Courts have started ruling on the underlying question. A judge dismissed Google’s scraping suit against SerpApi in July.

What happens next

The docket shows no judge assigned. Twitch and Amazon have not filed a response. Neither company commented when Courthouse News asked. The complaint attaches no exhibits.