A Connecticut streamer filed a nationwide class action against Twitch and its parent on August 20, 2026, arguing that a setting switched on for every channel cannot supply the consent Amazon needed to feed live broadcasts, recorded videos and chat logs into its generative models.

Warren Pandiscia, a Twitch broadcaster of roughly ten years with more than 900 followers and over 1,000 hours streamed, filed the complaint in the United States District Court for the Northern District of California on August 20, 2026. Docketed as 3:26-cv-08721 in the San Francisco Division, it names Twitch Interactive, Inc. and Amazon.com, Inc. as defendants and demands a jury trial.

The complaint runs 37 pages and asserts four causes of action: breach of implied contract and of the implied covenant of good faith and fair dealing; unjust enrichment, pleaded in the alternative; breach of express contract; and violation of California's Unfair Competition Law at Business and Professions Code section 17200. Jurisdiction is invoked under the Class Action Fairness Act, with the amount in controversy stated to exceed $5,000,000 and the putative class to exceed 100 members. Numerosity is pleaded more bluntly than that: the class, according to the filing, likely runs to millions of individuals. Counsel are Victor J. Sandoval of Almeida Law Group LLC and Arturo Peña Miranda of Sterlington, PLLC.

What Twitch changed on August 12

The dispute traces to a single day. On August 12, 2026, the official Twitch support account posted that a setting had been added allowing creators to opt out of having channel content used to train generative AI models across Amazon. There was no email, no in-product pop-up and no advance notice, according to the complaint, which credits the discovery of the setting to reporter Zach Bussey.

The control sits in account settings, under a section labelled Security and Privacy, as an option named Training for Generative AI. It ships enabled. Twitch support documentation, as quoted in the filing, describes the eligible material as a channel's streams, videos on demand, clips, stream chats, and the pictures and text on the channel page. The company's own FAQ offers a worked example in which creator audio helps refine speech-to-text models that improve captions on Twitch and elsewhere across Amazon.

Two structural features of that setting carry most of the legal weight. The first is scope: the toggle is channel-scoped rather than person-scoped. Twitch help documentation states that where a person chats on someone else's stream, the channel owner's opt-out preference governs whether that chat can be used for training. A viewer who switches the setting off on a personal channel therefore does nothing to protect messages typed in another broadcaster's chat room. The complaint argues that this makes all-party consent structurally unobtainable, since the systems are not built to ask the participants whose messages are captured. The second is reliability: users have reported the toggle reverting to the enabled position after being switched off, a claim the complaint records without resolving.

Twitch Chief Product Officer Mike Minton addressed the default during a live question-and-answer session. Asked why every account was enrolled automatically, he said: "If it was opt-in, nobody would opt in." Pressed on whether content had already been used before the setting existed, Minton said he did not know what Amazon had done in terms of model training, or what had and had not been used.

The 2024 acknowledgment

That answer sits awkwardly beside a public statement Minton made two years earlier. In 2024, when he held the title of Chief Monetization Officer, Minton appeared at an event hosted by The Information and was asked directly whether Amazon uses Twitch to train AI models. His answer, as reproduced in the complaint: "Yeah, for sure." He went on to frame the practice as bounded by user trust and by privacy regulations that vary internationally, and described the company as having a role to play in that. He also characterised the use at the time as prototyping rather than production-scale.

The complaint treats those remarks as the hinge of its willfulness theory. If defendants understood in 2024 that AI training on creator content was constrained by privacy law and by user trust, and proceeded for more than two further years without amending the Terms of Service or the Privacy Notice to disclose it, then the conduct that followed was knowing rather than inadvertent. The pleading also invokes the delayed discovery rule to toll any applicable limitations period, on the basis that creators had no way of learning what was happening.

The launch-day contract amendments

The most technically specific portion of the complaint concerns what Twitch did to its own governing documents on August 12, the same date the opt-out appeared.

Section 8(a)(i) of the Terms of Service grants Twitch and its sub-licensees rights to use, reproduce, modify, adapt and create derivative works from user content. Throughout the period from October 27, 2023 to August 11, 2026, that grant carried an express statement of purpose tying the exercise of those rights to "monetizing the Twitch Services." The August 12 amendment replaced that phrase with a reference to "Twitch and/or its affiliates' business."

The same amendment added a recital with no counterpart in any prior version, providing that by streaming or otherwise making content available, a user is directing Twitch to cache and store that content and every individual piece of it, itemised as each video clip, audio segment, image or text element, and to use any automated tools enabled on the account to generate and store additional content.

The Privacy Notice moved on the same day. The version in force immediately beforehand, last modified January 15, 2026, enumerated the uses to which Twitch put user information: operating and developing the service, security, account management, content and advertising recommendations, participation features, fraud prevention and understanding usage trends. Generative AI training appeared nowhere on that list. The August 12 revision added, for the first time, language about testing and improving technology including machine learning and artificial intelligence, and stated that this includes using data to develop or deploy generative AI models and services. It linked to an Amazon document titled Generative AI Development Disclosure.

Counsel checked that link on August 20, 2026 and found it broken. The complaint reproduces the resulting Amazon error page, which carries a stock apology and a photograph of a golden retriever named Brandi.

The pleading turns the amendments into evidence against the defendants. Had the operative license already conveyed the right to train Amazon's models, no expansion of the purpose clause and no newly minted recital of user direction would have been required. The opt-out is deployed the same way: a license already conveying the right would render the option meaningless.

Two commitments present in both the pre- and post-amendment Privacy Notice anchor the express contract count. One is a non-retroactivity covenant under which Twitch undertakes not to make its policies materially less protective of information collected under a prior version without the consent of affected individuals. The other is an advance-notice promise of an opportunity to decline before personal information is shared with third parties. The complaint alleges both were broken.

Twitch's own rules cut the other way

A subsidiary argument runs on the prohibitions Twitch imposes on everyone else. Section 7 forbids data mining, robots and similar extraction methods, along with downloading any portion of the service beyond page caching; Section 10 bars access by robot, spider, scraper, crawler or other automated means for any purpose. Creators, the complaint contends, reasonably read those clauses as a representation that bulk copying of platform content was impermissible.

Section 8(a) also provides that rights in live and pre-recorded audiovisual works terminate when a creator deletes the content or closes the account, subject to enumerated exceptions that do not include AI training. Once material enters a training corpus, the complaint argues, that termination right becomes unenforceable in practice.

The transmission theory

Sections D and E of the complaint read as interception pleading, even though no standalone wiretap count appears in the caption. The filing traces the Real-Time Message Protocol path from a broadcaster's encoder to a Twitch ingest server, cites Twitch engineering material on real-time transcoding and a high-throughput distribution control plane, and describes chat through Twitch developer documentation as a real-time system in which every message passes through Twitch servers before reaching connected clients.

From that description the complaint builds two propositions. First, that acquisition of live broadcasts and stream chats occurs contemporaneously with transmission, while the communications are in transit. Second, that Amazon is a legally separate entity and not a party to the communications exchanged between creators and viewers, which places it outside the participant exception that ordinarily defeats interception claims.

Those propositions surface in the Unfair Competition Law count rather than as an independent cause of action. The unlawful prong of section 17200 borrows predicates from elsewhere, and the complaint nominates the California Invasion of Privacy Act at Penal Code section 631, contractual rights, and Article I, Section 1 of the California Constitution. The structure carries wiretap-style allegations without pleading a wiretap count, which leaves the statutory damages formula of $5,000 per violation off the table for now.

Section 19(e) of the Terms of Service does the venue work. Twitch selected California law and either the Northern District of California or a San Francisco County state court for every dispute with every user nationwide. The complaint quotes that clause back at the defendants.

Why Amazon wanted the video

The commercial theory rests on Nova Reel, Amazon's text-to-video and image-to-video model. Developed by Amazon AGI and launched publicly in December 2024 as part of the Nova family of foundation models, it is sold through Amazon Bedrock on a per-second-of-generated-video basis. Published Amazon documentation for the model describes pre-training on curated data from licensed and proprietary sources, open-source datasets and publicly available data. The complaint alleges, on information and belief, that Twitch material sits inside the proprietary category.

Two supporting exhibits are drawn from Amazon's own record. A March 2024 paper by Amazon researchers, VidLA: Video-Language Alignment at Scale, discloses the use of 20 million videos drawn from the YT-Temporal-1B dataset to construct a proprietary corpus of roughly 800 million clips named YT-VidLA-800M. Separately, the complaint cites a first amended complaint in Ted Entertainment, Inc. v. Amazon.com, Inc. in the Western District of Washington, filed June 30, 2026, which alleges that Amazon obtained millions of YouTube videos for Nova Reel training by deploying tools designed to defeat blocking, verification and access-token systems. The contrast the pleading draws is that at Twitch, no circumvention was necessary, because the defendants controlled the platform and held the content.

Twitch streams and videos are not registered with the United States Copyright Office, which forecloses statutory copyright claims. The complaint responds that absence of registration does not render the works valueless, and pivots to contract and restitution instead.

For the unjust enrichment count, the benefit is quantified by reference to what the market charges. Google is reported to pay Reddit roughly $60 million a year for training rights, an arrangement PPC Land traced through Reddit's search rise and Gemini training role, and Reddit has reported more than $200 million in AI data-licensing revenue including a comparable deal with OpenAI. Against those figures, the complaint argues, the defendants obtained for nothing what competitors pay tens of millions of dollars per year to acquire.

Two internal discrepancies

The filing contains contradictions worth recording. Paragraph 31 identifies Twitch as a Delaware corporation headquartered at 350 Bush Street in San Francisco. Paragraph 102, in the choice-of-law section, states that Twitch is incorporated under California law. Both cannot hold. The heading above paragraph 31 also names the entity "Twitch Integrated, Inc.", while the caption and the body throughout use Twitch Interactive, Inc. Neither point is likely to prove dispositive, but both are the kind of drafting inconsistency that surfaces in early motion practice.

Platform scale and the audience at issue

By 2026, according to the complaint, Twitch carried more than 240 million monthly active users and roughly 26 to 30 million daily visitors. Between 3.2 million and 6.9 million unique creators go live each month, and average concurrent viewership sits between 2.1 and 2.37 million. Amazon acquired the platform in 2014 for approximately $970 million in cash, when Twitch drew more than 55 million unique monthly visitors watching over 15 billion minutes of content from more than a million broadcasters.

The proposed class covers all content creators who used twitch.tv and whose content was used to train Amazon's AI products without consent, excluding judges, the defendants and their affiliates, and counsel on both sides.

Why this matters for advertisers and platforms

Twitch inventory is not a separate advertising product. All Twitch advertising is transacted through Amazon DSP, a structure PPC Land documented in 2023 and which Amazon has since consolidated further alongside Fire TV, Prime Video and Alexa inventory. Any dispute that touches creator relations on Twitch therefore touches a supply source inside an advertising business that reported $19.8 billion in the second quarter of 2026, up 26% year over year.

The complaint also lands inside an accumulating body of litigation over consent architecture in AI products. A note-taking product was accused of recording participants without all-party consent while training models on the results, following a structurally similar 2025 case against Otter.ai over meeting recordings used for machine learning. A federal jury found in August 2025 that Meta violated the California Invasion of Privacy Act by collecting health data through an embedded SDK, and complaints filed in May 2026 applied the same statutes to tracking infrastructure inside ChatGPT.

European regulation points the same way. The European Data Protection Board adopted guidelines on July 7, 2026 concluding that consent will most likely not work as a legal basis for training-scale data collection, and that publishing material on an openly accessible page does not amount to agreeing to its use for model training. A French competition authority opinion in February 2026 found that 80 percent of video content creators hold weak or very weak bargaining power against the platforms they depend on, with Twitch among those examined.

The default-on pattern is familiar too. Google's Gemini app activated expanded Android permissions automatically in July 2025, and Meta's auto-enrollment of advertisers into generative creative tools drew criticism at Cannes Lions in June 2026 because it was opt-out rather than opt-in.

What distinguishes the Pandiscia complaint from the scraping cases is the absence of any perimeter to breach. Where an outside developer must circumvent a platform's controls to obtain training material, the defendants here are alleged to have held the content already, changed a default, and disclosed the arrangement afterwards. The question the case puts to the court is whether a platform's own license, drafted by the platform and amended on the day of disclosure, can retroactively supply the authorization the platform never asked for.

Neither defendant had filed a response as of the filing date, and no hearing has been scheduled.

Timeline

  • 2014 - Amazon acquires Twitch for approximately $970 million in cash
  • October 27, 2023 - Version of the Twitch Terms of Service takes effect tying the Section 8(a)(i) content license to monetizing the Twitch Services
  • March 2024 - Amazon researchers publish VidLA, disclosing use of 20 million YouTube videos to build an 800 million clip training corpus
  • 2024 - Mike Minton, then Chief Monetization Officer, confirms at an event hosted by The Information that Amazon uses Twitch content for AI development
  • December 2024 - Amazon launches the Nova family of foundation models, including Nova Reel
  • January 15, 2026 - Last modification of the Twitch Privacy Notice before the AI training disclosure
  • June 30, 2026 - First amended complaint filed in Ted Entertainment, Inc. v. Amazon.com, Inc. in the Western District of Washington
  • July 7, 2026 - European Data Protection Board adopts guidelines rejecting consent as a workable basis for training-scale scraping
  • July 30, 2026 - Amazon reports second-quarter advertising revenue of $19.8 billion, up 26%
  • August 11, 2026 - Final day of the prior Terms of Service and Privacy Notice
  • August 12, 2026 - Twitch adds the Training for Generative AI setting, enabled by default, and amends both the Terms of Service and the Privacy Notice
  • August 19, 2026 - Date on which counsel last accessed several Twitch developer and legal pages cited in the complaint
  • August 20, 2026 - Counsel finds the Privacy Notice link to Amazon's Generative AI Development Disclosure broken; class action complaint filed in the Northern District of California as case 3:26-cv-08721

Summary

Who: Warren Pandiscia, a Connecticut-based Twitch content creator, filed against Twitch Interactive, Inc. and Amazon.com, Inc., represented by Almeida Law Group LLC and Sterlington, PLLC. Twitch Chief Product Officer Mike Minton is quoted throughout the complaint.

What: A nationwide class action complaint asserting breach of implied contract and the implied covenant of good faith and fair dealing, unjust enrichment, breach of express contract, and violation of California's Unfair Competition Law, over the use of streams, videos on demand, clips, chat logs and channel imagery to train Amazon generative AI models. The pleading seeks class certification, declaratory and injunctive relief, damages, restitution and disgorgement, with a jury demanded.

When: Filed August 20, 2026. The Training for Generative AI setting and the amendments to both the Terms of Service and the Privacy Notice date to August 12, 2026. The complaint traces undisclosed use back to at least 2024.

Where: United States District Court for the Northern District of California, San Francisco Division, case number 3:26-cv-08721. Twitch is headquartered at 350 Bush Street in San Francisco; Amazon at 410 Terry Avenue North in Seattle.

Why: The complaint argues that a setting enabled by default cannot constitute consent, that the content license in force through August 11, 2026 authorised use only in connection with monetizing the Twitch Services, and that amending both governing documents on the day of disclosure demonstrates the prior text did not cover the conduct. For advertisers, the case tests whether creator content held by a platform can be moved into a parent company's commercial AI pipeline without a licence negotiation, at a moment when comparable rights trade for tens of millions of dollars per year.