Amazon and its streaming subsidiary Twitch are facing a class action lawsuit in the United States. It was filed last Thursday by the US streamer Warren Pandiscia at the US District Court for the Northern District of California. The allegation: the company used millions of videos from the platform as training material for generative AI models without acquiring licences or compensating the creators.
The 37-page complaint states that the defendants accessed the streams and videos in order to use them “as a massive dataset” instead of negotiating licences or asking for permission. Pandiscia argues that Amazon’s systems are designed in a way that makes it technically impossible to obtain consent from everyone involved in a broadcast. That includes guests, viewers in the chat and other people who appear in streams without owning the channel themselves
Streams, Clips, Chats and Channel Images
The conflict was triggered by a change Twitch put into effect in mid-August. Since then, content on the platform can be drawn on to train Amazon’s generative models. According to Twitch’s FAQ, this covers live streams, past broadcasts (VODs), clips and highlights, stream chats as well as channel images and text. The company names improved speech-to-text models and better captions as possible use cases, which could also show up in other Amazon services. Amazon is working on generative models for text, audio, images and video.
The arrangement is set up as an opt-out, which means every channel is included automatically unless its owner switches off the “Training for generative AI” setting under security and privacy. Anyone who does so protects only their own channel: if a person appears in a stream whose owner leaves the feature enabled, their own objection carries no weight there.
A remark by Twitch chief product officer Mike Minton added to the debate. In an official patch notes livestream, he justified the decision against an opt-in with the expectation that hardly anyone would actively agree. The quote, “if it was opt-in, nobody would opt in,” circulated widely in the streaming community and appears in several reports on the lawsuit.
Material From Before the Switch Is Also at Issue
One central point of the complaint concerns the time frame. Pandiscia argues that Twitch analysed video material going back to 2024, a period when the opt-out option did not yet exist. That raises the question of whether a setting introduced after the fact can legitimise use that took place beforehand.
Legally, the suit rests on breach of express contract, breach of implied contract, unjust enrichment and unfair business practices restitution, disgorgement of profits and injunctive relief that would include removing the material from the training datasets. The proposed class covers all streamers whose content ended up in AI training without their consent. Amazon and Twitch have so far made no public statement on the allegations
Part of a Broader Wave of Litigation
The case joins a growing number of proceedings around training data. A YouTuber has already taken Nvidia to court over the alleged use of his content for AI training. Anthropic settled a case involving books for a payment of 1.5 billion dollars, while Meta won a copyright suit over books used for AI training. How US courts assess the fair use question in the context of AI training remains inconsistent.
What makes the Twitch case distinctive is that a platform is passing its own community’s content on to its parent company, which uses it to build commercial AI products. Should the court adopt the reasoning of the complaint, that would send a signal to every platform that taps user-generated content for its own AI training by way of default settings.
Aus Datenschutz-Gründen ist dieser Inhalt ausgeblendet. Die Einbettung von externen Inhalten kann in den Datenschutz-Einstellungen aktiviert werden:
