Twitch is training Amazon's AI on streamers' content by default
A consent-by-default decision at the world's biggest livestream platform, a research result that turns LLM outputs into a privacy leak, and a $2 billion bet that AI implementation is the next big business.
Twitch will feed streamers' content into Amazon's generative AI models unless users dig into their security settings and opt out — a new toggle the platform switched on for every account by default. The setting, announced Wednesday, sits at the bottom of each user's security page and reads "Allow your channel content to train generative AI content models at Amazon." Flipping it off keeps streams, VODs, clips, chat, and channel pictures and text out of "future training" of any Amazon model that generates or synthesizes text, audio, images, or video. Opting out does not cover every AI feature — moderation tools like AutoMod and automated captions keep running, and if you chat on another streamer's channel, that streamer's opt-out choice governs your messages.
The move confirms rumors that circulated for a week and fulfills a plan Twitch has discussed for years: in 2024, Chief Monetization Officer Mike Minton said the platform had a "role to play" in training Amazon's generative AI, with streamer content already used "in a prototyping" capacity. 404 Media notes Amazon did not respond to a request for comment, and Twitch's press email bounced. The pattern is what matters here — the burden lands on users who happen to hear about the change and know where to look, which functionally means most of the platform's creators are opted in whether they intend to be or not.
Researchers at IIT Bombay and Adobe Research have shown that an LLM's prompt can be reconstructed from its output text almost word-for-word — no model weights, no API access, just the generated text. Their method, described in a new paper on arXiv, trains an "inverse language model" entirely from scratch on synthetic data produced by the target model, reversing next-token prediction into previous-token prediction. A single response yields the exact original prompt alongside several semantically equivalent variants: in one example, "How to reach out to competitors to find their pricing strategies?" was recovered verbatim. The inversion also transfers across models — an inverse model trained on a small Qwen-3-0.6B chatbot could capture the meaning of GPT-4o's responses, meaning an attacker would not even need to know which model produced a given output.
For companies, the risk is proprietary system prompts holding trade secrets, moderation rules, or specialized instructions; for individuals, sensitive queries could be extracted from anything they generate. The paper makes no explicit claims about attacking commercial systems, and if the method holds up on production models, labs will need answers fast. It lands in the same security thread as this morning's Attack exposes hidden reasoning traces inside Claude, GPT, Gemini — another way hidden inputs leak out through visible outputs.
Thrive Holdings, the OpenAI-backed firm that buys traditional businesses and installs AI into their workflows, has raised $2 billion at a $12 billion valuation from SoftBank, D1 Capital Partners, and Altimeter Capital. The New York Times DealBook was first to report the round. Thrive is effectively a private equity firm for AI: it has assembled more than 70 companies across Current, its accounting arm with over 50 firms and 2,000 professionals, and Shield, its information technology arm. The firm's numbers are aggressive — its TaxAI agents processed more than 7,000 tax returns at 98% accuracy and cut preparation times by over 30%, while Shield's AI products sped up help desk resolution by 36 times. Part of the new capital will fund a third vertical: regulatory services for physical assets like data centers, manufacturing, and power infrastructure.
OpenAI took an ownership stake in Thrive Holdings in December 2025 and seconded employees to its portfolio companies, part of a broader push into AI implementation — OpenAI and Anthropic have both launched billion-dollar enterprise joint ventures along the same lines. The thesis is that embedding AI into messy, regulated industries is where the real money is, and investors are paying a $12 billion premium to test it.
What to watch: whether creator backlash forces Twitch to reverse the default — and whether other platforms copy the template.
Should platforms have to ask before training on user content, or is an opt-out switch enough? Tell us in the comments.
