Twitch Trained Amazon's AI on Streamers by Default. Now Users Can Say No.

Today’s Focus

Twitch confirmed on Wednesday that content produced on its platform had been used to train artificial intelligence models built by its parent company, Amazon, according to the BBC.

The company said a setting permitting that use had been switched on by default for creators. Users can now turn it off, but the disclosure prompted anger from streamers who questioned why the arrangement existed in the first place.

Twitch chief product officer Mike Minton described the opt-out choice as a form of respect for users, the BBC reported. Pressed on why the default ran the other way, he offered a blunt rationale: “If it’s opt-in, nobody would opt-in. That’s the honest answer.”

Minton said any channel content belonging to users who stay opted in can be pulled for training. He added that the collected data would not be resold to other firms.

During the same livestream, Twitch head of community Mary Kish walked viewers through the steps to disable generative AI training inside channel settings. Switching it off blocks Amazon from using streams, clips, images, chat logs, and other channel material.

Kish told viewers she did not expect a warm reception, saying she did not anticipate anyone reacting favorably. She framed the practice as an industry norm.

The chat filling the livestream ran heavily against the move, according to BBC and coverage aggregated by outlets including TechCrunch, Business Insider, Engadget, and Ars Technica. One viewer wrote that nobody would opt in because nobody wants to feed AI with their creativity and content.

The Debate

Supporters argue

Twitch’s leadership framed the opt-out control as a concession that hands power back to creators. Minton argued the company was being straight with users rather than burying the practice, telling the livestream audience the setting reflects “respecting” them by giving them a way out.

Kish presented AI training on user content as a settled industry practice, not a Twitch quirk, saying data collection for training has become standard across the sector. In that view, a visible toggle is more transparent than the silent scraping many platforms allow.

Supporters of the broader approach note that Twitch pledged the data would not be sold to outside companies, keeping it inside Amazon rather than on an open market. Backers of default data collection also contend that training on real streams could improve tools that creators themselves use, from moderation to captioning. They argue a friction-free default is the only way to gather usable data at scale, since almost no one seeks out an opt-in switch.

Critics argue

Streamers reacting on the livestream and across social platforms rejected the idea that an opt-out amounts to consent. Many argued that the default should have run the other way, and that Minton’s admission that nobody would opt in proved the company knew users did not want it.

One viewer wrote that “nobody wants to feed AI with our creativity and content,” capturing a common complaint that creators’ labor was used without a clear yes. Critics said the practice had run quietly for years before Wednesday’s disclosure, undercutting claims of transparency.

Coverage from TechCrunch and Business Insider highlighted creator frustration that the burden falls on individuals to find and flip a setting, rather than on Amazon to ask first. Detractors also questioned the promise that data will not be resold, noting it does not limit how broadly Amazon itself can use the material. For working streamers, the fear is that their voices, faces, and styles could shape models they neither control nor profit from.

What the experts say

Research on data practices suggests defaults powerfully shape outcomes. Behavioral scientists Eric Johnson and Daniel Goldstein, writing in Science, documented that opt-out systems produce far higher participation than opt-in ones, with organ-donation consent rates near 90 percent in opt-out countries versus roughly 15 percent in opt-in ones.

That gap is the mechanism Minton described out loud. The Pew Research Center has found that most Americans feel they have little control over data companies collect, with about 81 percent saying the risks of corporate data collection outweigh the benefits in a 2019 survey.

The legal ground remains unsettled. Courts are still weighing whether training AI on copyrighted or user-generated material is fair use, with several cases pending against major AI developers. The U.S. Copyright Office, in a 2025 report on generative AI and training, noted that whether such use qualifies as fair use depends heavily on facts specific to each case. No clear federal rule yet governs whether platforms may train on creator content by default.

By the Numbers

Default on: Twitch confirmed the AI-training setting was enabled by default for creators before Wednesday’s opt-out announcement, per the BBC.

5+ content types: Streams, clips, images, chats, and other channel content can be used for training unless a user opts out, according to the BBC.

~90% vs ~15%: Consent rates in opt-out versus opt-in organ-donation systems, per Johnson and Goldstein in Science.

81%: Share of U.S. adults who said the risks of company data collection outweigh the benefits, per a 2019 Pew Research Center survey.

2005: Year Amazon-owned Twitch traces its lineage to, via Justin.tv, before Amazon acquired the platform in 2014.

No resale: Minton said data gathered for training would not be sold to other companies, per the BBC.