Twitch CPO: “If this was opt-in, nobody would opt in

Something shifted this month, and it lands squarely on anyone building creator features. Amazon set Twitch to feed streams into AI training by default. A guitar company got caught faking a demo with AI and denied it for two weeks. A movie studio scrubbed a video that showed AI concept art. Meanwhile, an OnlyFans investor bet the whole company on the one thing AI can't fake. Different stories, same fault line: who consented, and can you prove where the work came from. Let me catch you up.

The deep cut

• Consent and provenance are now product decisions. Twitch, D'Addario, and Warner Bros all lost trust on the same two questions.
• Default-on is a tell your users can read. Twitch's CPO admitted opt-in would get zero takers, so they buried the toggle.
• A denial that unravels costs more than the AI did. D'Addario spent two weeks lying before admitting Suno regenerated the track.

The default that says the quiet part

Twitch turned on AI training for every account and made people hunt for the off switch. Buried under Security and Privacy, not the creator dashboard. When streamers asked why it wasn't opt-in, Chief Product Officer Mike Minton gave the honest answer: "If this was opt-in, nobody would opt in."

That line is the whole problem. A default is a bet that your users won't notice or won't bother. When you know the answer would be no, the default becomes a way around consent, not a shortcut through it. Some streamers reported the toggle turning itself back on, which only deepened the sense that the deck was stacked. Others started tagging streams "NoAI" and "AIOptedOut." Your users will read the default. Set it like they're watching.

The cover-up that cost more than the code

D'Addario put out a guitar string demo, people flagged the audio as fake, and the company denied it for nearly two weeks. It floated excuse after excuse: low-quality exports, plugin noise, AI mastering tools. Then a YouTuber got the original Logic files straight from the company and the story fell apart.

The eventual admission was blunt. D'Addario said "We got this wrong" and confirmed Suno Studio regenerated the track. It also apologized for deleting critics' comments and blocking them. The AI use was recoverable. The two weeks of denial and the censored comments were not. Compare that to Saber, whose CEO went further and mocked the writer who accused him, saying "Stella who? I had to ask Claude." That response became the brand's first impression for a lot of people.

When you can't even tell

Warner Bros released a behind-the-scenes video for Supergirl showing a laptop with what looked like AI-generated concept art of Lobo, a character who has existed for 50 years. Then it deleted the video the moment big film accounts started sharing it. The takedown drew more eyes than the clip ever would have.

The pattern repeats on Steam. Saber's Rideshare game shipped a trailer that appeared to use AI writing and voices, with no disclosure on its listing, even though Steam requires it. Provenance is the missing piece across all of these. Nobody could say for sure what was AI and what wasn't, and that uncertainty did the damage on its own. If you ship creative work, you need a paper trail before someone else demands one.

The thing AI can't ship

While these companies got caught, an OnlyFans investor made the opposite bet. James Sagan called it "probably the least-likely platform to be disrupted by AI", because people pay for connection, not content quality. "The scarce asset is authenticity," he said. Most creator revenue comes from real messages between creators and fans.

Worth a grain of salt. Creators already use chatters and AI clones, and a class action over chatter "scams" got dismissed last year. But the strategic read holds. When AI makes content cheap and hard to source, the premium moves to what can be verified as human. That's the same lesson the other three companies learned the expensive way, just told from the winning side.

Three questions for your team

• Where in our product is AI training or generation turned on by default, and would our users say yes if we asked them straight?
• If a customer asked us to prove which parts of a feature or asset were AI-made, could we, and how fast?
• If we got caught tomorrow, is our first instinct to explain or to delete the evidence, and which one does our current process reward?