YouTube is about to let creators upload up to three different cuts of the same video and run them against real audiences. The winner is the one that holds attention longest. Thumbnail and title testing was one thing. Testing the cut itself is another.
Digiday’s Alyssa Mercante caught the mood yesterday: the creator economy spent the start of the year shouting down AI slop and Sora-style generation, then quietly folded AI into almost every layer of the job — CapCut, Firefly, analytics, scheduling. Rejecting AI-generated content was never the same as rejecting AI tooling. The blurry middle is where the new fights are.
YouTube’s Made on YouTube suite puts a conversational editing assistant in Shorts and Create, and Studio’s Ask Studio can coach pacing and hooks. The video A/B test is the sharper edge. Marques Brownlee called it the platform’s “craziest, riskiest, and possibly worst idea yet.” His worry isn’t sci-fi. It’s practical. Viewers may not know which cut they’re in. Timestamped comments break across versions. Someone watches cut A, comes back, and cut B won. The unspoken deal — we’re all watching the same video — gets a little softer.
And here’s the line that stuck: the tool doesn’t help you make a better video. It helps you find the version with the best retention. Those are not the same thing. (Side bar with me, the human part. This is no shit folks. Retention is NOT the same thing, IMHO. I make what I consider some damn good (and funny) videos on my tube channels (Couch Survivor for example). I can’t get views for nuttin because I just suck at the algo game.) Digiday’s sources are already talking about sameness — hooks that converge, “AI speak,” nuance draining out because the metric rewarded the average.
Been there, in a smaller way. Every platform that scores you eventually trains you.
Circle is being built as the opposite pressure. GAPOZ Press’s gated workshop isn’t optimizing for watch-time graphs. Members request assignments, write structured reviews of text or video, and earn credit toward having their own work reviewed. The unit of value is a named human judgment, not a retention curve. When AI shows up — and it will — disclosure is by role: what the tool did, what the human kept, who is accountable. Self-attested. No detector gates pretending a percentage is a verdict.
That matters more when the big platforms are teaching creators that the “right” cut is the one the model said kept people hooked. A workshop that asks who decided and what they kept is a different kind of infrastructure. The people’s publisher doesn’t need every video to flatten toward the same opening three seconds. It needs people who will say out loud what they used, what they meant, and what they stand behind.
Retention is a useful number. It is not taste. And no shit — the tools that confuse the two are already shipping.
Also on the beat
- SPUR releases a Content Telemetry standard so publishers can track when AI tools retrieve, ground, cite, show, and engage their work — and has invited OpenAI, Anthropic, Google, Meta, and Microsoft onto an AI Licensing Advisory Board (journalism.co.uk).
- Vulture on the Anthropic settlement windfall: publishers are set for big checks (Bloomsbury alone ~$21M), while many authors are fighting 100% publisher claims and rights-reverted titles.
- Pixability launches YouTube “AI slop” filtering for advertisers, pairing agentic channel analysis with Famous Birthdays’ human-creator verification data.
- Google is reportedly piloting payments to ~100 publishers for content used in AI Overviews, AI Mode, and Gemini.
- SCBWI’s updated GenAI policy bans generative AI in awards and showcases, requires a no-AI declaration, may ask finalists for process evidence, and says it will not use AI detectors.
How this was made
AI assisted research and a first draft of this Circle View. A human reviewed, edited, and is accountable for the judgment and the final take. Disclosure by role — not undetectable.
