The AI watched the video. That doesn't make it true.

"What happens in this video?" and "is what this video says true?" sound similar and are completely different jobs. CreatorBase has a model that genuinely watches YouTube videos — beat by beat, with timestamps, hook construction, cut rhythm, on-screen text. It's authoritative about what's on screen. It is deliberately treated as zero evidence that the story being told is accurate.

Say a competitor's video is blowing up around a claim — a statistic, a study, a story. The perception tool can tell you exactly how the video is built and why it holds attention. If you repeat the claim because "the AI analyzed it," and the claim is wrong, that's your face on the correction. Perception is not verification, so CreatorBase routes the second question to a different tool: research that comes back with sources you can check, and a workspace that keeps a ledger of which claims were verified and which search tools actually ran.

This is also why CreatorBase runs many models instead of one: the model that watches video, the one that researches with citations, the one that searches X live, and the one that writes in your voice are picked per job. Never because one model is "smarter" — because each question deserves the tool that can actually answer it. When you're the one on camera, that split is the difference between a take and a retraction.