I have taken a look at the developer documentation. It appears it works by downloading an AI model to a web browser window (on mobile, the model would be bundled within the app instead). The facial analysis is performed within the window/app, running on your own hardware, and then the estimation result is sent back to the web server. This means they don't necessarily need the footage uploaded to their servers, but it doesn't rule out the possibility that they might do it anyway.
I'm not sure, but the percentage of kids younger than 13 that would know how to manipulate that would be very little, I presume. Small enough to be considered an error margin I guess?