> You're basically giving people a playbook for how to game your system.
You aren't wrong, but accepting secrecy for this reason is a flawed premise. As it assumes that giving YouTube arbitrary and oversight free power over what people view and what it's creators make is better than trying to prevent an open rule set from being gamed.
We learned this was a bad idea in government, food, construction, vehicle safety, etc a long long time ago. Software is just overdue for it's own set of transparency rules.
Those are 2 separate and independent things.
Regardless, they are not controlling what creators make or view. Creators can make whatever they want and users can choose to watch whatever they want. YouTube also explicitly says when a video is recommended. Also note that this Youtube's platform.
Comparing this to the other things you list is also not a great comparison. It's quite a reach actually. For instance, not knowing what was in food is different than not knowing the formula.
Are you claiming recommendations do not affect views and views do not affect what content people create? This is obviously false.
> For instance, not knowing what was in food is different than not knowing the formula.
This is a great comparison actually, we all consume Youtube's recommendations without knowing either what was taken into account (the contents) or the exact weights of these things (the formula).
Recommendations for people I subscribe to are great, but don't need much magic.
Recommendations for creators I've liked or viewed many videos of before are generally good.
Recommendations for videos with tags that I tend to like are hit and miss.
Other recommendations are usually low quality garbage, in my experience.
I suspect their recommendation algorithm is overly complicated because it optimizes for ad revenue rather than value to viewer. I think a simple algorithm weighting viewer->creator score and creator->other creator score could produce higher quality recommendations than their current system. Maybe add a dash of keyword to the final sort to allow videos on the same topics to be grouped together.
That won't stop other actors from monitoring recommendations / search results and observing how they change. With a bit of ML they could train a model on that data and obtain all the answers they want. The information is hard to observe up close but easy to extract when you collect lots of observations.