[spartanatreyu]

The problem is that one wrong video leads to a cascade of bad suggestions.

Opening one video that looks legit but is really AI noise leads to your recommendations filling with other AI noise.

Opening one video with a completely misrepresented title leads to your recommendations filling with attention-bait, rage-bait, etc...

[riku_iki]

If I would be youtube ML engineer, I would use some engagement metric as a label in loss function: for example how many % of video you actually watched, not just fact that you clicked on it.