I came the opposite conclusion as you. Say your friend looks great in a piece of clothing. Based on just this, the next time they ask you what they should buy,

I came the opposite conclusion as you. Say your friend looks great in a piece of clothing. Based on just this, the next time they ask you what they should buy, you suggest the style and brand they looked good in before. A day later, you discover that the brand in question clubs baby seals. Does this suddenly make you liable for all the atrocities the brand commits? Superficially, a YT recommendation is based on metadata - video length? Did the user watch the video beyond threshold values? Did they comment or react to a video? Where do this user’s metrics like in comparison to metrics of other users who watched the same video? I concluded that since the algorithms aren’t moderating the content, just your access to it, they are not afoul of S230. Maybe the answer is to require all social media to provide a strictly timeline based view. Even here, the submission time stamp is purely metadata about the content, as is the watch duration or “comment/react?” flags.