You're actually surprisingly right. An AI could be trained to show varied enough content to avoid polarisation. E.g.: show videos of consequences of gun-acciden

You're actually surprisingly right. An AI could be trained to show varied enough content to avoid polarisation. E.g.: show videos of consequences of gun-accidents to gun lovers. Or lots of less obvious examples, of course. An AI could actually even determine how slowly it should recommend varied content. The problem is that if an organisation starts doing this, people will ask the opposite: "to be unbiased, an algorithm should show anti-black campaigns to BLM-supporters". Not that I agree with this sentiment, I'm just pointing out people will use this argument claiming anything else is biased. I do think you're onto something though; your comment is underrated.