Hacker News new | ask | show | jobs
by skosch 2144 days ago
Radiolab has a wonderfully deep episode on this topic: https://www.wnycstudios.org/podcasts/radiolab/articles/post-...

The main issue is the sheer volume of content. Four seconds – that's their estimate for how much time employees in Facebook's "censorship farm" centers have, on average, to make a decision about whether a reported post gets deleted or not. As a result, management decided to aim for the lowest bar: consistent enforcement of a limited set of rules. That's hard enough as is: is all blood gore? Are all references to race hate speech? Etc. etc.

So the censors simply have no capacity to fairly fact-check every single political post, and when in doubt, they err on the side of free speech. ML can help suppress certain viral posts, but not consistently so. I'm not a fan of Facebook by any means, but in this case they really are in a tough spot.

1 comments

Facebook seems to still have some profit, so it clearly has capacity. The insufficient number of people looking at reported violations is by design.
This assumes that there is a solution 'if you just throw enough money at it'.

I'm not familiar enough with how large scale moderation works, so is this solvable, with near unlimited funds?

It really only assumes the solution gets better if you throw more money at it. And that, frankly, is essentially certain given the fact that they need to pay for human eyeballs. Also, the problems facebook has are somewhat misrepresented often - they don't actually need to moderate everything equally intensively - most of the impact will be there with the things that get shared publicly and are viewed often. Comparing the amount of moderator time with total posting volume is thus misleading, trying to make it appear like a hopeless case. But look at it the other way around, not from the poster's side, but the reader's - clearly reads aren't distributed evently; so it's plausible you might catch 90% of post readungs with just 10% of post writings (I don't know the proportion, but I'd expect it to be heavily skewed).