Hacker News Reader

Comment by Havoc | Hacker News Reader

I really don’t get why there isn’t better automated filtering.

There can’t be that many suicide vids etc floating around the internet. Catalog that sht once and use AI similar to YouTube and copyright detection.

Sure there will remain a continued need for 2nd line humans but this really shouldn’t be this much of a mental sweatshop operation.

Vids especially.

Written text is a bit harder because anyone can easily produce unique content

Also FB being front and center on this issue seems to be a recurring theme…yet they aren’t even the big video destination

View on HN

I promise, if FB/other companies could automate this away, they 100% would.

In general, it's easier for a computer vision system to recognize and filter a video that has already been banned--though, there is a constant arms race here as well--than it is for it to judge the content of a completely new video. That means that, for a huge number of cases, a human being will have to see the footage at least once.

EDIT: To be clear, I'm not taking up for FB in this situation. I'm specifically clarifying the difficulty of using ML/DL in moderation systems.

> Written text is a bit harder because anyone can easily produce unique content

And video is somehow less unique, and easier for AI to categorise and identify?

> There can’t be that many suicide vids etc floating around the internet.

So very little training material for the AI, then?

> they aren’t even the big video destination

More than 4 billion video views take place on Facebook every day. [0]

I really despise FB. But you're trivializing the problem.

[0] https://99firms.com/blog/facebook-video-statistics/

>And video is somehow less unique

This type of video - yes. You and I are both writing unique content right now. Redditors produce thousands of unique posts daily.

There are only so many suicide videos on the internet though and the amount a person can produce per day is inherently limited.

It follows that fingerprinting said videos should be feasible.

>So very little training material for the AI, then?

And? Youtube does this with a single example for copyright purposes. Not all algos have to be backprop DL

Reply on news.ycombinator.com