,
9 tweets,
2 min read
Read on Twitter
No.
First: the story's title disagrees with the content of the story. This is the problem with modern journalism that title editors often put titles on stories totally at odds with the story.
Second, it's an important lesson why automated "detection" doesn't work.
First: the story's title disagrees with the content of the story. This is the problem with modern journalism that title editors often put titles on stories totally at odds with the story.
Second, it's an important lesson why automated "detection" doesn't work.
2/ The story claims detection of guns with a 90% confidence rating. That's meaningless. Is it saying 90% confidence that a video with a gun actually contains a gun? Or is it 90% confidence that a video without a gun doesn't in fact contain a gun?
3/ In other words, I have a product that detects 100% of all terrorist videos. It simply contains one line of code returning "true", triggering on ALL videos. Triggering on bad content is easy, NOT TRIGGERING on GOOD content is the hard, crinkly bits.
4/ We call this "false positives". Facebook has 300 million photos uploaded per day. If the false positive rate of their algorithms is 1%, then that means humans need to analyze 3-million photos per day to make the final judgement.
5/ As the story above points out, detection for "guns" is a complicated problem, because your automated detection will trigger on hunters, shooting competitions, the winter olympics biathlon, and Call of Duty video games.
6/ "Your account has been suspend for posting terrorist videos in violation of our community standards".
Uh, that was a Twitch video of me and Bob playing Call of Duty.
Uh, that was a Twitch video of me and Bob playing Call of Duty.
7/ The underlying premise is that people believe Facebook should do more to censor content they don't like, in violation of our free speech values. It's weird that journalists should be so pro-censorship.
8/ The title of the above article shows the problems we have in journalism, where there are only two allowable tropes:
a) Facebook doesn't do enough to censor bad content
b) Facebook does too much censoring of good content
Either way, Facebook is wrong.
a) Facebook doesn't do enough to censor bad content
b) Facebook does too much censoring of good content
Either way, Facebook is wrong.
9/ We aren't allowed stories explaining the intractability of the problem, the impossible standards Facebook is expected to meet, and that censorship of even "bad" content is not a good idea.
