Google Says AI Better Than Humans At Scrubbing Extremist YouTube Content (theguardian.com)

An anonymous reader quotes a report from The Guardian: Google has pledged to continue developing advanced programs using machine learning to combat the rise of extremist content, after it found that it was both faster and more accurate than humans in scrubbing illicit content from YouTube. The company is using machine learning along with human reviewers as part of a mutli-pronged approach to tackle the spread of extremist and controversial videos across YouTube, which also includes tougher standards for videos and the recruitment of more experts to flag content in need of review. A YouTube spokesperson said: "While these tools aren't perfect, and aren't right for every setting, in many cases our systems have proven more accurate than humans at flagging videos that need to be removed. Our initial use of machine learning has more than doubled both the number of videos we've removed for violent extremism, as well as the rate at which we've taken this kind of content down. Over 75% of the videos we've removed for violent extremism over the past month were taken down before receiving a single human flag."

Comments

Didn't we already have a post about training AI's?

[Anonymous Coward]

Who is training the AI and deciding what is illicit? Better by what measure?

Too little too late

[rsilvergun]

the advertisers pulled their ads because of some alt-righters getting a few ads. Everybody pulled them and then they noticed there was no appreciable drop in business / sales / brand recognition from the lost advertisements. That means google's pretty well boned. It's also why P&G just announced they're dropping $100 mil in digital advertising. They know it doesn't work...

Sad thing is I'm gonna miss the ad supported internet. I'm pretty good about ignoring ads (and will cheerfully click on any ad that annoys me to give 'em a false positive) so it never bothered me.

Re:Too little too late

[Applehu Akbar]

Advertisers could have saved themselves by agreeing to a non-offending standard (no popups, no autoplay videos, etc.) for advertising that would then be accepted by ad blockers. But no - now their business model is vanishing.

Maybe

[onyxruby]

Did the AI get programmed to only target politically incorrect content? Unfortunately double standards for YouTube, Twitter and other sites are all too common. Color me skeptical that this is anything other than an automated political censorship tool.

Re:Maybe

[AmiMoJo]

It's more like just "political content". Lots of people on the left are getting their videos taken down too, mostly due to people maliciously flagging them. Shaun & Jen, H. Bomberguy, Contrapoints... Some people have lost entire channels.

https://youtu.be/UgNhO8lMINw

How is accuracy measured?

[djinn6]

How does one know if AI is more accurate than humans? If the AI says the video is extremist, and the human says it's not, then who's right? Is there another machine out there that gets to be the arbiter?

The real answer is that there is some human who decided whether a video is extremist, then gave it to both the AI and another human, and the AI was able to learn how to agree better with the 1st human. Unfortunately, that doesn't actually tell us if the video is extremist or if the AI is any good, because the first human isn't more right than the second.

Re:How is accuracy measured?

[Kjella]

I would imagine that Google has some content policy experts who are making up the rules of what's acceptable and not. And then a whole lot of whack-a-mole workers trying to police by those rules. I guess that doesn't say if the expert is right, but the AI is doing the leg work more consistently in line with policy than the people. If you let say 5% through to a second opinion by humans or if there's some sort of appeal process there should be a pretty continuous learning. Though I suspect the AI will completely fail to understand parody or other use of extremist elements in a non-extremist way.

Re:Didn't we already have a post about training AI

[epyT-R]

As long as it's not muslim of course. That would be 'racist' and 'islamophobic.'

I also wonder if it allows 'extremist' content in the same vein as Galileo's observations about the solar system. Somehow I think this will end up as a simple anti popularity filter.

Re:"Controversial" = "Not SJW"

[ArylAkamov]

I don't see how this is complaining about not getting special treatment. If anything it is complaining about not getting equal treatment.

Re:Didn't we already have a post about training AI

[russotto]

I believe the training set is curated by a group headed by Coraline Ada, Christopher Poole, Deray McKesson, and Linda Sarsour.

Re:Didn't we already have a post about training AI

[AmiMoJo]

4chan is training their AI. People on the /pol board have been running a mass false-flagging campaign lately, getting videos and channels they don't like taken down. When the victims appeal or complain they just get a robot "reviewing" the strike and ultimately have to resort to tweeting at YouTube staff to have any chance of getting things resolved.

YouTube's content filtering is out of control, anti-free speech, kafkaesque and devoid of any human oversight to catch errors.

Re:"Controversial" = "Not SJW"

[AmiMoJo]

Plenty of progressive/left leaning videos get taken down too. 4chan users have been flagging them like mad lately, and some videos and even entire channels have been removed because of it.

It's not some political conspiracy by YouTube, it's just a shitty system that is wide open to abuse and utterly devoid of human oversight.