Fight Club: Could Technology Have Stopped The YouTube Cheerleader Beating?

The recent beating of a Florida girl, taped and posted on YouTube, has revived an old debate about Google's video site: Can YouTube be expected to thoroughly vet video uploaded to the site for violence, pornography, or even basic copyright violations?

Not really, is the broad consensus. YouTube tries to screen for some worrisome content, and is working with some copyright holders to try to keep their stuff off the site, or at least pay them for it. But YouTube is too big, the amount of content too vast (10 hours of video uploaded every minute) for the site to keep track of it all. As co-founder Steve Chen told the Sydney Morning Herald: "It's an impossible task for us to manually go through and solve this problem through just manual labour."

There is, of course, a counter argument. We talked to Waikit Lau, CEO of ScanScout, which uses its own technology to screen video on behalf of advertisers. That system could have recognized the YouTube cheerleader beat down, too, he says. Here's how:

First, Lau says, an algorithm would associate the video's keyword tags for clues like the association of "fight" and "blood." Then ScanScout crawls the Web associating keywords to make its technology smarter about what the connections between words mean.

Then ScanScout performs an analysis of audio clips, to search for sounds associated with violence. Fights may not have much dialog, but they do have signature sounds, and the software can learn to recognize them.

Last, the company performs an analysis Lau calls "visual clustering", that can recognize visual signatures associated with fighting. The basic version of this is used to recognize pornography, but it can also be more sophisticated and learn the visual signatures associated with other activities, he says.

No system is perfect, but depending on how well the software is trained, and how sensitive its thresholds are set, the number of videos that need a second look by a human could be reduced to a managable trickle. Human judgement might be required to, say, distinguish between a video made by terrorists and a parody of a video made by terrorists. "That's where the art rather than the science comes in," Lau says. "You are going to have false negatives and false positives. The key is to train the system to catch as many as possible."

Sounds great. Can it work? YouTube and other sites are already trying their own automated filtering systems, but no one seems to have a solution that works well. But the bigger problem may not be a lack of technological solution, but the lack of an incentive to apply one.

For video distributors, playing whack-a-mole with objectionable content works ok, for now -- because there's not much downside to having it up there. That might change if advertisers wanted to start user-generated video, but right now, they don't. "Ultimately, the publishers don't care," Lau says. "They'd rather react to it than proactively manage it."