It’s been a year since Google implemented YouTube’s video filtering in order to stop the upload of unauthorized clips and eventually monetize its users’ uploads. The folks over at ZDNet decided to cover the anniversary with a special 15-page report called The YouTube File (PDF available for registered users only), in which the system receives a largely sympathetic review from NBC Universal Chief Counsel Rick Cotton. The report has Cotton on record with the assessment that YouTube’s filters are “improving month by month” and now catch some 75-80 percent of all illegal uploads.
Others, however, disagree, among them the vendor of a competing filtering solution that puts the number at closer to 3 percent. Google’s response is that some of those videos may have been legitimate copies that were authorized by content owners, and viewing some of the disputed clips suggests that may in fact be the case.
Video filtering has always been a bit of a touchy subject. Vendors and site owners don’t like to discuss the accuracy of their systems, partly because of liability issues, and partly because, well, every net has its holes. We did an early test of video fingerprinting technologies deployed at both MySpace Video and MSN’s Soapbox platform back in June of 2007, for example, and found that it was easy to upload Viacom-owned Daily Show clips even if we identified them by title.
These filters seem to have improved quite a bit since then. ZDNet’s report is citing the Olympics as one prime example for YouTube’s system, with the site weeding out almost all uploads of official Olympic video coverage. In addition to its proactive filters, it also responded to 25,000 takedown notices from NBC, which had the exclusive rights for Olympic content in the U.S. YouTube was also largely successful in stopping the distribution of the second SNL clip featuring Tina Fey as Sarah Palin, which left some users scrambling to find other ways to get to the clip through their favorite video portal.
YouTube is working with technology from Audible Magic as well as Nexicon to identify and process unlicensed uploads in addition to its internally developed video fingerprinting technology. Some competitors seem to think that they could do a better job. ZDNet’s report references numbers from Mountain View, Calif.-based startup Anvato that claims to have identified more than 1,200 infringing clips on YouTube, only 37 of which have been taken down by Google.
A quick look at the Anvato web site casts some doubt on those claims. Anvato currently showcases part of an episode of the ABC show Eli Stone on its site that aired back in July, and links to dozens of supposedly infringing copies of the clip on YouTube. Problem is, most of the videos on YouTube are actually official 2-minute promo clips for the show that were uploaded months before the episode in question aired, and in some cases even before the show’s pilot aired on ABC. It’s hard to imagine that ABC ordered YouTube to take down these promo videos, and Google is under no obligation to take action against these uploads without an official takedown request.
Of course, all of this leads to another touchy subject that the YouTube File report completely fails to mention: fair use. The McCain campaign recently complained that YouTube had yanked some of its clips off the site because of dubious copyright claims, and there have been numerous other instances in which takedowns have actually been overreaching. Filters, unfortunately, don’t know much about fair use, and higher recognition rates inevitably also lead to more false positives. Maybe 75-80 percent accuracy really is good enough?