YouTube revealed how artificial intelligence helped filter "unwanted" videos

YouTube is full of extreme and misleading videos, and the company has been criticized for not doing enough to limit the damage. But one of the sites that Google has managed to clean is the homepage of this platform. Behind the scenes, Google has introduced an artificial intelligence software that badyzes many video sequences without human help, decrypts problematic clips and blocks them from the start page and the startup screen of the # 39; application. Your internal name is "junk video clbadifier", said sources related to the project to Bloomberg.

The system, which has not been disclosed before, plays a key role in attracting and retaining viewers on the YouTube homepage, creating a foundation for a series of new ads that reach the video service. Google has been testing the clbadifier since 2015, but has widely implemented it after a series of unpleasant incidents with children's videos in 2017, according to a former staff member of YouTube. An episode, nicknamed "Elsagate", featured popular videos showing the Disney Princess in various adverse situations.

YouTube develops "choose your own adventure" programs

That year, YouTube also faced a torrent of advertiser boycotts for inappropriate videos threatening sales. A Google spokeswoman confirmed that the company had a video clbadifier for the YouTube home page, as well as "below" signs recommending other clips. The system badyzes comments from users who report sensational videos that are misleading or only search for clicks. It also takes advantage of other data on public retention, likes and dislikes.

This shows that YouTube is able to regulate the spread of problematic content. However, current and former employees say the company has only seriously addressed the problem when the money is at stake or – in the case of terrorist content – when the external pressure forced him to act. The unwanted video clbadifier was partly motivated by financial incentives. As more and more people use smartphones to connect, more and more people are accessing the YouTube app and website directly, instead of watching YouTube clips in other online destinations. .

This was potentially good news for YouTube. When people see ads directly on the service, rather than anywhere on the Internet, Google often generates more revenue, says someone familiar with the business. People have asked not to be identified because they are private information. YouTube has also doubled its efforts by adding new features to the home page, including a "Trends" tab and custom clip carousels.

YouTube leaves Hollywood competition with Netflix and Amazon

The problem, however, was that some popular videos appearing on the home page were scary: clips showing people fighting or a person falling from a heavy fall, for example. A former engineer is reminded of an internal meeting involving a still image of a video with male bads. This type of content resulted in clicks, but also hijacked some people from the homepage and became a possible advertising nightmare. YouTube's terms of use prohibit badgraphy.

That's what led YouTube to adopt the junk food clbadifier. The system was a success, helping to remove from the homepage of many unpleasant clips and to bring back the spectators. Recently, Google told sellers that the "time spent in front of the TV" on the home page and the YouTube app had been multiplied by ten over the past three years.

The value of this artificial intelligence system has become even clearer this month, when Google, from Alphabet, has launched advertising features for YouTube's growing live audience. For the first time, marketers can make promotions on smartphones when users move in the YouTube app. The video service looks a lot like Facebook Inc.'s Instagram and could help Google get more advertising money from its competitors on social media.

YouTube turns off comments on videos with minors for critical

The company does not share YouTube's financial data, but RBC Capital Markets estimate that the video service generated sales of more than 20,000 million US dollars last year. The improved audience on the home page and the new methods of posting ads could generate billions of dollars in additional revenue. However, YouTube's AI software can only reach a certain point. Machines have trouble automatically badyzing what is happening and what is being said in the video clips, especially without the help of textual descriptions of the clips. More than 450 hours of video are uploaded to YouTube every minute. The company quotes this gigantic sum to justify repeated mistakes.

The latest artificial intelligence techniques improve the recognition of the content and context of videos, although the data signals remain "noisy", said Serena Yeung, researcher at Stanford University, advised by Google administrator, AI Fei-Fei Li. Could the software be trained to identify something like "people fighting" just for images? "You could," Yeung said. "It will probably have a lot of mistakes, it will work some of the time, but not all the time."