FakeTube: AI-Generated News on YouTube


Seit einem Jahr ungefähr fällt mir auf, wann immer ich nach News-Clips auf Youtube suche, etwa um ein Posting irgendeiner WTF-Meldung mit ’nem Video anzureichern oder wenn ich Rips von BBC-Clips suche oder ähnliches: Die Suchergebnisse sind geflooded mit Diashow-ähnlichen Videos, in denen Standbilder und Screenshots gezeigt werden und in denen eine automatisch generierte Robot-Stimme den Text einer Meldung vorliest.

Jonathan Albright (der Ende letzten Jahres bereits die hervorragende dreiteilige Analyse der Election2016 Micro-Propaganda Machine um Outrage-Memetics und „Fake-News“ veröffentlichte), hat sich die Sache nun genauer angesehen und ein Netzwerk aus 19 Youtube-Channels identifiziert, die bislang rund 80k Videos veröffentlicht haben, die alle demselben Video-Title-Schema folgen („A Tease: [Title]“): 📺 FakeTube: AI-Generated News on YouTube.

Ich kann mich nicht daran erinnern, dass meine Suchergebnisse mit „A Tease“ begannen und gehe davon aus, dass es noch weitere dieser Netzwerke gibt. Interesting Times, indeed.


Through a network analysis and channel detail API request, I found the 19 YouTube channels account for 78,349 similarly named, keyword-packed “A Tease…” videos. As you can see in the list above, this is what amounts to a large-scale “fake news” video content farm. YouTube is a big place, but ~80,000 videos is a large number by any measure. Especially given that it’s only 19 channels.

Each “T” AI-generated video consists of a progressive “slideshow” of still images related to the title of the video, which originate from various websites, WordPress blogs, and content delivery networks across the internet. […] A computerized voice “reads” out text from published news articles and other news-related content as the slideshow progresses. The narration is surprisingly accurate and coherent, except for the brief pauses during the transitions between the sources from which the video derives its “story.”

Some of the “T” channels — like the example above — list the source URLs for the images and spoken text underneath the video’s “…More” section. At the same time, many videos do not seem to contain a source list. I’m not sure this kind of data is readily available through YouTube’s API, but I can say it would be difficult to go to nearly 80,000 videos and manually check for lists of sources.

Everything about these “A Tease…” videos suggests SEO, social politcs amplification, and YouTube AI-playlist placement: In addition to the video titles being keyword-packed and URLs pasted all over the video descriptions, the spoken text that’s already published on news sites should help to boost the overall relevance of these videos and any associated news-related “channels” on YouTube.

Given the large number of these AI-generated YouTube videos, I thought it might be interesting to see the time between uploads. It appears some of these “A Tease…” channels (see “A T User” example below) post new “news” videos every three to four minutes.