번역에서 제공
flickr. Rego Korosi. CC BY.
상황을 하나 상상해보자. 당신이 궁금한 정보가 있어 유튜브에 키워드를 검색했다. 검색 상단에 오른 동영상 중 관심이 가는 썸네일을 클릭한다. 영상을 시청한다. 자연스럽게 ‘다음 동영상’ 목록에 오른 것을 시청한다. 그 후 또 한번 ‘다음 동영상’을 추천받아 시청한다. 이 과정을 반복한다.
위 상황은 우리가 유튜브에서 콘텐츠 추천 시스템을 접하게 되는 지점을 보여준다. 검색 결과 상단에 오르는 것, 그리고 ‘다음 동영상(Up next)’ 목록에 오르는 것. 이 두 가지는 모두 유튜브가 구성한 추천시스템에 의해 작동되며 이용자가 서비스를 계속 이용하도록 유도한다.
동영상 산업에서 콘텐츠 추천 시스템은 서비스의 성패를 좌우하는 핵심이 된다. 특히 유튜브는 계속해서 개인화 혹은 맞춤화를 강조해왔다. 이용자의 선호를 잘 파악하고, 양질의 콘텐츠를 제시해 이용자 편의를 추구한다는 것이다. 그렇다면 유튜브는 어떤 항목을 기준으로 콘텐츠 추천 시스템을 작동시킬까. 누구나 한번쯤 궁금해봤을 법한 문제다. 하지만 ‘기술’, ‘인공지능’, ‘알고리즘’ 같은 용어들이 더 이상의 궁금증을 막곤 했다.oo
의문은 계속해서 제기돼 왔다. 윤리적으로 도저히 납득하기 어려운 사건들이 연이어 터졌기 때문이다. 자살한 사람이 촬영된 ‘로건 폴’ 채널의 영상은 수백만건의 조회수를 기록했고, 디즈니 애니메이션 주인공을 소재로 한 자극적인 콘텐츠가 문제가 될 뿐만 아니라 폭력·테러에 관한 영상들도 항상 문제가 됐다. 그때마다 유튜브는 ‘구글의 머신러닝과 자동 알고리즘 시스템을 철저히 강화하겠다’라는 대응만 밝혀 비난을 샀다.
(사진=The Guardian)
“유튜브 추천 알고리즘은 사람들이 온라인에서 더 많은 시간을 보내도록 왜곡돼 있다.”
전 유튜브 추천시스템 담당자가 <가디언>에 추천 알고리즘 방식에 대한 의혹을 폭로했다. 기욤 샤스로 전 유튜브 엔지니어팀 직원은 구글에서 3년간 근무했으며 2013년에 해고당했다. 그는 유튜브의 추천 시스템이 결코 민주주의적이고, 진실에 가깝고, 균형적인 것을 최적화한 형태로 작동하지 않는다고 주장했다. 가장 우선순위는 시청 시간이다. 그는 자신이 일했던 엔지니어팀에서 사람들이 유튜브 내에서 동영상 시청 시간을 연장해 광고 수입을 늘리도록 하는 시스템을 계속해서 실험했다고 밝혔다.
앞서 설명했듯 유튜브는 이용자들에게 알고리즘에 관련한 어떠한 데이터도 공개하지 않는다. 이용자는 특정 동영상이 어떤 알고리즘을 통해 홍보됐는지, 유튜브의 추천 시스템 작동에 의해 홍보된 동영상이 어떤 성과를 얻었는지 전혀 알 수 없다. 때문에 유튜브가 알고리즘 내에서 편향된 추천 시스템을 작동시켰다거나 왜곡된 알고리즘 패턴을 적용했는지에 대한 추측도 할 수 없는 상황이다.
기욤 샤스로는 단순한 폭로에 그치지 않았다. 그는 지난 2016년을 시작으로 유튜브의 추천 알고리즘의 작동을 추측할 수 있는 데이터를 수집하기 시작했다. 웹사이트를 개설하고 스스로 설계한 프로그래밍을 통해 유튜브의 데이터를 수집했다. 이후 그가 수집한 데이터와 <가디언>이 손을 잡았다. <가디언>은 지난 2월2일 기욤 샤스로의 데이터베이스로부터 분석한 결과를 그의 인터뷰와 함께 보도했다.
기욤 샤스로의 유튜브 추천 데이터 웹사이트 화면 갈무리. 주제별로 가장 많은 추천을 받은 영상과 키워드 데이터를 제시한다.
#기욤 샤스로의 데이터 수집
기욤 샤스로는 유튜브의 완벽한 데이터 샘플을 긁어올 순 없음을 이미 알고 있었다. 때문에 2가지 추천 데이터를 단순히 스냅하는 방법을 선택했다. 유튜브의 초기 검색 결과에서 제공되는 리스트와 다음 동영상으로 추천하는 영상 리스트다. 수천번의 작업을 통해 기욤 샤스로는 유튜브가 특정 주제에 대해 추천 컨베이어 벨트에 올리는 데이터군을 수집할 수 있다는 것을 경험했다.
작업 과정은 이렇다. 그는 지난 대통령 선거 당시 도널드 트럼프 후보와 힐러리 클린턴 후보에 관한 유튜브 추천 시스템 작동을 주제로 잡았다. 먼저 각 후보 이름을 검색해서 나오는 결과 중 상위 5개 비디오를 수집하고, 해당 동영상을 클릭한 후 다음 동영상으로 추천하는 비디오 리스트를 캡처했다. 기욤 샤스로는 유튜브가 추천한 다음 동영상이 어떤 성격의 비디오인지를 분석하는 작업까지 프로세스를 반복했다.
이를 통해 기욤 샤스로가 공개한 대선후보 관련 데이터 리스트는 총 8052개다. 몇 가지 규칙이 있었다. 기욤 샤스로는 균형 있는 데이터를 수집하기 위해 개인화 시스템 작동을 최대한 방지해야 했다. 그는 자신의 검색 결과를 균형 있게 하기 위해 각 후보의 이름을 번갈아가며 한 번씩 검색했다. 또한 동영상 시청기록에 영향을 받지 않기 위해 추천된 다음 동영상으로 넘어가는 과정을 거치지 않았다.
#<가디언>의 분석
유튜브 검색 결과 및 추천 동영상
동영상을 시청하면 오른쪽에 ‘다음 동영상(Up next)’ 리스트를 제안한다.
<가디언>은 기욤 샤스로가 공개한 8052개 모든 동영상과 그중 가장 많이 추천된 상위 1천개의 동영상을 집중 분석했다. 각 개별 비디오에 대해 동영상의 성격도 분석했다. 추천된 동영상의 성격이 각 당을 지지하는지 여부를 판별하기 위해 내용 및 제목을 조사했다. 그 결과 전체에서 3분의 2에 해당하는 동영상 콘텐츠가 특정 후보에게 유리하게 작동하는 내용이었음을 판단했다. 총 643개의 편향 콘텐츠 중 551개가 트럼프를 지지하는 성격을 나타냈다. 나머지 3분의 1에 해당하는 추천 동영상은 선거와 무관하거나 정치적으로 중립적인 것으로 분석했다.
<가디언>은 전체 영상을 ‘추천 횟수’에 따라 순위를 매기는 작업도 했다. 검색 동영상 옆에 다음 동영상으로 추천된 횟수를 1회당 1추천으로 계산했다. 이에 따라 상위 추천 동영상 25개와 추천수가 많은 상위 유튜브 채널 10개를 추렸다. 대부분 중립적이라기보다는 특정 후보에 대한 성향을 드러내는 것들이다. 그중 몇 개 리스트는 아래와 같다. 제목을 통해 성향을 유추할 수 있다. (링크가 걸려있지 않은 것은 해당 보도 전후로 계정이 삭제됨)
– Trump supporter leaves CNN anchor speechless
– Must Watch!! Hillary Clinton tried to ban this video
– Can Donald Trump win the presidential election?
– Busted! Bill Clinton’s Face When Trump Brings Up The Rape Allegations is Priceless.
– BREAKING!!! JULIAN ASSANGE “DEAD MAN SWITCH” Goes Off after EXPOSING Hillary Clinton?
<가디언>은 정치 관련 데이터베이스 분석업체 그래피카와도 분석한 내용을 공유했다. 그래피카는 유튜브의 추천 동영상 데이터를 2016년 선거에서 활동한 트위터 계정 데이터와 통합하는 작업을 했다. 그 결과 51만3천개가 넘는 트위터 계정에서 기욤 샤스로가 공개한 유튜브 추천 동영상 중 적어도 1개 이상의 링크를 트윗했다는 것을 발견했다. 특히 그중 3만6천개 계정은 1개 이상의 링크를 10번 이상 트윗했다. 졸 켈리 그래피카 전무는 “선거 이전 수개월간 정치 요원들이 관리하는 수천개의 트위터 계정에서 유튜브 동영상이 업데이트됐다”라며 “가장 많은 연결이 있었던 건 도널드 대통령 후보를 지지하는 계정이었다”라고 밝혔다. <가디언>은 “해당 트윗 계정 중 가장 활동적으로 나타난 19개 계정은
http://www.bloter.net/archives/301890
The methodology Guillaume Chaslot used to detect videos YouTube was recommending during the election – and how the Guardian analysed the data
YouTube’s recommendation system draws on techniques in machine learning
to decide which videos are auto-played or appear “up next”. The precise
formula it uses, however, is kept secret. Aggregate data revealing
which YouTube videos are heavily promoted by the algorithm, or how many
views individual videos receive from “up next” suggestions, is also
withheld from the public.
Disclosing that data would enable academic institutions, fact-checkers and regulators (as well as journalists) to assess the type of content YouTube is most likely to promote. By keeping the algorithm and its results under wraps, YouTube ensures that any patterns that indicate unintended biases or distortions associated with its algorithm are concealed from public view.
By putting a wall around its data, YouTube, which is owned by 구글 protects itself from scrutiny. The computer program written by Guillaume Chaslot overcomes that obstacle to force some degree of transparency.
The ex-구글 engineer said his method of extracting data from the
video-sharing site could not provide a comprehensive or perfectly
representative sample of videos that were being recommended. But it can
give a snapshot. He has used his software to detect YouTube
recommendations across a range of topics and publishes the results on
his website, algotransparency.org.
The program simulates the behaviour of a YouTube user. During the election, it acted as a YouTube user might have if she were interested in either of the two main presidential candidates. It discovered a video through a YouTube search, and then followed a chain of YouTube–recommended titles appearing “up next”.
Chaslot programmed his software to obtain the initial videos through YouTube searches for either “Trump” or “Clinton”, alternating between the two to ensure they were each searched 50% of the time. It then clicked on several search results (usually the top five videos) and captured which videos YouTube was recommending “up next”.
The process was then repeated, this time by selecting a sample of those videos YouTube had just placed “up next”, and identifying which videos the algorithm was, in turn, showcasing beside those. The process was repeated thousands of times, collating more and more layers of data about the videos YouTube was promoting in its conveyor belt of recommended videos.
By design, the program operated without a viewing history, ensuring it was capturing generic YouTube recommendations rather than those personalised to individual users.
The data was probably influenced by the topics that happened to be trending on YouTube on the dates he chose to run the program: 22 August; 18 and 26 October; 29-31 October; and 1-7 November.
On most of those dates, the software was programmed to begin with five videos obtained through search, capture the first five recommended videos, and repeat the process five times. But on a handful of dates, Chaslot tweaked his program, starting off with three or four search videos, capturing three or four layers of recommended videos, and repeating the process up to six times in a row.
Whichever combinations of searches, recommendations and repeats Chaslot used, the program was doing the same thing: detecting videos that YouTube was placing “up next” as enticing thumbnails on the right-hand side of the video player.
His program also detected variations in the degree to which YouTube appeared to be pushing content. Some videos, for example, appeared “up next” beside just a handful of other videos. Others appeared “up next” beside hundreds of different videos across multiple dates.
In total, Chaslot’s database recorded 8,052 videos recommended by YouTube. He has made the code behind his program publicly available here. The Guardian has published the full list of videos in Chaslot’s database here.
The Guardian’s research included a broad study of all 8,052 videos as well as a more focused content analysis, which assessed 1,000 of the top recommended videos in the database. The subset was identified by ranking the videos, first by the number of dates they were recommended, and then by the number of times they were detected appearing “up next” beside another video.
We assessed the top 500 videos that were recommended after a search for the term “Trump” and the top 500 videos recommended after a “Clinton” search. Each individual video was scrutinised to determine whether it was obviously partisan and, if so, whether the video favoured the Republican or Democratic presidential campaign. In order to judge this, we watched the content of the videos and considered their titles.
About a third of the videos were deemed to be either unrelated to the election, politically neutral or insufficiently biased to warrant being categorised as favouring either campaign. (An example of a video that was unrelated to the election was one entitled “10 Intimate Scenes Actors Were Embarrassed to Film”; an example of a video deemed politically neutral or even-handed was this NBC News broadcast of the second presidential debate.)
Many mainstream news clips, including ones from MSNBC, Fox and CNN, were judged to fall into the “even-handed” category, as were many mainstream comedy clips created by the likes of Saturday Night Live, John Oliver and Stephen Colbert.
Formulating a view on these videos was a subjective process but for the most part it was very obvious which candidate videos benefited. There were a few exceptions. For example, some might consider this CNN clip, in which a Trump supporter forcefully defended his lewd remarks and strongly criticised Hillary Clinton and her husband, to be beneficial to the Republican. Others might point to the CNN anchor’s exasperated response, and argue the video was actually more helpful to Clinton. In the end, this video was too difficult for us categorise. It is an example of a video listed as not benefiting either candidate.
For two-thirds of the videos, however, the process of judging who the content benefited was relatively uncomplicated. Many videos clearly leaned toward one candidate or the other. For example, a video of a speech in which Michelle Obama was highly critical of Trump’s treatment of women was deemed to have leaned in favour of Clinton. A video falsely claiming Clinton suffered a mental breakdown was categorised as benefiting the Trump campaign.
We found that most of the videos labeled as benefiting the Trump campaign might be more accurately described as highly critical of Clinton. Many are what might be described as anti-Clinton conspiracy videos or “fake news”. The database appeared highly skewed toward content critical of the Democratic nominee. But for the purpose of categorisation, these types of videos, such as a video entitled “WHOA! HILLARY THINKS CAMERA’S OFF… SENDS SHOCK MESSAGE TO TRUMP”, were listed as favouring the Trump campaign.
Missing videos and biasRoughly half of the YouTube-recommended videos in the database have been taken offline or made private since the election, either because they were removed by whoever uploaded them or because they were taken down by YouTube. That might be because of a copyright violation, or because the video contained some other breach of the company’s policies.
We were unable to watch original copies of missing videos. They were therefore excluded from our first round of content analysis, which included only videos we could watch, and concluded that 84% of partisan videos were beneficial to Trump, while only 16% were beneficial to Clinton.
Interestingly, the bias was marginally larger when YouTube recommendations were detected following an initial search for “Clinton” videos. Those resulted in 88% of partisan “Up next” videos being beneficial to Trump. When Chaslot’s program detected recommended videos after a “Trump” search, in contrast, 81% of partisan videos were 페이버러블 to Trump.
That said, the “Up next” videos following from “Clinton” and “Trump” videos often turned out to be the same or very similar titles. The type of content recommended was, in both cases, overwhelmingly beneficial to Trump, with a surprising amount of conspiratorial content and fake news damaging to Clinton.
Supplementary countAfter counting only those videos we could watch, we conducted a second analysis to include those missing videos whose titles strongly indicated the content would have been beneficial to one of the campaigns. It was also often possible to find duplicates of these videos.
Two highly recommended videos in the database with one-sided titles were, for example, entitled “This Video Will Get Donald Trump Elected” and “Must Watch!! Hillary Clinton tried to ban this video”. Both of these were categorised, in the second round, as beneficial to the Trump campaign.
When all 1,000 videos were tallied – including the missing videos
with very slanted titles – we counted 643 videos had an obvious bias. Of
those, 551 videos (86%) favoured the Republican nominee, while only 92
videos (14%) were beneficial to Clinton.
Whether missing videos were included in our tally or not, the conclusion was the same. Partisan videos recommended by YouTube in the database were about six times more likely to favour Trump’s presidential campaign than Clinton’s.
Database analysisAll 8,052 videos were ranked by the number of “recommendations” – that is, the number of times they were detected appearing as “Up next” thumbnails beside other videos. For example, if a video was detected appearing “Up next” beside four other videos, that would be counted as four “recommendations”. If a video appeared “Up next” beside the same video on, say, three separate dates, that would be counted as three “recommendations”. (Multiple recommendations between the same videos on the same day were not counted.)
Here are the 25 most recommended videos, according to the above metric.
Trump supporter leaves CNN anchor speechlessThis Video Will Get Donald Trump Elected
Must Watch!! Hillary Clinton tried to ban this video
SR# 1271 NBC Crew – Crooked Hillary’s MASSIVE MELTDOWN at Commander-in-Chief Forum
10 Photos of MELANIA, TRUMP Wishes We’d Forget
Full Interview: Donald Trump, Melania & Family with George Stephanopoulos
Busted! Bill Clinton’s Face When Trump Brings Up The Rape Allegations is Priceless
Donald Trump Has Won The 2016 Presidential Election
Angry Ivanka Trump Walks Out Of Cosmo Interview
TRUMP: the COMING LANDSLIDE ~Ancient Prophecy Documentary of Donald Trump / 2016
ANONYMOUS WATCH - HILLARY CLINTON, YOU ARE FINISHED!
“Obama out:” President Barack Obama’s hilarious final White House correspondents’ dinner speech
Watch Live: The Final Presidential Debate
Can Donald Trump win the presidential election?
Michelle Obama’s EPIC Speech On Trump’s Sexual Behavior (FULL | HD)
ALL LEAKED TRUMP FOOTAGE Lewd comments Made on Daughter Ivanka Mini Documentary
Melania Trump - The Woman Behind Donald
BREAKING: VIDEO SHOWING BILL CLINTON RAPING 13 YR-OLD WILL PLUNGE RACE INTO CHAOS ANONYMOUS CLAIMS
BREAKING!!! JULIAN ASSANGE “DEAD MAN SWITCH” Goes Off after EXPOSING Hillary Clinton?
Bill Clinton’s Sexual Escapades
Anonymous Release Bone-Chilling video of Huma Abedin every American Needs to See
BREAKING: Michael Moore Admits Trump Is Right
BREAKING: FBI Reopens Hillary Clinton Email Investigation
Full monologue: Donald Trump roasts Hillary Clinton at Al Smith charity dinner
Hillary Cheats AGAIN?? Debate #3 Earphone AND Teleprompter?? BUSTED ON TV!
Chaslot’s database also contained information the YouTube channels used to broadcast videos. (This data was only partial, because it was not possible to identify channels behind missing videos.) Here are the top 10 channels, ranked in order of the number of “recommendations” Chaslot’s program detected.
- The Alex Jones Channel
- Fox News
- DONALD TRUMP SPEECHES & PRESS CONFERENCE
- The Young Turks
- MSNBC
- CBS News
- TheRichest
- The Next News Network
- CNN
- Right Side Broadcasting Network
We searched the entire database to identify videos of full campaign speeches by Trump and Clinton, their spouses and other political figures. This was done through searches for the terms “speech” and “rally” in video titles followed by a check, where possible, of the content. Here is a list of the videos of campaign speeches found in the database.
- Donald Trump (382 videos)
- Barack Obama (42 videos)
- Mike Pence (18 videos)
- Hillary Clinton (18 videos)
- Melania Trump (12 videos)
- Michelle Obama (10 videos)
- Joe Biden (42 videos)
The Guardian shared the entire database with Graphika, a commercial analytics firm that has tracked political disinformation campaigns. The company merged the database of YouTube-recommended videos with its own dataset of Twitter networks that were active during the 2016 election.
The company discovered more than 513,000 Twitter accounts had tweeted links to at least one of the YouTube-recommended videos in the six months leading up to the election. More than 36,000 accounts tweeted at least one of the videos 10 or more times. The most active 19 of these Twitter accounts cited videos more than 1,000 times – evidence of automated activity.
“Over the months leading up to the election,
나는 유튜브에서 미국 정부가 수작업으로 추천한 것이 특정인에게 추천으로 뜰 수도 있다고 생각한다. 그렇게 만드는 건 기술적으로 무척 쉽고 쓸모가 크니까.
레드튜브
사실 추천떠도 안보면 되는데 ㅋㅋ
다상관없으니까 그좆같은 티비플좀 추천에 안뜨게해라 뜰때마다 관심없음 처눌르고 3년째 안보는데 계속뜨네 이걸보면 유투브 알고리즘은 개좆쓰레기다