YouTube keeps deleting evidence of Syrian chemical weapon attacks

As Google-owned YouTube continues to delete millions of videos, the Syrian Archive is racing to back-up crucial evidence
A child receives treatment after a suspected chlorine attack on Saqba, in eastern Ghouta on April 7, 2018, in which 29n people were injuredKhaled Akasha/Anadolu Agency/Getty Images

At dusk on April 7, one of the worst chemical attacks in Syrian history hit the city of Douma in eastern Ghouta, killing more than 70 people. Minutes later, video reports began to appear on YouTube.

As darkness fell on the Syrian town, population 111,864, video shot inside a dusty building captured multiple bodies foaming at the mouth. Among the dead, a mother and father cradling their newborn, a thick white crust covering their noses and lips. The attacks on the rebel held town sparked international outrage.

YouTube footage is often the only evidence proving human rights violations are taking place in Syria. Yet this video, uploaded by a legitimate Douma based news agency, Kumait, has been removed by Google after it was flagged by its machine learning algorithm – along with thousands like it.

Some videos could be lost forever, according to Hadi Al-Khatib, founder of open source initiative the Syrian Archive, which aims to verify, back-up and preserve footage of the conflict. Other removed videos include hospitals being bombed, victims of chemical attacks, and ammunition remnants. “They are not propaganda; they are not extremist content,” says Al-Khatib. “This information might be the only source that indicates an attack has happened on a specific area, at a certain time.”

It is the most documented war in history, and most of it has taken place on social media. YouTube is citizen journalists’ platform of choice but preserving videos before Google removes them has proven to be a challenge. Between 2012 and 2018, Google has taken down 123,229 of the 1,177,394 videos the Syrian Archive has been able to back-up, Al-Khatib says. And these are just the videos we know about.

Following the Douma attack in April this year, the Syrian Archive launched a publicly-accessible chemical weapons database to specifically preserve videos documenting this type of attack. Videos on the database date back to 2014. Al-Khatib says 72 of the 861 videos have disappeared from YouTube – all of these legitimate.

Google has been removing huge numbers of online videos since it announced it would use machine learning to detect extremist content in June 2017. Two months later, the firm boasted that 75 per cent of videos removed for “violent extremism” were taken down “before receiving a single human flag”. By December last year, this figure had reached 98 per cent.

The Syrian Archive can preserve removed YouTube videos if the team are able to get to them first and back them up to its database. But there are huge implications for content taken down by Google’s algorithm before it can be archived. “I think we have already lost a lot of content since YouTube started using machine learning in 2017,” Al-Khatib says. “There will be a big impact on the Syrian accountability process if we aren’t able to retrieve it."

The algorithms are flagging videos that do not adhere to YouTube’s community guidelines, which ban “gratuitous violence, dangerous and illegal activities, and hate speech”. However, the company says it does make exceptions for material with “sufficient educational and documentary news value”.

But because footage of the Syrian war is often graphic, human rights content can be mistakenly flagged by the algorithm as violating YouTube’s guidelines. Google accepts it “doesn’t always get it right” and, according to a YouTube spokesperson, the firm takes the issue “incredibly seriously”.

Yet Al-Khatib and his team know a lot of the removed content is legitimate. Drawing from over 5,000 sources, they are able to verify videos by looking at landmarks, using Google Earth and satellite imagery. “First, we need to find the source: are they credible, do they have an online reputation, have they published reliable information before? Then we verify the date, time and location.”

The Syrian Archive and its partners have also had some success persuading Google to reinstate videos they have proven to be legitimate: 400,000 were put back online between mid-2017 and early 2018. And removals have slowed down since March 2018, Al-Khatib says. But no one knows exactly how the algorithm works – nor for how long Google is keeping the videos after taking them down.

“The response you get from Google depends on who you are talking to and how much they can actually help you,” says Dia Kayyali, programme manager, tech and advocacy at Witness – an organisation working alongside the Syrian Archive. “We have not been in communication as much as we would like in recent months.”

Human rights campaigners say Google’s opaque approach is a major obstacle. “Algorithmic transparency in general is a huge issue, but even more for content take downs,” Kayyali says.“They don’t have any outside auditing, and that’s a basic necessity of transparency. We don’t know the details of the algorithm: we have pushed and pushed, but we still don’t know.”

Eliot Higgins, founder of the open source and social media investigations platform Bellingcat has experienced Google’s strategy first hand. At one point, his entire YouTube account was shut down. Higgins managed to get his channel reinstated, but it still isn’t clear exactly why it was shut down. “If videos are being deleted before war crimes investigators can preserve them, evidence could be lost forever.”

Chris Woods, director at Airwars.org, an organisation monitoring the Syrian conflict, has also experienced issues caused by Google’s algorithm. Despite ensuring Airwars’ YouTube videos tracking attacks on civilians including children did not contain graphic images, Google stamped the content with an 18 certificate and refused to remove it.

“What troubled us deeply is: we had all assumed YouTube was a repository of videos – a permanent archive,” says Woods. “The last year has taught us how impermanent and vulnerable this material is, even on YouTube itself.”

Keith Hiatt, vice president at Silicon Valley-based non-profit organisation Benetech, was told by Google a year ago that a human reviews every piece of content flagged by the algorithm. “But does that mean a person watches each video and does a full, case-by-case analysis? Or does it just mean a human authorises removal?”

People viewing the videos need to be trained to spot human rights-related content, says Al-Khatib. “And, if it is removed, it’s really critical Google keeps it, so it can be reinstated.”

The evidence based on social media content has already proven to be invaluable. In August last year, the International Criminal Court (ICC) issued its first public warrant of arrest based mostly on video evidence and social media posts about war crimes in Libya.

“The scary thing to me is: we have no idea what is already missing,” says Kayyali of Witness. “The people doing open source investigations – such as the ICC and German government – will never know if those videos aren’t there, and it might stop them from being able to issue an arrest warrant or find a witness.”

But lawmakers, particularly in Europe, have made increasingly strong demands for platforms to identify and remove terrorist material within hours from initial upload. In response, social media firms have agreed to a series of voluntary commitments. This includes the terrorist hash database, which sees platforms using automated filters to find and remove duplicate material.

However, a video for example showing Isis recruitment can violate the law in one context, but also be legal and important for purposes such as documenting crimes for future prosecution, says Daphne Keller, intermediary liability director at Stanford's Centre for Internet and Society.

“The more we push companies to carry out fast, sloppy content removals, the more mistakes we will see,” Keller says. She thinks lawmakers should “slow down, talk to experts including both security researchers and members of the affected communities, and build on that foundation”.

YouTube says it’s committed to making its removal process “more transparent”, pointing to its Community Guidelines Enforcement Report and Reporting History Dashboard. The company also says it’s open to working with other organisations to “understand the context of videos to ensure important content remains available on the platform”.

Regulation will make things harder, but for now, humanitarian organisations such as the Syrian Archive can only continue to protest to Google when legitimate content is removed.

In the meantime, Al-Khatib’s team of eight are aiming to set up more separate databases focusing on airstrikes against civilian infrastructure including hospitals and medical facilities, and documenting ammunition. “We don’t have documents, crime scenes or physical evidence, so we want to add value to this content when it’s used for legal purposes,” Al-Khatib says. “YouTube is not the perfect place for it, but it is what people are using.”

Updated 26.06.18, 09:50 BST: YouTube states its algorithm is not automatically deleting videos, but rather flagging them for removal.

This article was originally published by WIRED UK