A day after Facebook finally announced that it would send home all its contract workers who moderate content and will rely on automated content removals, people around the world, including in India, reported that Facebook was marking legitimate news articles, including those about COVID-19, as spam. One such article that Facebook marked as spam was Huffington Post India’s report on the National Social Registry.

Facebook clarified that this is a case of correlation, not causation. Guy Rosen, Vice President (Integrity) at Facebook, tweeted that this was “a bug in an anti-spam system, unrelated to any changes in our content moderator workforce”. He again tweeted earlier this morning that the issue had been resolved and the company has since restored all posts that were “incorrectly removed” on all topics.

Megha Bahree, an Indian journalist, whose Facebook post about National Social Registry was marked as spam on March 17 told MediaNama that her post has now been restored. We were also able to share the article on Facebook without any problem today.

Screenshot of the Facebook prompt Megha Bahree got.

How does Facebook’s content moderation work? On a normal, non-COVID-19 day, automated systems take down content that it definitively violative of community standards, such as child pornography. Content in grey area, such as hate speech, is sent to human content moderators. The humans then either take the content down, or let it be. The machine learning algorithms recognise patterns from these takedowns and restorations to improve themselves. Since this happens at a very large scale, things fall through the cracks, which is why users are allowed to appeal decisions and report content.

What is happening now? Now that human oversight has been removed, the algorithms are presumably relying only on what they have learnt in the last three years, that is, since Facebook’s current content moderation systems were put in place.

Why is this a problem? While the need to restrict information about COVID-19 to only governments, international organisations and verified publications is an exceptional, but understandable step to take, the case about HuffPost report on National Social Registry is far more disturbing. This is a report from a reputed digital news publication that is critical of the incumbent Indian government. The roughly 3,500-word piece had only 11 links within the story itself, not counting links on the HuffPost website that are a standard across all its pages.

The marking of this article as spam suggests that articles about Aadhaar, surveillance in India, etc. had been taken down by human content moderators in the past, perhaps even from verified news publications which is why the HuffPost article got flagged. Taking Facebook’s statement that this was a bug in the anti-spam filter at face value, why did this happen with an article critical of the government that deals with surveillance? That is not to suggest that the Indian government has ordered it, but that there are human biases at play throughout this chain of moderation that have unintended consequences.

Over-reliance on algorithms, that are trained on biased data sets, governed by human biases (this is unavoidable), is something that needs to be reconsidered. Biases, both in the algorithms and of humans, can at best be mitigated. Then should algorithms be allowed to censor? The Intermediary Guidelines (Amendment) Rules 2018, which are currently under deliberation, insist that intermediaries must use automated mechanisms to “proactively” identify and remove unlawful content. There is no provision for review of such takedowns, or for content restoration in the Rules.

Twitter and YouTube are also relying on automated content removal

Facebook is not the only one that to rely on automated content takedowns. Twitter is also increasing its “use of machine learning and automation”, but clarified that it would not “permanently suspend any accounts based solely on our automated enforcement systems”. YouTube announced that it would “temporarily” rely on automated content takedowns which will happen without human review, but strikes usually won’t be issued in these cases (YouTube follows a three-strike system).

All three platforms warned that “mistakes” might happen as it is the first time fully automated content takedown systems are being used. YouTube clarified that decision appeals might take longer because of the circumstances, while Facebook said that fewer people would be available which is why the company would prioritise “imminent harm”.