How to Offend Facebook’s AI Moderators

Physics of Shadow and Light
Physics of Shadow and Light
5 min readJan 10, 2021


Facebook didn’t ban a Trump ally who called for a beheading but it banned Christians for love.

Facebook is relying more heavily on Artificial Intelligence (AI) for content moderation. During the 2020–2021 U.S. election season and pandemic this effort has failed. Facebook should invest more in helping humans safely moderate during a pandemic. As it is right now, AI moderation is wildly inaccurate and anyone can be put into “Facebook jail” for mysterious reasons including:

1.) Reference real history A Facebook post quoting General Philip Sheridan on the 152nd anniversary of the Washita Creek…incident put the author into Facebook exile for most of December. General Custer and General Sheridan did not meet community standards of 1868 or 2020. Facebook’s AI moderators probably maintain a list of racist historical quotations to ignore. Custer, Nixon, George Wallace and LBJ are well represented but Sheridan’s most racist quote is almost always misquoted as an even more racist community standards version!

2.) Contradict a friend who shares false information they got from some famous propagandist or random fake account. If a friend posts false information about Covid-19 or who won the 2020 U.S. presidential election, do not post a rebuttal unless you’re okay with losing their friendship and/or your Facebook account.

Photo of a March 15th Facebook post countering common U.S. falacies about Covid-19
Many Americans were in denial in the early days of Covid-19 and remain in denial after 11 months and 360,000+ deaths. They happily spread that denial over social media. Facebook moderators were also in denial.

3.) Contradict a disinformation profiteer Most ordinary humans don’t have thousands of Facebook friends but politicians and professional propagandists tell people what they want to hear and quickly attract hundreds of thousands or millions of followers. They gain their popularity by dividing us and pitting members of their cult against the rest of us. Some of these cultists would do anything to preserve their illusion including destroy, dox and hurt you. It’s easy for a few to gang up and weaponize Facebook’s moderation AI to get you banned and drive you to Parler, 8chan or another cult social media platform.

Photo of popular photoshopped social media meme which attempts to link fire and violence to BLM and democrats.
A version of this image spread far and wide promoting a racist narrative about BLM protestors.

4.) Post evidence of racism in the past or present. Social media companies reference community standards from communities which are deeply racist but want to believe that they are not. White fragility is real and it’s so easy to end up on the wrong side of it, progressive black Facebook pastors have had to use whyte instead of white to talk about certain members of a deeply insecure race that believe themselves to be self-made and self-assured but are not.

Photo of police with gun pointed towards a little girl on a BLM protestor’s shoulders which triggered fb AI moderator
Thousands of Facebook users saw and shared photoshopped images trying to push the narrative that Black Lives Matter protestors were more dangerous than organizers of a plot to overthrow the U.S. government and that the police treated BLM protestors better. A single contradictory image triggered Facebook’s moderation and disabled me from posting for 30 days.

5.) Speak Truth to Power One of the most powerful allies of an authoritarian U.S. president posted a call for the beheading of an infectious disease expert to his 175,000 followers. Mark Zuckerberg implicitly blessed this post by allowing him to stay on Facebook. The moderator of the Christian for Love Facebook group has been banned twice for mysterious reasons. Why? Because Facebook makes far more money off Stephen Bannon’s 175,000 followers than off Christians4Love’s 18 friends.

Hat tip to the hard-working human moderators who put me back in touch with family and friends who have been bullied, slandered and lied to by this guy and his accomplices.

6.) Explain the First Amendment to someone who is able to communicate with tens of thousands of groupies through TV, Radio and social media platforms. Tell them that the first amendment does not require that anyone give them a megaphone and a soapbox. It doesn’t require anyone to give them a platform or infrastructure to allow their stream of consciousness to become the conscience for hundreds of thousands of Americans. It doesn’t allow them to scream fire in a crowded theater. It wasn’t designed to help scam artists steal money from vulnerable Americans. It wasn’t designed so the powerful could lie to the weak but so the people could speak truth to the powerful.

FCC regulations do need to change. The last time Americans confused free speech with mob-incitement anarchy, a fake doctor killed hundreds of Americans with his goat balls quackery and sweet-talked followers into lock-step with Hitler’s fascism. FCC reform should treat social media as an loosely regulated communications conduit until the number of followers exceeds the typical audience of micropower AM/FM broadcast station. Above this threshold, Social Media, TV and Radio should demonstrate that they are not abusing our public resources (the internet and the airwaves) and profiteering from dividing us and destabilizing our nation’s moral and social structure.

7.) Post original content. It shouldn’t be a surprise that Facebook didn’t become a $250 billion+ company by sharing selfies of you and your pet goldfish. They profit from viral content and the superspreaders who spread it. They are able to devote more moderation resources to this profitable content. So when a foreign power with thousands of bots and complicit helpers or a disinformation profiteer with millions followers pushes a photoshopped meme slandering 20 million BLM protestors and more than 80 million democrats, Facebook can afford to devote human moderation resources to vetting that meme. But if you use an unusual hiberno-english colloquialism which is a Matryoshk word that happens to contain an offensive word within it, or if you share a photo of your own cancer-struck child, the AI moderators might flag it and the hard-working underappreciated human moderators may never see it. From Facebook’s November 2020 transparency report:

“Due to a temporary reduction in our review capacity as a result of COVID-19, we could not always offer our users the option to appeal. We still gave people the option to tell us they disagreed with our decision”

Facebook’s message:“We could not process your request. Please try again later” to review an AI moderation
If Facebook allows logins at all you may be able to see the cause of the trouble in the support inbox and a link which used to provide an opportunity to disagree with the decision and submit it for human review. But Facebook has not hired enough human reviewers to sort toxic content from bugs in their own AI moderation.

Approximately 4 million robo-blocked accounts have had no option since the first quarter of 2020, almost a year ago.

Graph from Facebook’s November 2020 transparency report showing a huge dropoff in moderator content restoration.
Facebook’s transparency report shows that actioned content restoration dropped from 2 million+ to near zero after Q1 2020. It appears that appeals also dropped to near zero because Facebook removed the appeal mechanism for most flagged content.



Physics of Shadow and Light
Physics of Shadow and Light

Physics of Shadow and light is about science, history, good and evil, truth, beauty, love and infinity.