AI Used in Fact-Checking Process on Facebook Posts About Masks and Wildfires

Meta used artificial intelligence to apply existing fact-checks to new articles and videos shared on Facebook in 2020 and 2021, according to a guest essay published by Hot Air. The essay, adapted from Matthew Hoy’s book “Fact-Checking Frauds: How Fact-Checkers Distract, Deceive, and Distort Our Politics,” focuses on two cases.

In April 2021, City Journal published John Tierney’s article “Much to Forgive,” which examined studies from European countries on the effects of mask mandates for children and teens during the Covid-19 pandemic. The piece concluded that the mandates provided little benefit while causing harm to children’s health and education.

Shortly after publication, Facebook labeled shares of Tierney’s article as containing “Partly False Information. ” The label linked to a Science Feedback fact-check of a different article from GreenMedInfo that discussed a German preprint study on masks in children.

Science Feedback is described in the essay as an organization staffed by activists with science degrees that participated in Meta’s since-discontinued program with International Fact Checking Network signatories. The essay states that Science Feedback did not appear to review Tierney’s specific article or the studies it cited.

It suggests Meta’s SimSearchNet++ AI, introduced months earlier to detect modified images and text, likely made the connection between the pieces based on shared references and sources. ” The video attributed wildfires in western states primarily to poor forest management, while also referencing climate change.

Facebook applied a “false information” label linking to a Science Feedback fact-check that addressed the claim “Forest Fires are caused by poor management. ” The essay notes that exact phrasing did not appear in Stossel’s video or in statements by his main source, Michael Shellenberger.

The Science Feedback fact-check predated Stossel’s video by weeks and was itself a repost drawn from an earlier review of a Shellenberger article. When Stossel appealed the label, the organization did not respond. He later interviewed two experts listed on the fact-check page, Stefan Doerr of Swansea University and Zeke Hausfather of The Breakthrough Institute.

According to the essay, neither had watched the video before speaking with him, and both appeared to question its application to his content. The essay presents these examples to illustrate how AI-assisted fact-checking linked disparate content and applied generalized reviews without direct analysis of the new material.

U.S. program at the time. The essay reports that its reviewers held science degrees but characterizes them as activists. It notes that the organization did not typically link directly to the original claims it reviewed. In both the Tierney and Stossel cases, clicking the Facebook warning led users to fact-check pages that did not mention the specific article or video in question.

The essay compares the handling of evidence in Tierney’s piece to tactics once used by the tobacco industry against early smoking research.