As indicated by engineering experts consulted for this analysis and official statements from Facebook, Facebook’s basic artificial intelligence mistakenly classified a Kansas Reflector article on climate change as a safety issue, resulting in a series of errors that led to the blocking of domains associated with news outlets featuring the article.
This evaluation is consistent with an internal review by States Newsroom, the parent company of Kansas Reflector, which criticized Facebook for the inconsistencies in its AI algorithms and the lack of transparency regarding its errors.
Authorities have suggested that Facebook may not fully comprehend the reasons behind the failure, leaving uncertainty as to why Facebook’s AI interpreted the structure or content of the article as a potential threat.
Chris Fitzsimon, the leader and editor of States Newsroom, commented, “It appears that Facebook utilized overly aggressive and uncertain AI to incorrectly flag a Reflector article as a phishing attempt.” He also raised concerns about Facebook’s handling of the situation, describing it as convoluted and difficult to understand. The dissemination of inaccurate information to users, suggesting that their content posed security risks without correcting the misinformation for their followers on the platform, is troubling.
On April 4, Facebook restricted Kansas Reflector from sharing Dave Kendall’s climate change opinion piece and promptly deleted all posts mentioning any content from the website. Although these posts were eventually restored about a week later, the restriction on the paragraph persisted.
Subsequently, The Handbasket, a publication overseen by independent journalist Marisa Kabas, and News From the States, a newspaper under States Newsroom, both tried to publish the column the following day. However, Facebook rejected these posts and removed all links to both websites, mirroring its actions towards Kansas Reflector.
Twitter also took steps by deleting posts and informing users flagged for posing security risks due to fabricated identities.
Meta, Facebook’s parent company, issued a public apology for the “safety issue.” However, a Meta representative stated that Facebook would not actively contact clients to address the misinformation provided.
Visitors to the Kansas Reflector website are still unsure about the situation. Facebook’s interventions not only disrupted Kansas Reflector’s journalistic activities during the final days of the parliamentary session but also had a chilling effect on other media outlets.
Daniel Kahn Gillmor, a senior staff technologist at the American Civil Liberties Union, highlighted the risks of relying heavily on a single communication platform to determine the importance of various topics, noting that this is not Facebook’s core competency. Gillmor emphasized the challenges of expecting Facebook, originally a social networking tool for college students, to arbitrate factual accuracy in today’s information landscape.
Adam Mosseri, the head of Meta’s Instagram, attributed the error to machine learning algorithms, a specific type of AI trained to identify characteristics associated with phishing scams. Mosseri recognized that these classifiers assess millions of content pieces daily, leading to occasional mistakes.
Jason Rogers, CEO of Invary, a cybersecurity firm linked to the University of Kansas Innovation Park, reviewed Kendall’s column and suggested that Facebook’s sensors may have been triggered by factors such as the number of hyperlinks or page resolution. He highlighted the inherent limitations of AI systems and cautioned against overestimating their capabilities.
Sagar Samtani, director of Indiana University’s Kelley School of Business’ Data Science and Artificial Intelligence Lab, emphasized the learning curve for Facebook in discerning appropriate content globally to protect against malicious actors. Samtani stressed the inevitability of occasional misclassifications in this learning process.
In light of these developments, questions about accountability and transparency come to the forefront. While Meta’s spokesperson Andy Stone denied content-related motivations for Facebook’s actions, ACLU’s Gillmor expressed doubts about the platform’s ability to program AI to overlook semantic nuances in articles. The lack of clarity on Facebook’s accountability, understanding of the mistake’s origins, preventive measures, and Oversight Board’s involvement highlights the complexities of holding AI systems responsible for such errors.
Gillmor underscored the broader implications of heavy reliance on a single information ecosystem like Facebook, emphasizing the importance of accountability and transparency in addressing such incidents.
