Gmail's email classification algorithms systematically diverted nearly 90% of political emails from advocacy groups and candidates to promotions or spam folders, causing significant drops in open rates and reduced donations for political organizations.
Between 2018 and 2020, multiple political advocacy groups and candidates reported that Gmail's automated email classification system was diverting their emails away from users' primary inboxes. In early 2018, advocacy groups including Democracy for America, CREDO Action, and Change.org noticed drops in open rates for Gmail subscribers of at least 47%, with Change.org experiencing a 79% drop. To investigate this systematically, The Markup created a new Gmail account in October 2019 and subscribed to emails from 231 political entities including presidential candidates, House representatives, and advocacy organizations. Over four months, they collected 5,134 emails from 171 groups. The results showed Gmail diverted 89.37% of political emails from the primary inbox, with 50% going to promotions and 40% to spam. Only 10.64% reached the primary inbox. Some candidates like Elizabeth Warren, Joe Biden, and Kamala Harris had zero emails reach the primary inbox, while Pete Buttigieg had 63% reach primary. The advocacy groups that initially reported problems had between 60-90% of their emails sent to promotions. This algorithmic filtering significantly impacted political fundraising and engagement, as emails in promotions folders have much lower open rates than those in primary inboxes.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
AI-driven concentration of power and resources within certain entities or groups, especially those with access to or ownership of powerful AI systems, leading to inequitable distribution of benefits and increased societal inequality.
AI system
Due to a decision or action made by an AI system
Unintentional
Due to an unexpected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed