This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
An investigation by The Markup found that Gmail's automated inbox sorting algorithm consistently diverted a large majority of political emails from the 'Primary' inbox to 'Promotions' or 'Spam' folders. Advocacy groups and political campaigns reported that this categorization significantly reduced email open rates, thereby hindering fundraising and political mobilization. While Google maintains that the algorithm adjusts to user preferences, the study highlights the outsized impact of Gmail's curation on democratic processes.
Google's Gmail inbox sorting algorithm for political emails was reported by presidential candidates, nonprofits, and advocacy groups for having negative impact on call-to-actions, allegedly suppressing donations and impeding political actions.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
AI systems that fail to perform reliably or effectively under varying conditions, exposing them to errors and failures that can have significant consequences, especially in critical applications or areas that require moral reasoning.
Also classified as
The algorithm treated different political candidates and organizations highly unevenly, resulting in disparate outcomes for their campaigns.
The incident stems from Gmail's classification algorithm failing to robustly and accurately distinguish between commercial marketing and political/advocacy communication, leading to high rates of misclassification.
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
The incident was caused by Gmail's deployed classification algorithm, which unintentionally filtered political communication into promotions and spam, impacting democratic engagement.
Gmail's automated email sorting algorithm unintentionally diverted a large majority of political emails to spam or promotions folders, unevenly affecting political candidates and suppressing fundraising. While raising concerns about the algorithmic curation of democratic discourse, the national security impact is minor and manageable within standard regulatory and platform governance procedures.
The report describes Gmail's mail classifications as automatically adjusting to match users' preferences and actions. The algorithm operates independently to sort incoming emails into default tabs like promotions or spam in real-time.
Minimal or No Risk: The EU AI Act explicitly lists 'spam filters' as an example of AI systems that pose low or negligible risks to users and society, falling under Risk Level 4.
The primary function of Gmail's algorithm is to filter spam and curate/categorize incoming emails into default tabs like promotions to help users organize their inboxes.
The reports provide detailed empirical data from a structured four-month experiment analyzing over 5,000 emails across 171 political groups. The findings are supported by statistical drops in open rates and direct statements from affected organizations and Google representatives.
The filtering could be interpreted purely as a standard spam/promotions filter working as intended to reduce inbox clutter, rather than a systemic threat to democratic engagement.
Technology
Algorithmic Sorting Flaws
Algorithm misclassifies political emails as promotions or spam.
Inconsistent Classification
Similar emails treated differently across candidates and groups.
Lack of Explicit Primary Label
Primary inbox is determined by excluding other categories.
Data Inputs
User Interaction Bias
Algorithm relies heavily on aggregate and individual user actions.
Mislabeled Senders
List sharing led to incorrect attribution of email senders.
Metadata Classification Limits
Algorithm relies on metadata fields like X-Gmail-Labels.
Human Factors
Default Settings Reliance
Users rarely disable default tabbed layout settings.
User Interaction Variance
Varying user habits cause unpredictable sorting behavior.
Lack of User Training
Users do not manually correct sorting errors often.
Process and Methods
Opaque Sorting Logic
Lack of transparency on how the algorithm classifies emails.
Inadequate Edge Case Handling
Official non-campaign emails were sent to promotions.
Lack of Appeal Mechanisms
Senders have no way to contest incorrect sorting.
Management
Monetization Conflict of Interest
Suggesting buying ads incentivizes poor free reach.
Prioritizing Organization Over Reach
Focusing on user inbox organization over sender visibility.
Regulatory Environment
Absence of Platform Regulations
No rules govern algorithmic curation of political speech.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
