Coffee Meets Bagel’s Algorithm Reported by Users Disproportionately Showing Them Matches of Their Own Ethnicities Despite Selecting “No Preference”

Users selecting “no preference” were shown by Coffee Meets Bagels’s matching algorithm more potential matches with the same ethnicity, which was acknowledged and justified by its founder as a means to maximize connection rate without sufficient user information.

Risk classification

• Primary risk domain: 1 Discrimination & Toxicity
• Primary risk subdomain: 1.1 Unfair discrimination and misrepresentation

The algorithm treated users unequally based on their race, resulting in racial siloing and unfair representation of potential matches.

Additional risk subdomains

• 7.3 Lack of capability or robustness: The system failed to robustly handle the 'no preference' input, defaulting to biased empirical averages instead of a diverse distribution.

Causal factors

• Entity: AI
• Intent: Unintentional
• Timing: Post-deployment

The racial siloing was an unexpected outcome of the AI matching algorithm's optimization process after it was deployed to users.

EU AI Act risk tier

• Risk tier: 4 Minimal or No Risk

Minimal or No Risk: The AI system is a matchmaking algorithm for a dating application, which is a consumer entertainment and social platform posing low or negligible risk to safety or fundamental rights under the EU AI Act framework.

AI system and alleged parties

• AI system: Coffee Meets Bagel matching algorithm
• AI purpose: Matchmaking; Content Recommendation
• Behaviour type: Assistant
• Alleged developer: Coffee Meets Bagel
• Alleged deployer: Coffee Meets Bagel
• Alleged harmed parties: Coffee Meets Bagel users having no ethnicity preference, Coffee Meets Bagel users

Harm severity

Severity ratings are based on this scale.

Highest direct severity in any category: Minor. Severity is scored from Negligible to Catastrophic in each harm category, for harm the reports describe as caused directly or indirectly by the AI system.

• Physical: direct Negligible, indirect Negligible
• Infrastructure: direct Negligible, indirect Negligible
• Property: direct Negligible, indirect Negligible
• Financial: direct Negligible, indirect Negligible
• Environmental: direct Negligible, indirect Negligible
• Malicious content: direct Negligible, indirect Negligible
• Differential treatment: direct Minor, indirect Minor
• Civil rights: direct Negligible, indirect Negligible
• Democracy: direct Negligible, indirect Negligible
• Privacy: direct Negligible, indirect Negligible
• Psychological: direct Negligible, indirect Negligible
• Epistemic: direct Negligible, indirect Negligible
• Child sexual exploitation and abuse: direct Negligible, indirect Negligible

Differential treatment

Reported: Yes

Directly caused: The algorithm systematically provided different match recommendations to users based on their race, resulting in racial siloing.

Indirectly caused: Users felt pressured to manually exclude certain racial groups in their settings to bypass the algorithm's biased defaults.

Inferred additional harm: A significant portion of the 60,000 users likely experienced racially biased match suggestions due to the algorithm's design.

People affected

• Occurrences reported: 1
• People reportedly harmed: 10
• People reportedly exposed: 60000

Potential causes

Management

• Prioritizing Connection Rates: Management prioritized maximizing connection rate over user experience.
• Defensive Stance on Bias: Management dismissed user bias concerns as empirical user preferences.

Technology

• Flawed Match Optimization: Algorithm prioritizes overall connection rates over match diversity.
• Game Theory Feedback Loop: System forces users into exclusionary preferences to get diverse matches.
• No Preference Misinterpretation: Algorithm treats 'no preference' as a license to dump skewed matches.

Data Inputs

• Skewed Demographic Data: User base is highly skewed toward white, Asian, and Jewish demographics.
• Empirical Bias Default: Using historical system-wide matching data to predict individual desires.
• Lack of Diverse Match Data: System lacks diverse interaction data for users selecting no preference.

Human Factors

• Explicit Racial Preferences: Many users limit their preferences to their own ethnicity.
• User Expectation Disconnect: Users expect 'no preference' to yield a diverse mix of matches.
• Game Theory Pressure: Users feel forced to adopt exclusionary settings to bypass algorithmic bias.

Process and Methods

• Inadequate Algorithmic Testing: Failure to test how 'no preference' settings behave under skewed demand.
• Lack of Diversity Metrics: No mechanism to measure or preserve diversity in match recommendations.

Regulatory Environment

• Absence of Algorithmic Standards: No regulatory standards governing bias in dating recommendation algorithms.
• Lack of Oversight: No external auditing for discriminatory outcomes in matching systems.

Information quality

• Classification confidence: High
• Reason for confidence: The reports provide clear, first-hand accounts of the matching algorithm's behavior, along with direct explanations from the company's co-founder (Dawoon Kang) confirming the empirical data-driven matching strategy. The mechanics of the racial siloing are well-explained through both user experience and game theory analysis.
• Ambiguities identified: None significant; the algorithm's behavior and the company's justification are clearly outlined.
• Alternative interpretations: The matching behavior could be interpreted as a reflection of existing user biases rather than an algorithmic failure, though the system's design actively reinforced and amplified these biases.

In 2013, the Coffee Meets Bagel dating app algorithm was found to disproportionately suggest same-ethnicity matches due to optimization for connection rates. This algorithmic bias caused user frustration but represents a consumer-level issue with negligible implications for national security.

• Overall national security impact: Negligible
• Scope: Single nation
• Primary target: United States
• Alleged perpetrator: Unknown

Threat characteristics

• Imminence: Long-term. The incident represents a past consumer-level algorithmic issue with no immediate or near-term national security threat.
• Autonomy: Human-controlled. The AI acts as a recommendation assistant, but human users retain full control over final matching decisions.
• Novelty: Established threat. Algorithmic bias and feedback loops in recommendation systems are well-documented and established issues.

Impact by dimension

• Physical security: Negligible. The incident involves a consumer dating application and has no impact on physical systems, critical infrastructure, or human safety.
• Information security: Negligible. There is no indication of information warfare, state-sponsored disinformation, or compromise of intelligence capabilities.
• Sovereignty: Negligible. The incident does not affect state authority, electoral systems, or core government operations.
• Economic security: Negligible. The algorithmic bias in a dating app does not threaten financial systems, strategic industries, or national technological advantages.
• Societal stability: Negligible. While the algorithm exhibited racial bias and created user frustration, it occurred within a limited consumer user base of 60,000 and poses negligible threat to national societal stability or systemic human rights.