Algorithmic bias

Machine learning reproduces the distortions in its training data — and amplifies them through the recommendation loop. In dating apps this is paired with an economic incentive that actively devalues fairness.

Racial bias

Cornell studies document: platforms with explicit ethnic filters or recommender models extrapolating historic biased preferences show measurable racial segregation. [37]

Filter tools — “only show people of ethnic group X” — are not neutral personalization. They reproduce and legitimize structural prejudice in an area (partner selection) where most societies anchor anti-discrimination in law.

Research recommendation: redesign selection affordances, push cross-group recommendations, drop filters — not a technical question, a design choice. [37]

Popularity bias

Collaborative filtering amplifies majority preferences exponentially:

1. Profile A receives many likes → flagged as “relevant”.
2. Algorithm shows it to more users → even more likes.
3. The peak concentrates further, the middle slides down.

In economic simulations an unbiased algorithm (each person shown with equal probability) yields markedly higher average match rates — but lower platform revenue, because the illusion that “the top is always one swipe away” disappears. [38] Consequence: platform operators have active economic incentives to push popular profiles even though it lowers the average user’s match probability.

Socio-economic amplification

As discussed in algorithmic mechanics, recommender systems implicitly weight income and education signals differently by gender. [22] This weighting is not coded actively but learned from training material — and thus solidifies patterns that are sociologically problematic and politically contested.

What fairness frameworks do

FAIR-MATCH and similar multi-objective approaches try to mitigate bias via: [26, 28]

• Demographic balance constraints: no group may be systematically under-recommended.
• Reciprocity weighting: score not for “X likes Y” but for “X and Y mutually like each other”.
• Diversity bonus: recommendation list includes profiles outside the historic cluster.

In practice such frameworks are rarely implemented — they reduce the engagement KPIs described in monetization.

What this means for users

• The people the app shows you frequently are not necessarily the most compatible — they are the most popular within your filter cone.
• Filters marketed as “comfort features” (ethnic, religious, education) are, in aggregate, documented drivers of segregation. [37]
• What you do not see is decided algorithmically — not by your explicit preferences.

Bias in recommender systems is not a matter of “bad code”. It is the sum of data distortion, business model, and design choice — and therefore correctable if the incentive exists.