Algorithmic bias

Inequality, bias and discrimination

These concepts overlap but are not identical. Unequal outcomes can reflect preferences and opportunities. Missing exposure can bias measurement before ranking is even assessed. Discrimination requires closer examination of groups, rules and context.

Hutson and colleagues discuss how intimate-platform design can amplify or reduce exclusion, including through filters and representation. Their work does not establish a percentage of all apps that discriminate. 16

Popularity bias and business objectives

Celdir, Cho and Hwang combine data from an Asian platform with an economic model. Our previous account was incorrect: equally frequent recommendations produced lower revenue and fewer matches in the studied model. Revenue and matching therefore need not inherently conflict. 14 15

The result is contextual. It justifies neither every concentration of visibility nor a claim that redistribution must always hurt. Assessing an actual platform requires its data, objectives and side effects.

Requirements for a fairer comparison

Exposure, mutual interest, contact quality and safety can be assessed separately. Improvement in one dimension may worsen another. Missing or inaccurate group information also limits comparisons.

CupidLeaks therefore promises no universally fair algorithm. A lower simulated Gini initially means only a more even distribution of simulated likes. It proves neither satisfaction, absence of discrimination nor real matchmaking success.