Do Perceived Gender Biases in Retrieval Results Affect Relevance Judgements?

Gender stereotypes in search results influence how people rate relevance: The study found that when documents contained gender-stereotypical content, participants gave different relevance scores to the same information depending on which gender was represented, suggesting that bias in search results can shape user judgments rather than users making purely objective assessments.

The problem affects both male and female evaluators: Experiments comparing judgments when participant gender was known versus unknown revealed that gender stereotypes influenced relevance ratings across different demographic groups, indicating this is a systemic issue rather than one limited to particular users.

Relevance judgments are more complex than they appear: Because relevance in search systems is "multi-dimensional," the presence of gender bias in retrieved documents can affect how people evaluate whether results actually answer their questions, potentially creating a feedback loop where biased results are rated as more relevant.

Current information retrieval systems may amplify gender stereotypes: If biased documents are consistently rated as more relevant due to stereotype confirmation, search algorithms could learn to prioritize and promote these biased results, making the problem worse over time.

This research highlights a blind spot in AI evaluation: The study raises awareness that the people who judge whether search results are "good enough" may themselves be influenced by stereotypes, meaning quality control processes for search systems could inadvertently perpetuate gender bias rather than catch it.