How Fake News Affect Trust in the Output of a Machine Learning System for News Curation

Users struggle to identify contaminated recommendations: When a single fake news story was mixed with four trustworthy articles, users rated the batch similarly to one containing five trustworthy stories, suggesting people have difficulty spotting misinformation when it appears alongside credible content.

Trustworthy context can shield untrustworthy content: The study indicates that fake news may "benefit from appearing in a trustworthy context"—meaning users' trust in an algorithm's overall quality can inadvertently make them more accepting of false information within those recommendations.

User trust ratings have significant limitations: While participants could distinguish between fully trustworthy and fully untrustworthy recommendations, their ability to evaluate mixed batches was unreliable, revealing a gap between users' perceived ability to assess algorithmic outputs and their actual performance.

Algorithm design choices have real consequences for misinformation spread: Since people increasingly rely on ML systems for news curation, the way these systems present recommendations directly affects whether users accept or reject false information, making responsible curation design a critical responsibility.

Current user interfaces may not adequately support informed decision-making: The findings suggest that existing interactive ML systems for news may not provide users with sufficient tools or information to make trustworthy judgments about recommendation quality, highlighting a design challenge for responsible AI systems.