The Casual Conversations v2 Dataset

Consent-driven approach sets ethical standard: The dataset includes 26,467 videos from 5,567 paid participants who explicitly agreed to have their data used for AI fairness research, establishing a model for responsible data collection that respects participant autonomy and compensates contributors.

Global representation enables bias detection across regions: By collecting data from seven countries (Brazil, India, Indonesia, Mexico, Vietnam, Philippines, and USA), the dataset allows researchers to identify whether computer vision and audio models exhibit different biases across geographic and cultural contexts, rather than assuming bias patterns are universal.

Comprehensive demographic labeling supports intersectional analysis: The dataset captures 11 different attributes including self-reported age, gender, language/dialect, disability status, and skin tone (using both Fitzpatrick and Monk scales), enabling researchers to examine how AI systems may discriminate based on multiple overlapping characteristics simultaneously.

Multiple annotation methods improve reliability: Using both self-reported data from participants and trained annotators for attributes like skin tone and voice timbre provides cross-validation opportunities and reduces the risk that a single labeling approach introduces systematic errors into fairness evaluations.

Practical scale supports real-world model testing: With nearly 5 videos per person on average, the dataset provides sufficient data volume to rigorously test whether commercial computer vision and speech recognition systems perform equitably across different demographic groups.