Casual Conversations features 45,186 videos of 3,011 participants based in the US who have explicitly provided their age and gender themselves, without being labelled by third parties or machine learning models.

A team of trained annotators then labelled the lighting conditions and used the Fitzpatrick Scale to determine participants’ apparent skin tones, with the purpose of helping AI developers examine how their technology performs across a diverse range of people and lighting environments.

The conversations in the videos are non-scripted and allow AI researchers to examine whether their computer vision and audio models fairly represent people based on their age, gender, skin tone and ambient lighting.

Facebook said it hoped the study would help address concerns about implicit bias in AI and algorithms, which can unfairly impact on certain individuals and groups.

A blog written by the research team behind Casual Conversations said: “As an industry and research community, we are at the beginning of understanding the multifaceted, ongoing challenges of fairness and bias.

“By making fairness research more transparent and normalising subgroup measurement, we hope this data set brings the field one step closer to building fairer, more inclusive technology.”