Do Fictive Sexual Orientations Induce In-Group Bias in Emotion Recognition?

Deep learning techniques can now recognize sexual orientation from faces pictures more accurately than humans (70% vs. 60%). This advance in artificial intelligence can be problematic due to its lack of accuracy and the natural in-group bias of humans. More visible social characteristics (race, age, or gender) are known to induce such biases especially when it comes to emotion recognition. However, it is unknown whether the same would occur with less perceptible characteristics such as sexual orientation. In the present study, we investigated the effect of fictive sexual orientations on emotional face recognition. In our first experiment, we found no in-group bias when target pictures were randomly associated with fictive sexual orientations. In a second experiment, we observed that faces pre-evaluated as homosexual were recognized faster than heterosexual faces, but only when paired with an incongruent fictive sexual orientation. While it remains uncertain what factors guided participants' responses, our results highlight several limitations to further understand how AI and humans judge persons' characteristics from faces, as well as the potential biases that may result.