Auditing Facial Emotion Recognition Datasets for Posed Expressions and Racial Bias
[ABSTRACT READ IN FULL at the OpenAlex work record, reconstructed from its deposited inverted index. FULL TEXT NOT READ: closed access, no open-access location at OpenAlex. PEER-REVIEWED CONFERENCE PROCEEDINGS, which is a real venue and weaker than a journal, so call it a 2025 conference paper rather than a study] Khan, R., & Stinson, C. (2025). 'Auditing Facial Emotion Recognition Datasets for Posed Expressions and Racial Bias.' 2025 3rd International Conference on Foundation and Large Language Models. doi:10.1109/fllm67465.2025.11391216. ADDED ON THE CLOSING PASS OF 2026-10-21 BECAUSE IT KILLS AN ABSENCE CLAIM THIS FILE WAS ABOUT TO SHIP. The brief asked whether any published evaluation reports differential emotion-recognition accuracy by skin tone as opposed to by gender. This one reports differential outcomes by skin tone. The authors audit two state-of-the-art facial emotion recognition datasets, sample images to test whether they are posed, and then, verbatim: 'We also observe the skin color of individuals in the samples, and test three models trained on each datasets to predict facial expressions of people from various races and skin tones. We find that the audited models were more likely to predict people who were not labeled as white or determined to have dark skin tones, as showing a negative emotion such as anger or sadness even when they were smiling.' HOLD IT AT EXACTLY THIS ANGLE: it is not an accuracy table by Fitzpatrick type and it attaches no percentage to anything we can see. It is a directional finding about which label the model assigns. The skin-tone labels are the authors' own observation of sampled images, which is the same methodological soft spot that the Dominguez-Catena dataset audit names about its own predicted protected attributes. ATTACH NO NUMBER TO IT.
[ABSTRACT READ IN FULL at the OpenAlex work record, reconstructed from its deposited inverted index. FULL TEXT NOT READ: closed access, no open-access location at OpenAlex. PEER-REVIEWED CONFERENCE PROCEEDINGS, which is a real venue and weaker than a journal, so call it a 2025 conference paper rather than a study] Khan, R., & Stinson, C. (2025). 'Auditing Facial Emotion Recognition Datasets for Posed Expressions and Racial Bias.' 2025 3rd International Conference on Foundation and Large Language Models. doi:10.1109/fllm67465.2025.11391216. ADDED ON THE CLOSING PASS OF 2026-10-21 BECAUSE IT KILLS AN ABSENCE CLAIM THIS FILE WAS ABOUT TO SHIP. The brief asked whether any published evaluation reports differential emotion-recognition accuracy by skin tone as opposed to by gender. This one reports differential outcomes by skin tone. The authors audit two state-of-the-art facial emotion recognition datasets, sample images to test whether they are posed, and then, verbatim: 'We also observe the skin color of individuals in the samples, and test three models trained on each datasets to predict facial expressions of people from various races and skin tones. We find that the audited models were more likely to predict people who were not labeled as white or determined to have dark skin tones, as showing a negative emotion such as anger or sadness even when they were smiling.' HOLD IT AT EXACTLY THIS ANGLE: it is not an accuracy table by Fitzpatrick type and it attaches no percentage to anything we can see. It is a directional finding about which label the model assigns. The skin-tone labels are the authors' own observation of sampled images, which is the same methodological soft spot that the Dominguez-Catena dataset audit names about its own predicted protected attributes. ATTACH NO NUMBER TO IT.
Written from the aBSTRACT READ IN FULL at the OpenAlex work record, reconstructed from its deposited inverted index. FULL TEXT NOT READ: closed access, no open-access location at OpenAlex. PEER-REVIEWED CONFERENCE PROCEEDINGS, which is a real venue and weaker than a journal, so call it a 2025 conference paper rather than a study.
What our sources record
Recorded when this source was checked, figures included, with page references and quotation marks removed.
2025 3rd International Conference on Foundation and Large Language Models. doi:10.1109/fllm67465.2025.11391216