Scholarlink

<p>Generative Artificial intelligence (AI) technologies allow users to create novel images and modify existing images easily and rapidly. However, the training sets used to create the generative technologies may be biased; they may be primarily trained on images of White men (Karkkainen and Joo (2021) FairFace: face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. Proceedings/IEEE Workshop on Applications of Computer Vision, 1547–1557. 10.1109/WACV48630.2021.00159, <a href="https://doi.org/10.1109/WACV48630.2021.00159">https://doi.org/10.1109/WACV48630.2021.00159</a>). When the training is biased, the output will also be biased. In this case, generating realistic faces of White men may be done with ease, but varying race, emotional expression, or gender may reduce realism. To determine the degree to which this bias exists, AI-generated faces were created using Leonardo.ai’s image-to-image algorithm, with two genders (man and woman), three races (Asian, Black, and White), and four emotional expressions (anger, fear, happiness, and sadness). N = 138 participants were presented with real and AI-generated faces and classified them as such. A 2 × 3&#xa0;×&#xa0;4 repeated measures ANOVA was run with sensitivity (d’) as the dependent variable. Participants were significantly better at identifying AI-generated faces of people of colour, more specifically, men of colour. Criterion values were also examined; participants had a bias towards classifying all faces except for those of White men as AI-generated, which they tended to classify as real. Our results support previous findings showing that demographic biases exist in generative AI tools and expand on those by showing how such biases interact with the emotional expression of the face.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Not all AI-generated faces are created equal: impacts of model gender, race, and emotional expression on classification accuracy

  • 期刊论文

  • 发表日期：

    2025年10月14日

查看全文
  • Megan Lawrence,
  • Kasey N. E. Cimermanis,
  • Charles A. Collin

摘要

Generative Artificial intelligence (AI) technologies allow users to create novel images and modify existing images easily and rapidly. However, the training sets used to create the generative technologies may be biased; they may be primarily trained on images of White men (Karkkainen and Joo (2021) FairFace: face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. Proceedings/IEEE Workshop on Applications of Computer Vision, 1547–1557. 10.1109/WACV48630.2021.00159, https://doi.org/10.1109/WACV48630.2021.00159). When the training is biased, the output will also be biased. In this case, generating realistic faces of White men may be done with ease, but varying race, emotional expression, or gender may reduce realism. To determine the degree to which this bias exists, AI-generated faces were created using Leonardo.ai’s image-to-image algorithm, with two genders (man and woman), three races (Asian, Black, and White), and four emotional expressions (anger, fear, happiness, and sadness). N = 138 participants were presented with real and AI-generated faces and classified them as such. A 2 × 3 × 4 repeated measures ANOVA was run with sensitivity (d’) as the dependent variable. Participants were significantly better at identifying AI-generated faces of people of colour, more specifically, men of colour. Criterion values were also examined; participants had a bias towards classifying all faces except for those of White men as AI-generated, which they tended to classify as real. Our results support previous findings showing that demographic biases exist in generative AI tools and expand on those by showing how such biases interact with the emotional expression of the face.