[论文解读] A Taxonomy of the Biases of the Images created by Generative Artificial Intelligence

Adriana Fernández de Caleya Vázquez, Eduardo C. Garrido‐Merchán|arXiv (Cornell University)|May 2, 2024

Impact of AI and Big Data on Business and Society被引用 7

一句话总结

本论文提出了AI生成图像中偏见的分类法，分析了它们的社会影响，并讨论了缓解策略与监管考量。

ABSTRACT

Generative artificial intelligence models show an amazing performance creating unique content automatically just by being given a prompt by the user, which is revolutionizing several fields such as marketing and design. Not only are there models whose generated output belongs to the text format but we also find models that are able to automatically generate high quality genuine images and videos given a prompt. Although the performance in image creation seems impressive, it is necessary to slowly assess the content that these models are generating, as the users are uploading massively this material on the internet. Critically, it is important to remark that generative AI are statistical models whose parameter values are estimated given algorithms that maximize the likelihood of the parameters given an image dataset. Consequently, if the image dataset is biased towards certain values for vulnerable variables such as gender or skin color, we might find that the generated content of these models can be harmful for certain groups of people. By generating this content and being uploaded into the internet by users, these biases are perpetuating harmful stereotypes for vulnerable groups, polarizing social vision about, for example, what beauty or disability is and means. In this work, we analyze in detail how the generated content by these models can be strongly biased with respect to a plethora of variables, which we organize into a new image generative AI taxonomy. We also discuss the social, political and economical implications of these biases and possible ways to mitigate them.

研究动机与目标

• 识别并对AI模型生成的图像中存在的偏见进行分类。
• 解释训练数据、模型和用户解读如何导致偏见。
• 讨论这些偏见在图像生成中的社会、政治和经济影响。
• 提出缓解策略并指出政策和未来研究的空白。

提出的方法

• 在文化、社会经济、生物以及人口统计维度上建立结构化的图像生成偏见分类法。
• 解释偏见如何源自训练数据、模型参数和输出解读。
• 使用概率框架 p(Y|X,θ) 描述偏见产生的技术直觉，并讨论正则化/数据增强作为缓解手段。
• 回顾相关工作和监管考量，将该分类法置于伦理与政策讨论之中。
• 提供一个面向开发者的实用清单，以减少图像生成中的偏见。

实验结果

研究问题

• RQ1 AI生成的图像中表现出的类别及具体偏见有哪些？
• RQ2 数据集、模型架构和提示如何促成这些偏见？
• RQ3 带有偏见的AI生成影像的社会、经济和政治影响是什么，如何进行缓解？
• RQ4 开发者和政策制定者可以采取哪些实际步骤来减少和规范图像生成中的偏见？

主要发现

• 一个全面的分类法识别了AI生成图像中的文化、社会经济、生物学和人口统计偏见。
• 偏见源自训练数据分布、模型优化和提示解读，可能助长刻板印象。
• 缓解措施可包括数据增强、训练中的正则化，以及偏见感知评估测试集。
• 对于透明度、问责制和缓解措施有效性，需要在伦理、政策和监管方面进行更广泛的讨论。
• 研究强调将偏见缓解扩展到性别和种族之外的更广泛的偏见集合，在图像生成中。