[Paper Review] 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 Society7 citations

TL;DR

The paper proposes a taxonomy of biases in AI-generated images, analyzes their social impacts, and discusses mitigation strategies and regulatory considerations.

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.

Motivation & Objective

• Identify and categorize biases present in images generated by AI models.
• Explain how training data, models, and user interpretation contribute to biases.
• Discuss social, political, and economic implications of these biases in image generation.
• Propose mitigation strategies and identify gaps for policy and future research.

Proposed method

• Develop a structured taxonomy of image generation biases across cultural, socioeconomic, biological, and demographic dimensions.
• Explain how biases originate from training data, model parameters, and output interpretation.
• Describe technical intuition of why biases arise using a probabilistic framing p(Y|X,θ) and discuss regularization/data augmentation as mitigation.
• Review related work and regulatory considerations to situate the taxonomy within ethics and policy discussions.
• Offer a practical checklist for developers to reduce biases in image generation.

Experimental results

Research questions

• RQ1 What are the categories and specific biases that manifest in AI-generated images?
• RQ2 How do datasets, model architectures, and prompts contribute to these biases?
• RQ3 What are the social, economic, and political implications of biased AI-generated imagery, and how can they be mitigated?
• RQ4 What practical steps can developers and policymakers take to reduce and regulate biases in image generation?

Key findings

• A comprehensive taxonomy identifies cultural, socioeconomic, biological, and demographic biases in AI-generated images.
• Biases originate from training data distributions, model optimization, and prompt interpretation, and can perpetuate stereotypes.
• Mitigation can involve data augmentation, regularization in training, and bias-aware evaluation test batteries.
• Broader discussions on ethics, policy, and regulation are essential to address transparency, accountability, and effectiveness of mitigations.
• The study emphasizes extending bias mitigation beyond gender and race to a wider set of biases in image generation.