No Bias, No GAN: Why Generative Models Necessarily Produce Unexpected and Uninterpretable Biases

Recent advancements in generative AI have led to remarkable developments in text and image generation. However, a significant challenge is their propensity to generate biased outputs, including stereotypes and underrepresentation of certain groups. In this paper, we will focus on Generative Adversarial Network (GAN). GANs have revolutionized image generation with their ability to efficiently produce highly realistic images.

 

However, GANs, while fast and capable of producing high-quality outputs, often generate only a subset of classes in the training data, limiting the diversity of its outputs. As a result, the study of biases in GANs is particularly mature, providing a wealth of information about their mechanisms and potential solutions. This understanding is crucial because these issues in GANs can be generalized to other generative models. Therefore, focusing on GANs offers insights that are applicable to the broader field of generative AI. There are two important types of algorithmic biases. The first one is dataset bias (DB), which happens when training data for some groups are over- or under-represented in a dataset relative to their actual prevalence in the population. The second one is generative bias (GB), the phenomenon where a generative model, such as a GAN, generates a set of synthetic data that are biased in their representation (such as underrepresentation and misrepresentation) of different groups.

 

Empirical literature often suggests that while GBs are an inevitable outcome of GANs, they primarily reflect and exacerbate well-known societal biases in the dataset (Arora et al., 2018; Jain et al., 2022; Maluleke et al., 2022). In this paper, we argue that GANs not only replicate well-known DBs but also produce a plethora of unexpected and uninterpretable GBs due to mode collapse. This is because GBs reflecting well-known societal biases are merely special cases of a more general phenomenon. A much larger number of them are inevitably produced by GAN because:

​

1. GBs can reflect and exacerbate unknown DBs;

2. GBs can result from an imbalanced dataset with no DBs;

3. GBs can result from a dataset deemed balanced and unbiased by the relevant stakeholders.

4. The patterns of GBs can often be difficult to interpret at some level of description.

 

That is, we will show that GAN can generate an enormous amount of unexpected and uninterpretable GBs. These biases, overlooked in current literature, hold significant moral implications as they tend to systematically misrepresent and underrepresent minoritized groups. Importantly, these biases are not merely a reflection of human biases in the training data but a consequence of machine learning processes that can diverge from human perspective (e.g., our interpretive resources and background knowledge). As a result, these biases can be epistemically opaque and challenge our ability to detect, evaluate, and debias them effectively. They amplify the moral risks associated with GANs, risks that are more pervasive than previously acknowledged. We conclude by discussing the heightened moral risks posed by GANs and other generative models and propose new approaches to mitigate them.