Year: 2023
Type: dissertation
Abstract: The utilization of artificial intelligence (AI) is significantly increasing. Cognitive models have begun to perform tasks such as image, voice, and text recognition - tasks previously considered difficult to implement with traditional computers - at a human level. Discriminative models are widely used to assist or automate human decision-making processes. Through generative models, humans are now able to communicate naturally with AI, and AI can create artistic works. Alongside the increase in these applications, there is growing concern about the risk of AI models reproducing existing societal biases embedded in training data, leading to discriminatory impacts. \nOne of the major approaches to mitigating such risks involves conducting fairness audits on AI systems. Fairness auditing is considered a crucial regulatory mechanism for ensuring AI fairness in compliance with anti-discrimination laws, personal data protection laws, and AI regulatory laws. Consequently, there is a pressing need to establish standards that precisely define the concept of fairness, and to develop methodologies to assess whether the AI under audit meets these standards. To this end, researchers have proposed various operational definitions for fairness metrics to be employed in audits. However, with the demonstration of the impossibility theorem, which states that several fairness standards cannot be simultaneously satisfied – especially for classification models – the issue of selecting appropriate fairness criteria for audits has become increasingly complex. This study investigates which fairness criteria can be normatively justified within the specific context of AI use cases, considering classification and generative models separately. \nIn the case of classification models, audits typically focus on whether the models predictions result in disparate treatment or impacts for socially marginalized groups that have historically faced discrimination. However, applying legal standards developed for anti-discrimination laws to AI-driven discrimination poses several challenges. Additionally, technically defined fairness metrics seem to conflict with existing anti-discrimination laws. To overcome these challenges, this study adopts a novel approach using simulations. Hypothetical discriminatory training data was generated across various scenarios to train a classification model. Following this, the ways in which the model violates certain fairness criteria were investigated, thereby deriving suitable fairness criteria for each scenario. The simulation results suggest that, while it is valid to consider parity of false positive and negative rates as primary evaluation indicators for cognitive models, parity of positive and negative predictive values (often referred to as 'calibration') is a suitable fairness criterion for discriminative models. Furthermore, this study points out the limitations of previously suggested fairness metrics and proposes new directions for alternative fairness audits. \nOn the other hand, fairness audits for generative models present different characteristics compared to those for classification models. While classification models, which typically assist or automate the decision-making process, could potentially lead to allocative harms, generative models pose the risk of causing representative harms. These harms are cultural and pervasive across society, rather than influencing specific decisions. Furthermore, representative harms can trigger various problems, such as the propagation and reinforcement of social bias and stereotypes, as well as the creation and dissemination of harmful expressions. In such cases, it becomes challenging to identify victims or quantify damages, thus complicating regulation via traditional legal remedies and posing difficulties for regulation within existing administrative systems. As a result, an ex ante self-audit by the providers of generative models is considered inevitable. In this regard, this study reviews concrete metrics and evaluation methods for fairness audits of generative models, categorizing them according to dimensions of bias, fairness in a narrow sense, and toxicity. Furthermore, it underscores the need to establish benchmark data and evaluation metrics capable of assessing the multifaceted legal and ethical risks posed by large-scale language models and image generation models, in order to effectively regulate the legal and ethical risks that generative models can instigate.
Author: 김병필
Institution: Seoul National University
Language: Korean
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7072852728
