Adultification Bias in LLMs and Text-to-Image Models
Project Overview
The document explores the implications of generative AI, particularly large language models (LLMs) and text-to-image (T2I) models, in education, focusing on the critical issue of adultification bias. It reveals that these AI models often portray Black girls as older and more sexualized compared to their White counterparts, which can lead to detrimental effects in educational settings, including biased disciplinary actions. The findings indicate that current alignment techniques fail to adequately address these biases, raising concerns about the potential reinforcement of stereotypes against marginalized groups. Consequently, the document underscores the necessity for rigorous evaluation and ethical considerations in the deployment of generative AI in educational contexts to mitigate harms and ensure equitable treatment for all students.
Key Applications
Measurement and definition of adultification bias in AI models
Context: Educational research on AI ethics, particularly concerning biases in interactions with minors, relevant for educators, policymakers, and researchers.
Implementation: This study involved prompting language models to define adultification bias and measuring explicit and implicit biases in popular LLMs and T2I models. Outputs related to Black, White, Asian, and Latina girls were compared to reveal biases.
Outcomes: Evidence of adultification bias was found in LLMs, where Black girls were subjected to harsher treatment and portrayed as older and more sexualized in T2I model outputs. While models correctly identified adultification bias, they still exhibited biased outputs, highlighting the need for awareness in AI applications.
Challenges: Existing alignment techniques are ineffective in comprehensively addressing biases, potentially leading to harm in educational settings. Models may produce biased outputs despite being trained to recognize biases, reflecting the complexity of AI ethics in education.
Implementation Barriers
Technical
Current alignment methods are insufficient to address various forms of bias in generative AI.
Proposed Solutions: There is a need for improved alignment techniques that specifically target biases like adultification.
Ethical
The existence of biases in models can lead to systemic inequalities and stigmatization, particularly affecting minors. Generative AI models can perpetuate biases, such as adultification bias, which can affect educational outcomes and fairness.
Proposed Solutions: Regulatory frameworks should address the issues of adultification and sexualization of minors in AI systems. Implementing rigorous evaluation and bias mitigation techniques in AI training and deployment.
Project Team
Jane Castleman
Researcher
Aleksandra Korolova
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: Jane Castleman, Aleksandra Korolova
Source Publication: View Original PaperLink opens in a new window
Project Contact: Dr. Jianhua Yang
LLM Model Version: gpt-4o-mini-2024-07-18
Analysis Provider: Openai