The paper "Beyond Behaviorist Representational Harms: A Plan for Measurement and Mitigation" by Jennifer Chien and David Danks addresses the issue of representational harms, which are less tangible but equally significant in their impact compared to allocative harms. The authors critique existing definitions of representational harms, which often focus on observable behaviors and neglect cognitive, affective, and emotional changes. They propose an expanded definition that includes changes in mental representations, such as attention, engagement, beliefs, and emotions, which can lead to harm. The paper outlines the challenges in measuring and mitigating representational harms, emphasizing the need for interdisciplinary expertise and community involvement. It highlights the unique vulnerabilities of large language models (LLMs) in perpetuating these harms, particularly when they go unmeasured and unmitigated. The authors provide a case study to illustrate the potential harms and measures, and discuss proposed mitigations, including seamful design and counter-narratives, to reduce the likelihood and severity of representational harms. The paper concludes by discussing morally permissible cases where representational harms may be necessary to achieve a more ethical state, and emphasizes the importance of measuring and mitigating these harms to ensure fair and equitable outcomes.The paper "Beyond Behaviorist Representational Harms: A Plan for Measurement and Mitigation" by Jennifer Chien and David Danks addresses the issue of representational harms, which are less tangible but equally significant in their impact compared to allocative harms. The authors critique existing definitions of representational harms, which often focus on observable behaviors and neglect cognitive, affective, and emotional changes. They propose an expanded definition that includes changes in mental representations, such as attention, engagement, beliefs, and emotions, which can lead to harm. The paper outlines the challenges in measuring and mitigating representational harms, emphasizing the need for interdisciplinary expertise and community involvement. It highlights the unique vulnerabilities of large language models (LLMs) in perpetuating these harms, particularly when they go unmeasured and unmitigated. The authors provide a case study to illustrate the potential harms and measures, and discuss proposed mitigations, including seamful design and counter-narratives, to reduce the likelihood and severity of representational harms. The paper concludes by discussing morally permissible cases where representational harms may be necessary to achieve a more ethical state, and emphasizes the importance of measuring and mitigating these harms to ensure fair and equitable outcomes.