From Symptoms to Systems: An Expert-Guided Approach to Understanding Risks of Generative AI for Eating Disorders

Paper Title

From Symptoms to Systems: An Expert-Guided Approach to Understanding Risks of Generative AI for Eating Disorders

Publication Info

• Topic area: Risks of generative AI in mental health, specifically eating disorders.
• Keywords: Generative AI, eating disorders, mental health, risk taxonomy, thinspiration, AI safety, clinical insights, digital technologies, user vulnerability, participatory evaluation.

Background and Problem

• Problem / challenge: Existing safeguards in generative AI systems fail to address subtle but clinically significant risks for individuals vulnerable to eating disorders. Current evaluation methods often overlook nuanced, context-specific risks.
• Significance: Eating disorders are life-threatening conditions, and generative AI systems may exacerbate vulnerabilities through harmful outputs, perpetuation of stereotypes, and reinforcement of disordered behaviors.
• Motivation and related work: Previous research has explored AI risks broadly and the impacts of digital technologies on eating disorders, but these efforts lack clinically grounded, context-sensitive taxonomies. This paper builds on sociotechnical and HCI research to address this gap by involving domain experts.

Solution

• Proposed approach: Development of an expert-guided taxonomy of generative AI risks for eating disorders, informed by semi-structured interviews with 15 clinicians, researchers, and advocates.
• Novelty:
  1. Creation of a taxonomy with seven risk categories specific to eating disorders.
  2. Identification of how generative AI amplifies existing risks and introduces novel harm pathways.
  3. Demonstration of the importance of early engagement with domain experts to design contextually relevant safety evaluations.
• Procedure and key techniques:
  • Conducted semi-structured interviews with 15 experts in eating disorders.
  • Used abductive qualitative analysis to identify and categorize risks.
  • Iteratively refined the taxonomy based on expert feedback and alignment with prior research.
  • Developed seven risk categories: generalized health advice, encouraging disordered behaviors, symptom concealment, thinspiration, reinforcing negative self-beliefs, excessive focus on the body, and perpetuating narrow views of eating disorders.

Results

• Concrete findings:
  • Seven risk categories were identified, each with specific examples (e.g., AI providing calorie estimates, offering thinspiration images, or reinforcing negative self-beliefs).
  • AI systems may unintentionally validate distorted thinking, amplify cultural biases, and create harmful feedback loops.
• Advantage over baselines:
  • The taxonomy provides a more nuanced and clinically informed framework than existing AI risk assessments, which often generalize or mischaracterize risks for vulnerable populations.
• Experiments / evaluation:
  • Interviews with 15 experts provided qualitative insights.
  • Examples of real-world AI outputs (e.g., BMI advice, thinspiration images) were analyzed to illustrate risks.
• Limitations and future work:
  • Did not include individuals with lived experience of eating disorders, limiting insights into real-world interactions.
  • Did not systematically evaluate the prevalence of identified risks in current AI systems.
  • Future research should involve participatory approaches with diverse stakeholders and develop operationalizable metrics for risk assessment.

Summary

This study identifies and categorizes the risks generative AI systems pose to individuals vulnerable to eating disorders, creating a taxonomy of seven key risk areas. Through expert interviews and qualitative analysis, it highlights how AI interactions can exacerbate vulnerabilities via harmful outputs, cultural biases, and feedback loops. The taxonomy provides a clinically grounded framework for assessing and mitigating these risks, emphasizing the importance of engaging domain experts in AI safety design. Future work should focus on participatory research, systematic evaluations, and the development of targeted interventions to ensure AI systems are safe and supportive for at-risk users.