← CHI 2026 map

CHI '26 · Honorable mention · full-paper review · confidence medium-high

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

Amy Winecoff , Kevin Klyman

DOI PDF Program page Page JSON

This is a strong qualitative CHI paper because it turns a familiar safety question into a clinically grounded systems problem. The main value is not a detector or benchmark, but a carefully argued taxonomy that can guide future evaluation and safeguard design for a high-risk population.

Axes Lens

Rare contribution shape, typical evidence profile. The point here is not a score. It is to show what kind of claim the paper makes, and whether the evidence pattern is unusual or baseline in this 268 -review set.

Review Summary

This paper’s contribution is best understood as a field-level reframing rather than a technical system advance. Instead of treating generative-AI harm as a matter of obviously unsafe outputs, it uses expert interviews and abductive analysis to map how apparently ordinary interactions can become risky when they intersect with eating-disorder symptoms, clinical vulnerabilities, and model affordances. That is a meaningful CHI move because it expands the unit of analysis from content to interactional and clinical context. The taxonomy is therefore novel as a synthesis and framework: it organizes risks such as generalized health advice, symptom concealment, thinspiration, and reinforcement of negative self-beliefs into a clinically informed structure that can support later design and evaluation work. At the same time, the paper is careful about its own limits. It does not claim to measure prevalence, severity, or model-to-model variation, and it does not test mitigations or benefits. So the strongest reading is that it offers a high-value conceptual map for a sensitive domain, with clear downstream utility for participatory evaluation and safeguard design, but not an operational safety solution by itself.

What Changed

Canon before

Prior CHI work on generative AI safety and eating-disorder-related harms would typically frame risk as overtly harmful content, policy violations, or generic misinformation. This paper shifts the lens toward clinically situated, expert-defined risk pathways that depend on user vulnerability and system affordances.

Departure from common sense

The paper shows that advice that seems acceptable for a general audience can still be harmful in an eating-disorder context because it can be inaccurate for that population and can reinforce rigid, restrictive thinking. The key move is that safety cannot be judged only by whether content is broadly correct or guideline-conforming.

Actual novelty

The paper’s novelty is an expert-guided taxonomy that organizes generative-AI risk around overlapping pathways linking user vulnerabilities, clinical features, and AI affordances, rather than treating harms as isolated bad outputs. It reframes the problem as a systems issue and surfaces mechanisms such as symptom concealment, thinspiration, and reinforcement of negative self-beliefs.

Evidence

Evidence supports a qualitative, expert-guided conceptual contribution: interviews with 15 domain experts and abductive analysis produced a seven-category taxonomy of risks. The paper explicitly notes that it did not systematically evaluate current systems, so the contribution is mapping and interpretation rather than benchmarked detection or intervention performance.

Limits

Method limits

The study relies on semi-structured interviews and abductive qualitative analysis with 15 experts, so the taxonomy reflects expert interpretation rather than direct measurement of model behavior or prevalence of harms. The paper does not provide a systematic evaluation of current generative AI systems against the taxonomy.

Deployment limits

The findings are not a ready-made detector or safeguard; they require translation into concrete assessment protocols, participatory review processes, and domain-specific mitigation design before deployment. The taxonomy is most useful as a conceptual lens for risk review rather than as an operational safety guarantee.

Boundary conditions

The claims are bounded to eating-disorder-related risks in generative AI and to expert-informed conceptual mapping. They do not establish how often the risks occur across models, populations, or settings, and they do not assess benefits of generative AI for early symptoms or recovery support.

Manual check: No manual check required from the provided evidence packet; the main caveat is that the evidence spans are coarse and the review should remain bounded to the qualitative claims explicitly supported in the paper text.

Position in field

This sits at the intersection of CHI safety, health-sensitive AI, and qualitative risk analysis. Its contribution is to move from generic content-safety framing toward clinically grounded, expert-defined risk pathways for a vulnerable population.

Abstract

Generative AI systems may pose serious risks to individuals vulnerable to eating disorders. Existing safeguards tend to overlook subtle but clinically significant cues, leaving many risks unaddressed. To better understand the nature of these risks, we conducted semi-structured interviews with 15 clinicians, researchers, and advocates with expertise in eating disorders. Using abductive qualitative analysis, we developed an expert-guided taxonomy of generative AI risks across seven categories: (1) providing generalized health advice; (2) encouraging disordered behaviors; (3) supporting symptom concealment; (4) creating thinspiration; (5) reinforcing negative self-beliefs; (6) promoting excessive focus on the body; and (7) perpetuating narrow views about eating disorders. Our results demonstrate how certain user interactions with generative AI systems intersect with clinical features of eating disorders in ways that may intensify risk. We discuss implications of our work, including approaches for risk assessment, safeguard design, and participatory evaluation practices with domain experts.