What Counts as Harmful Eating Disorder Content? A Delphi Study Developing Consensus-Based Harm Identification Rules with Lived Experience and Professional Experts
This study utilized a modified Delphi methodology with 41 experts (combining professional and lived experience perspectives) to develop 45 consensus-based rules for identifying both explicit and context-dependent harmful eating disorder content on social media, aiming to support automated detection and personalized moderation while preserving beneficial material.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Social media has become a constant companion for many people, shaping how they see themselves and interact with the world. For those struggling with or at risk of developing an eating disorder, these digital spaces can be particularly powerful, sometimes reinforcing harmful thoughts about body image, food, and weight. Algorithms, the computer programs that decide what content to show users, often learn from a person's clicks and viewing time. If someone lingers on a post about dieting or extreme thinness, the system may show them more of the same, creating a cycle where dangerous ideas feel normal and unavoidable. While social media companies have tried to stop the most obvious harmful content, such as posts that openly encourage starvation, they struggle with the "grey areas." A video about what someone ate for lunch, for instance, might look harmless on the surface but could secretly promote restrictive eating habits or trigger someone to compare their own body. The challenge for platforms is knowing exactly where to draw the line without accidentally removing helpful recovery stories or normal health information.
To solve this puzzle, a team of researchers brought together two distinct groups of people to build a new set of rules for identifying harmful content. They did not rely on computers or guesswork alone; instead, they used a structured process called a Delphi study, which is a method for building agreement among experts through repeated rounds of feedback. The team gathered twenty-one professionals, including doctors, researchers, and clinicians who work in the field of eating disorders, and twenty individuals who had personally experienced eating disorders in the past but were no longer currently diagnosed. By combining professional knowledge with the real-world insights of those who have lived through the illness, the researchers aimed to create a clear, shared definition of what makes online content dangerous.
The process began with the researchers reviewing existing studies and conducting interviews to draft a long list of potential rules. They started with sixty-seven different ideas about what might be harmful, ranging from posts that show visible ribs to encourage thinness, to videos that promote specific weight-loss drugs without medical advice. These initial ideas were then tested and refined through a series of surveys. First, a small steering committee reviewed the rules, and then the larger group of forty experts voted on them. In each round, the experts rated whether they agreed a rule should be kept, and they shared their thoughts on why a rule might be too vague or too strict. The researchers used these comments to rewrite the rules, making them more precise, before asking the group to vote again. This cycle happened several times until the group reached a strong agreement on a final set of guidelines.
The result was a finalized list of forty-five rules that can be used to spot harmful eating disorder content. The experts found that some types of content were easy to agree on as harmful. Rules that targeted clear, dangerous behaviors, such as promoting extreme fasting, showing visible bones to suggest thinness is desirable, or selling unregulated weight-loss pills, received near-universal agreement. These are the types of posts that are objectively unsafe. However, the group found it much harder to agree on content that depends heavily on context. For example, a post about exercise or nutrition is not inherently bad; it only becomes harmful if it frames these activities as a way to punish the body or if it suggests that a person's worth depends on their physical appearance. The experts noted that distinguishing between a healthy recovery story and a triggering post often requires understanding the intent of the creator and the vulnerability of the viewer.
Ultimately, the study concluded that there is no single, simple filter that can catch every harmful post without also catching helpful ones. The final rules show that harm is not just about what is said or shown, but about how it is framed and who is seeing it. While the experts could easily identify posts that glorify starvation or promote unsafe medical shortcuts, they struggled to create rigid rules for more ambiguous topics like body measurements or recovery narratives. The researchers suggest that these findings provide a foundation for better automated detection systems, but they also highlight the need for human judgment and context-sensitive moderation. By understanding exactly what constitutes harm through the eyes of both experts and those with lived experience, social media platforms can move toward protecting vulnerable users while preserving the beneficial content that supports recovery.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.