The Promise and Peril of AI in Eating Disorder Support
Eating disorders kill more people per illness rate than any other mental health condition. Someone dies from an eating disorder–related cause roughly every 52 minutes in the United States, and yet twenty to twenty-five percent of those affected never receive professional care at all (Deloitte Access Economics, 2020; Solmi et al., 2024). That's a gap most of us don't think about until we're standing in it. And here's the thing: artificial intelligence could, in theory, help close it.
AI can't replace specialized human care for eating disorders — these are complex, clinically serious conditions that demand nuanced, individualized treatment. But early research suggests these tools may offer short-term support by helping slow symptom escalation while individuals wait for treatment. In 2025, a rigorous clinical trial tested an AI chatbot called Therabot with adults at high risk for eating disorders. Results showed large reductions in eating disorder–related symptoms compared to waitlisted controls, and participants rated the experience as comparable to working with a human therapist (Heinz et al., 2025). A separate randomized controlled trial of Brazilian adolescents aged 13 to 18 found that a brief 72-hour chatbot intervention improved body satisfaction, with the greatest benefits among those who reported higher baseline body-related distress (Matheson et al., 2023).
AI is also showing significant promise as a clinical tool for identifying and predicting eating disorders. A review of 75 studies found that machine learning models could predict eating disorder onset, support diagnosis, and forecast treatment outcomes with meaningful accuracy — including detecting binge-eating episodes approximately 82 percent of the time (McClure et al., 2025). Emerging approaches are improving model transparency, allowing clinicians to understand not just what a system has flagged but the underlying factors driving those predictions (Brizzi et al., 2025). A clinical trial in Australia found that a single-session chatbot intervention was beneficial for adolescents and adults on waiting lists, reducing both eating disorder symptoms and psychosocial impairment (Sharp et al., 2025). Participants who engaged with the therapeutic chatbot were also more likely to enter treatment once it became available.
So the promise is real. It's just not complete.
When AI Gives Dangerous Diet Advice
Despite these promising developments, significant concerns remain about AI's ability to safely support individuals at risk for eating disorders. And the failures aren't theoretical.
In one study, UK researchers created ten fictitious adolescent personas ages 10 to 15 representing different gender identities and weight statuses, then engaged ChatGPT and Claude in conversations about eating, weight, and appearance concerns. Most chatbot responses treated potential eating disorder symptoms as general health or lifestyle worries rather than recognizing possible clinical risk. Responses were inconsistent, occasionally reinforced harmful weight-related ideals, and rarely directed vulnerable teens toward appropriate mental health resources (Sheen et al., 2025).
Other studies have demonstrated even more alarming failures. Posing as a 13-year-old girl, researchers at the Center for Countering Digital Hate found that ChatGPT generated an extreme calorie-restricting plan with as few as zero calories per day and advised the user to conceal these restrictive eating behaviors from family members (Ahmed, 2025). A study by Bilen and colleagues (2026) found that five leading AI chatbots — including ChatGPT-4o, Gemini 2.5 Pro, Claude 4.1, Bing Chat-5GPT, and Perplexity — generated meal plans for adolescents that underestimated daily energy needs by an average of 695 calories compared with plans developed by a registered dietitian.
The risks extend beyond dietary recommendations. A 2025 study (Choi et al., 2025) found that while individuals with eating disorders often reported feeling empowered by interactions with an LLM-based chatbot, many failed to recognize harmful or problematic responses because they placed a high degree of trust in the chatbot. That trust is precisely what makes this dangerous — these tools feel supportive, which means users are less likely to question advice that should raise red flags.
The Tessa Precedent: A Defining Cautionary Tale
Perhaps the most well-known cautionary example is Tessa, the rule-based chatbot deployed by the National Eating Disorders Association (NEDA) (Hoover, 2023). Intended to provide support and recovery resources, the chatbot instead began dispensing weight-loss and calorie-restriction advice to users seeking help for eating disorders. The tool was suspended in 2023, and the incident has since become a defining example of AI-related harm in the eating disorders field.
Tessa matters because it wasn't a fringe failure. It was deployed by one of the most established organizations in the field, with clear clinical intent and oversight. And it still went wrong — badly. The incident underscores the risks of deploying automated support tools without adequate safeguards, even when the developers' intentions are sound. It's a reminder that good intentions don't override system design flaws, and that the stakes in this domain are higher than almost anywhere else in mental health technology.
For context, you might also want to read our earlier piece on the broader limits of AI in mental health, where we explore why the same patterns of overconfidence and underperformance show up across diagnostic domains.
The Capability Gap: Why AI Misses Early Warning Signs
Research is beginning to point to a specific weakness in AI's ability to identify eating disorders: it often misses the early warning signs.
A 2026 study from mpathic AI (Douglas et al., 2026) tested six leading AI models using realistic conversations about eating, weight, and body image. The conversations ranged from no eating disorder risk to severe risk, and licensed clinicians reviewed every exchange to evaluate how well the chatbots responded. The results were striking.
Most models avoided giving obviously harmful advice when someone was clearly in crisis. But they struggled to recognize the kinds of subtle warning signs that often appear first — comments about "eating healthier," exercising more, losing weight, or being more disciplined with food. In many cases, the AI treated these statements as ordinary wellness goals rather than possible signs of an emerging eating disorder.
The models also had difficulty recognizing when risk increased over the course of a conversation. As a person's concerns became more serious, the chatbots often responded to each message as if it were a stand-alone question rather than noticing the larger pattern developing over time.
This matters because eating disorders rarely begin with obvious cries for help. They often start with behaviors that look healthy on the surface: careful eating, strict exercise routines, or an intense focus on nutrition. One of the most important skills clinicians develop is recognizing when those seemingly healthy habits are masking something more serious. Current AI systems, according to this research, are not yet consistently able to make that distinction.
For readers interested in the clinical side of this challenge, our article on why human presence remains non-negotiable in trauma therapy explores similar themes around the limits of algorithmic assessment.
What This Means for Patients, Families, and Clinicians
At their best, AI tools for eating disorders offer a low-barrier, judgment-free first step toward support for people who might otherwise receive none. At their worst, they hand a struggling teenager a starvation plan and tell them to keep it secret.
The concern isn't limited to eating disorders. A G7 report released this year, co-authored by Children and Screens and endorsed by researchers and institutions around the world, examined generative AI's broader impacts on child development. Its message was similar: the opportunities are real, and so are the risks, and both deserve serious attention from researchers, developers, and policymakers alike.
What emerges from the research is not a simple verdict of "good" or "bad." AI is becoming increasingly capable of identifying risk, expanding access to information, and supporting clinical care. At the same time, it remains prone to errors that can have serious consequences when the topic is food, weight, body image, and mental health.
For families, clinicians, and young people, the lesson is clear: AI may be a useful tool, but it is not yet a trusted guide. The question is no longer whether young people will rely on AI for advice about food, weight, and body image. Many already are. The question is whether the systems answering them are prepared for the responsibility. The evidence so far suggests that, while they are improving, they are not there yet.
Until the capability gap closes — and we have reason to believe it will, gradually — the most responsible use of AI in this space is as an adjunct to, not a substitute for, specialized human care. Clinicians should be aware of these tools, families should discuss them openly with young people, and developers need to treat safety guardrails not as an afterthought but as a clinical imperative.