Therapists say AI Chatbots are feeding patients dangerous diet advice

This one’s personal. Repeat after me: Don’t use AI for nutrition, fitness, and wellbeing advice. 

Eating-disorder therapists are raising alarm over a pattern showing up in their sessions: patients uploading photos of themselves to AI chatbots and asking for weight-loss or muscle-building tips, and then returning to therapy with extreme, bot-generated meal plans (that fall way under the body’s nutritional needs) or excessive fitness routines (that overtax an already undernourished body) in hand, according to a paywalled Wall Street Journal report (accessed via AI Weekly).

The concern isn’t that diet or fitness advice exists online—it’s context. Delivered without awareness of a user’s diagnosis, medical history, nutritional needs, or vulnerability, seemingly reasonable advice can land directly on top of an active eating disorder. Patients may log what they eat and ask a chatbot how to “improve” their diet, only to receive guidance encouraging steep cuts to calories, fat, or carbohydrates. Clinicians, at least according to those in the WSJ report, say undoing the damage in session can pull time away from actual treatment, which is already under-covered and often doesn’t meet a patient’s needs.

This isn’t simply a matter of AI occasionally saying something obviously dangerous. A new research paper, “From Symptoms to Systems: An Expert-Guided Approach to Understanding Risks of Generative AI for Eating Disorders,” identified seven overlapping ways generative AI can create or amplify disordered eating risk, including giving apparently benign diet advice that becomes harmful in context, facilitating disordered behaviors, helping users conceal symptoms, generating “thinspiration,” reinforcing negative self-beliefs and social comparison, encouraging obsessive attention to the body, and reproducing stereotypes about who gets eating disorders. The researchers’ larger warning is that AI safety can’t be measured simply by asking whether a chatbot refuses an obviously dangerous prompt — which they actually do. The risk can emerge gradually across a conversation, as a user’s vulnerabilities interact with AI’s ability to personalize, validate, and endlessly engage.

That distinction matters, as the problem isn’t that a chatbot can calculate calories or design a workout for leg day but that it doesn’t reliably know when those capabilities have become harmful. A request for a “healthy” calorie deficit can look identical on the surface whether it comes from someone casually preparing for a 5K or someone with an active eating disorder who is already severely undernourished. The paper therefore argues for context-sensitive, multi-turn safety evaluation developed with eating-disorder experts—and cautions that guardrails also need to avoid blocking legitimate recovery-oriented conversations.

The situation echoes a cautionary precedent from 2023, when the National Eating Disorders Association pulled its Tessa chatbot after newly added generative-AI capabilities led it to recommend the outdated nutrition advice of losing one to two pounds a week and capping daily intake at 2,000 calories, approximately what the body needs just to keep vital organs functioning. The advice is particularly inappropriate for people with active eating disorders, who may require substantially more nutrition to restore physiological stability. But the current wave looks different: rather than a single specialty tool malfunctioning, potentially harmful advice is now coming from general-purpose consumer chatbots that people already use every day.

Notably, Anthropic says its Claude models run classifiers built to recognize signs of disordered eating and pull back from responses that could reinforce harmful patterns—evidence, AI Weekly notes, that guardrails against this kind of harm are technically achievable. But the research suggests that “refusing the dangerous prompt” isn’t enough to minimize risk. Companies need to account for the quieter ways an AI can participate in an eating disorder: validating a user’s increasingly restrictive rules, helping conceal behaviors, intensifying body comparison, or turning a seemingly innocuous wellness conversation into an iterative optimization exercise.

AI isn’t a dietitian, personal trainer, therapist, or eating-disorder clinician, and increasingly sophisticated chatbots don’t change that. If anything, their ability to sound empathetic, confident, personalized, and authoritative makes knowing when not to use them even more important.

This piece touches on eating disorders and diet-related harm. If you’re personally struggling with disordered eating, the National Alliance for Eating Disorders helpline is a good place to start for more information.