How AI Advice Is Undermining Eating-Disorder Therapy
Eating-disorder therapists are getting fed up when patients turn to chatbots for advice.
Even a chatbot trained on nutrition and fitness research that dispenses reasonable-sounding guidance can become a deadly influence
- Eating-disorder therapists say patients are increasingly using chatbots for diet and exercise advice, sometimes challenging professional guidance.
- Therapists warn that chatbot advice delivered without clinical context can be dangerous and delay vital treatment for eating disorders.
- AI companies say they are working to improve how their models recognize and respond to signs of disordered eating.
Women and men are uploading photos of themselves to artificial-intelligence bots, asking for tips on how to lose weight or bulk up on muscle. Some list what they’re eating and ask how they could eat better, sometimes getting dangerous responses about cutting out too much fat and carbs. Others prompt their chatbots to produce ultralow-calorie meal plans or excessive fitness routines.
While advice on diet and exercise could be helpful to some users, it can be harmful to someone with an eating disorder. Trying to countermand the chatbot’s advice takes valuable time and effort away from treatment, therapists say.
It’s getting so bad that some say patients are even challenging them during therapy sessions.
“I’ve had patients hold up their phone and use their mic to ask ChatGPT what was wrong with what I just said,” says Dr. Anne O’Melia, chief clinical and quality officer of the Eating Recovery Center, which provides treatment in several states. “Half my patients are fact-checking me daily.”
While chatbots are the new Dr. Google for a lot of people seeking medical advice, eating-disorder therapists say advice offered to their patients without the context of their diagnoses is downright dangerous. Eating disorders have the second-highest mortality rate of any psychiatric illness, behind opioid addiction.
Chatbot advice also differs from that found on social media, which has already been linked to the development of eating disorders among teens. “This is one-on-one engagement,” O’Melia says. “Chatbots learn what you like to hear and are programmed to agree with you.”
With eating disorders, patients often believe that extreme exercise or bingeing and purging are helping them cope with problems such as anxiety or depression, says Allie Weiser, director of helpline services for the National Alliance for Eating Disorders.
Therapists are now spending a lot of time trying to convince their patients that their guidance is based on clinical evidence, while chatbots might synthesize advice from general nutritional or fitness sources intended for those without eating disorders, or pull information from unreliable sources.
Making better models
AI companies say they’re working to improve the way their chatbots recognize and respond to signs of eating disorders, and bots now refer patients to professionals more often.
OpenAI continues to use input from clinicians and real use cases to refine its eating-disorder policies and safeguards, a company spokeswoman says. OpenAI pointed to independent research, which found that GPT-5.4, the model that previously powered ChatGPT, consistently recognized eating disorders and recommended evidence-based treatment.
An Anthropic spokesman says its Claude chatbot models have classifiers designed to recognize signs of disordered eating. When that happens, he says, Claude avoids providing information that could encourage or reinforce harmful patterns (calorie counts, weight-loss guidance, etc.). Earlier this year Anthropic introduced a new evaluation dedicated to disordered eating. Its latest models, Opus 4.8 and Fable 5, scored 99.8% and 99.7%, respectively, on identifying potential risks in prompts.
A Google spokeswoman says Gemini encourages people to consult with medical professionals. “We’re continuously working to improve the safety and accuracy of our models,” the spokeswoman said in a statement.
Correcting chatbot misinformation can delay vital treatment, says Hannah Lindsey, a therapist in Louisville, Ky. “When a patient says a chatbot told her to go to the gym five times a week and all she eats is oatmeal,” she says, “I have to put the trauma processing to the side and address the meal plan that’s not appropriate.”
But training AI to know when and how to offer advice to those struggling with disordered eating has proven difficult. When leading experts tried to offer their own chatbot to help those at risk, it dispensed problematic advice. In 2023, a bot on the website for the National Eating Disorders Association offered unapproved weight-loss tips. It turned out the bot was programmed with generative AI without the nonprofit’s knowledge. It was taken offline. Chief Executive Jessica Scheer says NEDA has no plans to bring back a chatbot.
Downsides and upsides
Even when therapists break through to their patients, the patients often fall back to the always-available chatbot once they’re outside the doctor’s office. Therapists acknowledge that chatbots can provide support for people when a therapist isn’t necessarily available, such as between visits. People who don’t see therapists also now turn to chatbots.
“We know that upward of 70% of people with eating disorders will never get access to care,” says Johanna Kandel, CEO of the National Alliance for Eating Disorders. “The positive part of AI is that it’s pushing a lot of people to call our helpline directly”—one staffed by licensed human therapists.
Last month, 7.6% of calls to the helpline originated from ChatGPT and other AI platforms, up from less than half a percent a year ago.
Kandel says the uptick in referrals suggests chatbots are getting better at detecting signs of disordered eating. But such disorders are complex, and people don’t usually disclose them to their chatbot. When an AI model is dispensing advice, it doesn’t know the difference between a body builder who wants to lift heavier weights and is consuming enough protein and a malnourished person who isn’t.
When an AI model is dispensing advice, it doesn’t know the difference between a body builder who wants to lift heavier weights and is consuming enough protein and a malnourished person who isn’t.
“I really feel like AI platforms have an opportunity here to learn from what happened with social media and eating disorders, and to do better,” Kandel says.