Chatbots Feed Diet Advice That Undermines Eating-Disorder Care

A small item in a Wall Street Journal column caught my eye this week, because it's the shape of a story consumer AI companies are going to keep getting hit with. Eating-disorder therapists told the Wall Street Journal that their patients are uploading photos of themselves to AI chatbots and asking for tips on how to lose weight or bulk up on muscle, then coming back to session with ultralow-calorie meal plans or excessive fitness routines the bot produced. Some patients list what they are eating and ask the chatbot how to eat better, and reportedly receive dangerous responses about cutting out too much fat and carbs.

The frustration in the piece is not that the advice exists. Diet and exercise pointers can help some users. It is that, delivered without any context of a diagnosis, those pointers land straight on top of an active eating disorder, and undoing them in therapy pulls valuable time away from actual treatment.

The precedent everyone in this space remembers is Tessa, the National Eating Disorders Association's chatbot, which was taken offline in 2023 after generative-AI capabilities added later caused it to recommend losing 1 to 2 pounds per week and eating no more than 2,000 calories a day. What is different now is that the harmful advice is not coming from a single specialty tool. It is coming from general-purpose consumer models people already use daily. Anthropic's Claude reportedly runs classifiers designed to recognize signs of disordered eating and back off when it detects them, which at least establishes that vendors can build for this.

The honest caveats matter. The reporting is anecdotal and therapist-sourced. It does not put a number on how widespread this behavior is or which bots come up most often in patients' phones, and it does not tell us whether classifier-style guardrails actually reduce harm versus plain refusals. Take the specifics as reported, not settled.

Still, the direction of travel is clear. Any consumer AI that will produce a calorie plan on request is one screenshot away from a clinical-harm story, and the labs that have already invested in eating-disorder specific safety are the ones with an answer when a regulator, insurer, or plaintiff comes asking.