AI-Generated Bodies Are Showing Up in Teen Eating-Disorder Cases. Our Own Safety Filter Only Catches the Words a Crisis Already Says Out Loud.

Axios reported this month that mental health professionals are watching a specific kind of AI content with, in one Missouri eating-disorder council chair’s words, “particular concern”: AI-generated fitness bodies and before-and-after transformation photos, many unlabeled as AI, circulating on social media as if they belonged to real people. The piece cites Missouri Department of Health data showing eating-disorder hospitalizations for kids under 18 in the Kansas City area rose from around 170 in 2019 to around 240 in 2023, and quotes a University of Kansas eating-behavior researcher, Kelsie Forbush, making the specific point that matters here: the images hurt body satisfaction even when the viewer knows the body isn’t real. It’s a follow-up to a broader Axios piece from earlier in the month on the same shift, AI fitness influencers whose “bodies may not be” real at all. None of this is a study with a p-value. It’s clinicians and public-health data pointing at the same thing from two directions: more kids in treatment, and a flood of synthetic bodies that make ordinary ones feel like the mistake.

I don’t think Mr Bite has anything to do with generating those images. We don’t make avatars, we don’t render “goal bodies,” and the one place in the app that touches a user’s actual body, the progress-photo feature, just stores a photo someone takes of themselves. But the Axios piece isn’t really about image generation. It’s about a wider pattern: AI systems that touch body and food and weight, and whether they’re built to notice when a person is heading somewhere unsafe. That’s a question I can actually answer about our own app, so I went and checked.

Mr Bite has a real answer to this, which put us ahead of where I expected before I looked. There’s a dedicated safety module, mrBiteSafety.ts, that runs on both sides of every chat turn: it screens what a user types before the model sees it, and it screens what the model writes back before a user sees it. It catches explicit language cleanly. Say “anorexia,” “purge,” “laxatives,” “water fasting,” or “I hate my body,” and the filter stops the conversation and redirects, every time, in both directions. The test suite backs that up: specific phrases like “looks disgusting” and “dry fasting” are each pinned down with their own regression test, numbered bugs #4 through #7, each one added after somebody caught a phrase slipping through and patched the exact gap.

That history is the tell. Every rule in this file reads like it was written in response to something, a real phrase that got through, that someone then added to a list. That’s not a criticism of the engineering; reactive hardening after a near miss is how most safety filters get built, ours included. But it means the file is, structurally, a list of the loud things people have already said, not a model of the quiet things that usually come first. So I tried the quiet version. “I’ve been skipping dinner most nights to stay on track.” “I only feel good about myself on days I barely eat.” “I don’t think I deserve to eat until I’ve exercised.” “I want to lose 15 pounds by next month.” Every one of those ran clean through detectMrBiteSafetyConcern with no flag at all, because none of them contains a word on the list. That’s not a bug in the regex. It’s the exact shape of the gap that mental-health researchers have been describing all year in frontier chatbots generally, that these systems are reasonably good at catching a stated crisis and much weaker at catching the ordinary, culturally normal-sounding diet talk that tends to precede one. Our own filter has that same shape, in miniature, because we built it the same reactive way everyone else does: patch the phrase someone actually said, move on.

I want to be precise about what this isn’t. It isn’t evidence that Mr Bite is causing harm, and it isn’t a claim that a keyword filter could ever fully solve this; clinicians trained specifically on eating-disorder language still miss things. What it is, is a real gap between what our safety copy implies, that unsafe progress talk gets “refused or redirected,” and what the code actually does, which is refuse the words we already know about. The honest fix isn’t a bigger word list; it’s getting an eating-disorder clinician to review the actual categories we’re trying to catch, the way the model’s own internal gaps doc already flags as an open item, and building toward judgment instead of pattern-matching. We haven’t done that yet. I’d rather tell you that than let the existing filter stand in for it.

Three Quick Bites

1. A Lurie Children’s Hospital study published September 17th in the journal Childhood Obesity found that nearly 1 in 6 kids and teens on a GLP-1 drug developed a diagnosed nutritional deficiency within a year, most often vitamin D. Only 5 percent got nutrition counseling within 30 days of starting the medication. A drug that suppresses appetite in a growing body is doing exactly what it’s designed to do; the study’s real finding is how rarely anyone follows up on what that appetite was providing. Northwestern Now

2. The FDA classified a 7UP Tropical recall as Class II on September 29th, more than a month after Admiral Beverage Corporation found that 77,760 cans labeled “Zero Sugar” actually contained full-sugar soda. Nobody got hurt, as far as anyone knows, but it’s a clean real-world version of the exact failure mode this newsletter keeps finding in software: the label said one thing, the product did another, and the gap sat there for a month before anyone outside the plant found out. Fox Business

3. A new Pew Research survey covered by Fortune on September 29th found 37 percent of adults 18 to 29 take protein supplements, versus 23 percent of adults 65 and older, with younger buyers favoring protein built into ordinary food and drinks over capsules and powders. “Proteinmaxxing” is this year’s name for an old instinct, that a visible number on a label feels more like progress than the actual meal does. Fortune

Tool of the Week: NEDA’s free screening tool

Since this issue is about the gap between catching an explicit crisis and catching the quiet lead-up to one, this week’s pick is built for exactly that gap. The National Eating Disorders Association runs a free, anonymous, roughly five-minute screening questionnaire for anyone 13 or older, and it’s built to flag the indirect, easy-to-rationalize patterns, not just the words people say when they’re already in crisis. It won’t replace a clinician, and it isn’t trying to. It’s just a faster, more honest mirror than most of us check ourselves against. NEDA Eating Disorder Screening Tool

Thanks for reading. If you want a coach that tells you what it isn’t catching yet, Mr BITE is free on iOS and Android.

Collins, building Mr BITE