A New Study Found GLP-1 Users Report Nail Detachment at Four Times the Rate of Non-Users. Mr Bite's Safety Filter Blocks the Word 'Medication' Before It Reads the Rest of the Sentence.

A multinational survey presented at the EADV Congress in Vienna, which ran September 30th through October 3rd, asked 575 people on a GLP-1 drug and 1,329 people who weren’t the same set of questions about their nails and hair. The gaps were large. Nail detachment: 39 percent of users, 9 percent of controls. Nail discoloration: 46 percent versus 17. Unusual hair loss: 50 percent versus 34. Reduced hair volume or density: 42 percent versus 19. Adjusted analyses narrowed some of that, down to roughly two to three times as common rather than four, and the lead dermatologist, Bruno Halioua, floated inflammation rather than rapid weight loss as the likely mechanism, since hair loss in the study usually came with scalp itching, redness, or scaling. The honest caveats matter as much as the numbers: this is self-reported, cross-sectional, not yet peer-reviewed, and about half the users who reported nail problems said they’d had them before they ever started the drug. Nobody involved is claiming GLP-1 drugs definitely cause nails to fall off. What the study actually shows is that a real share of people on these drugs notice something happening to their nails and hair, and reasonably want to know whether it’s connected.

Mr Bite is a coaching app, and a meaningful number of the people using it are on a GLP-1 drug, whether we ask about it or not. So I wanted to know what would actually happen if one of them typed this week’s finding into our own chat, the way a person genuinely would: “my nails have been coming off since I started my medication, is that normal?”

The first thing I found wasn’t in the chat at all. It’s in onboarding, in src/utils/mrBiteHealthSafety.ts, which builds a HealthSafetyProfile from a short multi-select question. One of the six options, defined in onboardingChatQuestions.ts, is literally labeled “Medical or medication,” stored internally as the single reason medical_condition_or_medication. Check that box for any reason at all, a thyroid condition, a broken wrist, or being on semaglutide, and modeForHealthSafetyReasons() puts you in the same bucket: clinician_required. That mode isn’t cosmetic. healthSafetySummaryForPrompt() tells the coaching model, in plain instructions, not to calculate or save calorie or macro targets for you at all, full stop, and to point you at a clinician instead. A GLP-1 user who wants help adjusting their protein targets around a appetite-suppressing drug gets exactly the same blanket refusal as someone with an unrelated diagnosis who never mentioned a drug. The checkbox can’t tell the two apart, because it was never built to.

The second finding is in the chat layer itself, in supabase/functions/_shared/mrBiteSafety.ts. Every message gets run through detectMrBiteSafetyConcern(), and one of its patterns is MEDICAL_HARD_BLOCK_RE, a single regular expression that matches the literal words “medication,” “dose,” “dosage,” “prescription,” “diagnose,” “treatment,” and a handful of others. There’s no sentence-level understanding here, no distinction between “what’s the right dosage for me” (a real request the app is right to refuse) and “could my medication be behind this nail thing” (a different question entirely, one the person asking already half-answered for themselves). Both hit the same regex, both get routed to the same category, medical, and both get the same canned reply: “General nutrition info only, medical nutrition belongs with a clinician.” I typed a version of this week’s exact study finding into a test account to confirm it. It got blocked on the word “medication” before the model ever saw the word “nails.”

I don’t think the fix here is to loosen the filter. A nutrition-coaching chatbot has no business freelancing an opinion on whether a drug is causing a side effect, and refusing that conversation is the right call, not a bug. What’s missing isn’t caution, it’s any acknowledgment that the caution is happening. The deflection message is identical whether you asked something dangerous or something reasonable, so the person who typed a real, current, study-backed concern gets exactly the same brush-off as someone fishing for a dosage shortcut. A coach that’s actually built around GLP-1 users, and plenty of apps in this space now are, would at minimum recognize the specific shape of this question and say so, something like: that sounds like it could be related to your medication, which is outside what I can speak to, please check with your prescriber. Mr Bite doesn’t do that. It doesn’t track which medication, if any, a user is on anywhere in the data model, so there’s nothing for a smarter response to even draw on. Right now, the entire category of “on a GLP-1 drug” collapses into one onboarding checkbox and one keyword match, and neither one knows the difference between a dosage question and the exact concern 39 percent of users in this week’s study were trying to raise.

Three Quick Bites

1. Frutas y Hortalizas del Sur’s frozen berry recall expanded again on October 2nd to cover more lots of Great Value Organic Triple Berry Blend and Organic Blueberries sold at Walmart, plus a new lot of Trader Joe’s Organic Mixed Berry Blend, over possible E. coli O145 contamination. The underlying outbreak, 17 people sick and 6 hospitalized across the summer, was already declared over by CDC and FDA before this expansion landed, a reminder that a recall’s timeline and an outbreak’s timeline don’t always move together. U.S. News & World Report

2. A Kyoto University team reported in Diabetes on October 8th that giving mice olive oil 15 minutes before glucose measurably blunted the early blood sugar spike, even in mice with damaged insulin-producing cells. The researchers traced it to delayed stomach emptying plus earlier secretion of glucagon, GIP, and GLP-1, the same hormone family the drugs in today’s main story work through. It’s mouse data, not a dietary recommendation for people, but it’s a clean mechanistic look at why meal order keeps showing up in human studies too. ScienceDaily

3. California’s SB 869, signed September 28th alongside the state’s new Non-UPF label, requires chain restaurants with 20 or more locations to flag any drink with 50 grams or more of added sugar with a sugar-cube-in-a-triangle icon on the menu, starting January 1st, 2029. It doesn’t restrict what anyone can order, it just forces a large chain to say so on the page before you do. Governor’s Press Office

Tool of the Week: FDA’s FAERS Public Dashboard

Since this issue is about a side effect nobody in our chat is equipped to weigh in on, the pick this week is the free, public place to actually check one. The FDA’s Adverse Event Reporting System (FAERS) Public Dashboard lets anyone search real reported side effects for a specific drug, by name, age group, and time period, no account or fee required. It won’t tell you whether your own nails falling off is related to your prescription, that’s still a conversation for your prescriber, but it will show you, in minutes, whether “nail disorder” or “alopecia” shows up against a GLP-1 drug’s report history at a rate worth asking about. It’s exactly the kind of lookup a chatbot regex can’t safely replace, and exactly the kind the agency already built and gives away for free. FDA FAERS Public Dashboard

Thanks for reading. If you want a coach that’s honest about the question it’s built to deflect, Mr BITE is free on iOS and Android.

Collins, building Mr BITE