Open Food Facts Flags Its Own Bad Data. Our Scanner Never Asks to See the Flag

On September 1st, Sibylle Kranz, a registered dietitian and nutrition researcher at the University of Virginia, published the results of hands-on testing she’d run across several AI diet-tracking apps, the same category Mr BITE sits in. She signed up for free trials, built different user profiles and goals, and logged real meals to compare what came back. Photo recognition was one problem: “I could give you dozens of examples where I took a photo of something I ate, and then I had to correct it because it misinterpreted what’s there,” she said. But she flagged something less obvious too: apps that lean on crowdsourced databases built from user-submitted entries, where the numbers can shift as more people add more, and sometimes wrong, data. The independent validation research she pointed to puts the typical per-meal calorie error for these tools at 250 to 345 kilocalories, in either direction, which is not a rounding error over a real week of logging.

I want to be fair about what this actually is: one dietitian’s hands-on test of a handful of apps, not a peer-reviewed trial, and she’s careful to say these tools still help people log more consistently than a paper diary or occasional dietitian visits ever did. But the crowdsourced-database point landed somewhere specific for me, because it isn’t a hypothetical about “AI apps” in general. It’s a description of exactly where a large share of Mr BITE’s own numbers come from.

Every barcode scan in our app resolves through Open Food Facts, an open, crowdsourced product database that anyone with an account can add to or edit, the same kind of database Kranz was describing. So I went and checked what our own lookup does with that. Our by-barcode fetch calls the full Open Food Facts product endpoint with no fields= filter at all, so we get back the entire product document Open Food Facts holds, not just macros. Buried in that same response, on every single product, are the fields Open Food Facts itself uses to flag trust: data_quality_tags and data_quality_errors_tags, which list specific problems editors or automated checks have found with an entry; states_tags, which says whether a product’s data has actually been marked complete and checked by a contributor, or just created and left half-filled; and unique_scans_n, a real count of how many shoppers have actually scanned this exact barcode, which Open Food Facts uses as its own default sort order when you search their database directly. Our own mapper, normalizeOffProduct, reads none of it. Calories, protein, carbs, fat, fiber, sugar, a name, a brand, an image. That is the entire list of fields we pull out of a response that already contains Open Food Facts’ own answer to “should you trust this entry.”

It’s worse for text search than for a scan. When you search for a food by name instead of scanning a barcode, our own ranking function scores every candidate purely on how closely its text matches your query, brand and product name string distance, nothing else. If two products share a name, one fully checked and complete, one a half-finished duplicate a contributor abandoned with a typo’d fat number, they rank identically as far as our code is concerned, because nothing in that ranking function has ever looked at quality, completeness, or how many other people have actually scanned it. The better-documented product doesn’t win. The better-matching string does.

I’m not going to pretend surfacing this is trivial product work. Deciding what a “this entry looks unfinished” flag should look like, whether to suppress incomplete entries outright or just warn on them, how to weight real scan counts against text relevance without breaking search for genuinely obscure but correct products, all of that is a real design job I don’t have finished. What I can tell you plainly is that the excuse “we don’t have that information” doesn’t apply here. We are already downloading it with every single lookup. We just never taught our own code to look at it.

Three Quick Bites

1. The Environmental Working Group’s analysis of 12 major fast-food chains, published this month, found at least one chemical from its “Dirty Dozen” food-chemical list on the menus of 11 of the 12 chains, with artificial dyes the single most common additive (on every menu except Chipotle’s and Subway’s) and sodium nitrite, a preservative linked to cancer, on more than half. Chipotle was the one chain that came back clean, built from a self-reported 53 total ingredients across its entire menu, which says less about how hard this is and more about how few other chains have bothered to try. EWG

2. A randomized clinical trial from Washington University School of Medicine, published August 27th in Cell Metabolism, found that a ketogenic diet cut liver fat by 67% in adults with obesity, prediabetes and fatty liver disease, more than either a Mediterranean or low-fat diet, even though all three groups lost the same 10% of body weight. Same weight loss, different metabolic outcome depending on what the calories were made of, which is exactly the kind of distinction a diary that only totals calories has no way to make. ScienceDaily

3. Abbott and Google announced a partnership on August 11th to feed real-time glucose data from Abbott’s over-the-counter Lingo sensor directly into Google’s Gemini-powered Health Coach, which will use it alongside activity, sleep and recovery data to shape its nutrition and coaching advice. It’s the clearest signal yet that the next round of AI nutrition coaching is going to be judged on what real physiological data it’s actually grounded in, not just how good its food-photo guess is. MobiHealthNews

Tool of the Week: EWG Food Scores

Since this issue is about how much a food’s trustworthiness depends on who actually checked it, this week’s pick is a free database built to check that for you. EWG’s Food Scores rates packaged foods on nutrition, ingredient concerns and degree of processing, with a plain letter-style grade and a full ingredient breakdown for each product, searchable by name or barcode. It won’t fix what our own barcode lookup currently ignores, but it’s a real second opinion sitting one search away. EWG Food Scores. Free, no signup required.

Thanks for reading. If you’d rather your coach’s database asked who actually checked its own data, Mr BITE is free on iOS and Android.

Collins, building Mr BITE