My GI dietitian wanted a record of everything I ate and how I felt afterward. Stopping to type into a spreadsheet at mealtimes never worked for me. On February 13 I took that problem to ChatGPT and talked it through for about three hours.
That became the idea: an app that would flag potentially irritating foods before you ate them, at the grocery store or on your plate. It would rate FODMAPs, short-chain carbohydrates that gut bacteria ferment. In some people that brings gas, bloating and pain. It would keep each rating with the meal and let you log symptoms afterward. Over time it could look for patterns between the two.
A few days later we started coding with ChatGPT at 9 a.m. The layouts were in place by noon and the app worked by dinner. ChatGPT kept calling my ideas brilliant and truly unique. I felt invincible.
Then came the food. In an early version an AI read each meal and gave it a FODMAP rating. On my first day of logging it rated a snack of almonds and cashews low. A latte with regular milk and tortillas filled with caramelized onions came back low too. Every rating arrived in the same calm voice. Nothing on the screen said where it came from.
Cashews, milk and onions are all well-known FODMAP sources. Whether they bring on symptoms depends on the portion and the person.
Making it work right has taken months of testing ratings and fixing how the app matched foods to data. It has also meant learning how each engine behind the app reaches a decision.
How a meal earns a rating
By the summer the AI no longer rated meals by default. Today its default job is to turn your words or your photo into a list of ingredients. The app breaks a dish into its parts and looks up each one. It checks our curated food table first. Then it checks a much larger set of records from national food databases and a published FODMAP dataset.
Before the app settles on a low or moderate rating, a second and more cautious check looks again. Your own corrections override the app's rating. When the data has nothing, the app says so instead of rating the meal. Asking the AI to guess is off by default. Its guesses carry a GUESS badge.
What measurement taught us
Ordinary wording kept changing the answer.
- Plain "bread" came back low. The national food databases we started from don't measure fructans, a FODMAP found in wheat.
- "Sourdough toast" as one phrase came back low. Split into "sourdough" and "toast" the same slice came back high.
- "Gluten-free pretzels" and "gluten free pretzels" got different answers. The only difference was a hyphen.
- In August the app rated a cheese pizza by its cheese and missed the crust. Most of the dish names we tested came back low.
Takeaways
- Confidence isn't evidence. The AI sounded the same on every rating, so its tone couldn't tell you which ones to trust.
- Every rating should show where it came from. One you can't trace is an answer you can't check.
The next time an app rates your meal, check three things: the food it matched, the serving size and where the rating came from.