Why I keep deleting calorie tracking apps (and what AI finally fixes)

I have deleted MyFitnessPal four times. Lose It twice. Cronometer once. Each install lasted somewhere between three days and three weeks. Each time I told myself the same thing on the way out: maybe I just don't have the discipline.

I don't think that's true anymore. I think those apps are fine for a small slice of users — bodybuilders, dietitians, people who genuinely enjoy logging food. For the other 90% of us, calorie tracking is a tax. Thirty seconds per meal, three meals a day, every day, forever. Almost no one signed up for that life.

The numbers back this up. App-retention research consistently puts calorie-tracker churn somewhere between 70% and 80% within the first three months. The dropoff is sharpest in the first 30 days. You can blame discipline if you want, but if 75% of customers churn from any other product category, we'd call it a product problem, not a customer problem.

So what's the actual product problem?

The friction stack

When I trace back why I quit, it's never one thing. It's a pile of small frictions that add up to "not worth it":

• Search-and-pick. I type "chicken breast." I get 47 entries. Some are weighed in grams, some in ounces. Some are raw, some cooked. Some are user-submitted with the calorie count off by 30%. I stand there picking.
• Portion math. Was that 4 oz or 5? Was the rice 1 cup or 1.5? When I log a homemade dish, I'm doing third-grade arithmetic at the dinner table.
• Restaurant meals. Half the database entries for chain restaurants are wrong or stale. Independent restaurants don't exist in the database. I'm guessing.
• Anything homemade. My mom's lasagna is not in the database. Neither is the soup I made on Sunday.
• The 11-tap problem. I once counted: logging a single chicken sandwich in MyFitnessPal took me 11 taps and 38 seconds. That's not a "quick log."
• The streak break. Miss one Sunday brunch and the daily streak resets. Watch the motivation evaporate by Tuesday.

None of these are catastrophic on their own. The product hasn't shipped a hostile feature. It's just death by a thousand little asks. The 8 PM version of me — tired, eaten, sitting on the couch — is not going to open the app and reconstruct what I had for lunch six hours ago.

What AI actually fixes

Here's what's changed in the last two years that's worth getting excited about, and being specific:

• Photo recognition removes the search step. Snap a plate. The model identifies the items. You don't type "chicken breast" — the chicken breast types itself.
• Multi-item detection collapses entries. A photo of dinner is one log, not five. The mental overhead of "what counts as a meal" drops to zero.
• Vision-based portion estimation. This is the hard one. The model looks at your plate, infers size from context (fork, plate diameter, table edges), and estimates grams. It's not perfect — I'll get to that — but it's close enough that you stop doing math at the table.
• Voice and text input. "I had a turkey sandwich and an apple." Done. No search, no scrolling.
• Bulk catch-up. Forgot to log all day? Type the whole day in one paragraph. The model parses it. The 8 PM version of you can finally win.

The thing I built (Cali) is just a bet on this stack. No barcodes. No food database to scroll. Snap, talk, or type. The whole product is a refusal to ask the user for anything more than they want to give.

What AI doesn't fix

This is where I want to be honest, because the marketing in this category is getting wild.

AI doesn't make you want to lose weight. It doesn't fix the harder problem, which is that we live in a culture that engineers cheap, hyperpalatable food and then asks individuals to white-knuckle their way through it. A frictionless tracker can't substitute for an absent goal. It can only amplify a present one.

AI also can't see invisible things. The 15 g of olive oil coating your salad doesn't show up in any photo. The butter your restaurant cooked your fish in isn't on the plate. There's a perception ceiling, and any vendor claiming sub-10% accuracy on real meals is either overfitting to a benchmark or lying. (I ran the benchmark on 200 real meals — the honest number is around 29% median error, and I think that's the floor for this category until someone solves portion estimation in a way nobody has yet.)

And AI doesn't fix the dopamine machine. Streaks, badges, red numbers when you "go over" — these are the same gamification patterns that drove the 75% churn. Putting AI input on top of a habit-shaming UX gives you the same outcome with a faster log step.

What I actually think you should look for

Three rules, after building one of these and using all the others:

1. Default action: nothing. The best calorie tracker is the one you can ignore for two days and pick up again without penalty. No streak you can lose. No daily verdict.
2. Friction budget under 10 seconds. From "I want to log this meal" to "logged" should fit inside the time it takes to take a sip of water. If the app makes you confirm five things, the friction won.
3. Honest about accuracy. If the marketing says "97% accurate" without telling you accurate at what, accurate vs what ground truth, on which kinds of meals — assume marketing.

Whatever you pick, the test is day 47, not day 1. Day 1 every app feels great because the novelty is doing the work. Day 47 is when you find out whether the product respects your time.

I built Cali because day 47 is where I always quit. The whole goal is to get past it. Try it on the App Store — free, takes three seconds to log a meal, and the default action is doing nothing.