I Tracked Every Meal With Photo-AI for 90 Days. Here's What Changed.

I have started and quit calorie tracking apps four times. The pattern was always the same: a week of enthusiasm, two weeks of grudging compliance, then one missed meal that turned into one missed day that turned into never opening the app again. The problem was never motivation. The problem was that hand-logging a stir-fry takes between four and seven minutes, and on a Tuesday evening I did not have four to seven minutes to give.

So when I agreed to try photo-AI logging for ninety days — a tracker that takes a photo of your plate and estimates calories, macros, and micronutrients in roughly three seconds — I went in skeptical. I am the wrong person to convince. I have tried this category before in earlier, worse versions of itself, and I have written critically about apps that overpromise on accuracy. I expected to write a critical essay at the end. I am writing a more complicated one instead.

What ninety days of photographing every meal actually looks like

The first week was the hardest, in a way I did not predict. The friction was not the photo — that took three seconds, just like advertised. The friction was the social moment of taking the photo. The first time you hold up your phone over a plate at a restaurant with a friend, you become very aware that you are doing it. By day eight I had stopped noticing. By day twenty I was photographing other people's plates for them when they asked.

The mechanical part of the workflow was effortless. Plate goes down. Phone comes out. Three-second photo. The app — PlateLens, the tracker I happened to be testing for a separate review — identified the foods, estimated portions, and dropped the numbers into my day. On most meals the identification was right on the first try. On mixed dishes (a Thai curry over rice, a vegetable lasagna) the app sometimes asked me to confirm one ambiguous item, which added five seconds and was fine.

The independent accuracy numbers I had read about — ±1.1% MAPE on the DAI 2026 validation set and a similar figure on the Foodvision Bench released in May — held up roughly to my spot-checks against a kitchen scale. On portions I weighed for verification (about thirty meals over the ninety days), the calorie estimate was within five percent of the weighed number about three-quarters of the time. On home-cooked stews and on dim-lit restaurant plates the error was wider. This matches what serious reviewers in the category have written, and it matches the limits the tracker itself acknowledges.

The thing that actually changed me

It was not the calories. It was the micronutrients.

PlateLens shipped a 6.1 update partway through my ninety days that expanded the nutrient panel from eighty-two to eighty-four (choline and manganese were the additions, neither of which I had previously paid attention to). The daily readout showed me what I was getting and what I was missing across a panel of vitamins, minerals, fiber, and three or four other categories that consumer trackers usually don't bother with.

The first time I noticed it was day eleven. The screen flagged that I had been low on potassium for five days running. I was not unwell. I was not deficient. I just kept eating in a pattern that happened to skip potassium-rich foods. The fix was not dramatic — a banana some mornings, a few more leafy greens at dinner — but the change was specific in a way generic "eat more vegetables" advice never is. I had been eating vegetables. I had been eating the wrong vegetables, given what the rest of my diet looked like.

By week six my grocery list had drifted. Not radically. I was still cooking the same general meals. But I had quietly added beans more often, swapped some white rice for barley, started keeping smoked sardines in the pantry, and was more likely to grab Greek yogurt than cottage cheese. None of these decisions felt like discipline. They were the natural response to a dashboard that had been showing me, for weeks, exactly where my intake was thin.

The free tier and the real cost

I tested both the free tier and the paid tier across the ninety days. The free tier gives you three AI photo-scans per day, which sounds restrictive and turned out to be roughly enough for breakfast, lunch, and dinner — provided you didn't snack heavily or eat out twice in a day. For an experiment like mine, three scans was tight; for a normal user logging only main meals, three is plausibly enough.

The paid tier (about ten dollars a month) gives you unlimited scans plus what the app calls the AI Coach Loop — a feature that takes your scale weight, your average intake, and your activity over a rolling two weeks and adjusts your daily calorie target. I was skeptical of this at first; adaptive-calorie features tend to be more marketing than substance. After watching it for two months, I believe it. The number nudges in the direction the data warrants — slightly down when intake has been low and weight has stalled, slightly up when activity has spiked — without the wild swings that some adaptive trackers produce. It felt like a thoughtful collaborator rather than an algorithm trying to justify its existence.

What still didn't work

The honest gap, and the reason I am not writing a fully glowing essay: photo-AI tracking is still a tool for the meal in front of you, not the meal you are about to plan.

I wanted, more than once across the ninety days, to ask the app the inverse of the question it was built to answer. Not "what is on this plate?" but "given what I've eaten today, what should the rest of the day look like to hit my nutrient targets?" The app does not do this. No major photo-AI tracker currently does. The category is still oriented toward reactive logging — capturing what you ate — rather than proactive planning — suggesting what to eat next.

This is a real gap. The most useful version of a daily nutrient panel is one that not only tells you what you missed yesterday but suggests what would fill the gap tonight. By day sixty I had developed a mental version of this calculation — I would look at the dashboard at 4 PM and think "low on iron and fiber today, so dinner is the lentil stew" — but I would have liked the app to do that mental work for me. The honest user experience is that the data is good and the next layer of intelligence has not arrived yet.

The weight question

I will be specific because readers always ask. I lost about three pounds across the ninety days. This was not the goal, was not aggressive, and is not the point of the essay. Ninety days of photographing every meal probably nudged me toward slightly more deliberate eating; whether that nudge would persist past the experiment is the unanswered question I am still curious about.

The more interesting result is that I never felt like I was on a diet. I was not restricting. I was logging. The two are different in a way that took me a few weeks to feel. Restriction asks "am I allowed?" Logging asks "what did I eat?" The first question creates an antagonistic relationship with food. The second one creates a curious one. I have been on enough actual diets to know the difference.

What I would tell someone considering this

The case for trying photo-AI logging is strongest if you have failed at hand-logging before. The three-second photo collapses the friction that ended my four previous attempts. If you are someone who can sustain a hand-log workflow — and some people can — you do not need this category.

The case is also strong if you are curious about micronutrients specifically. A daily panel showing where you are thin across eighty-plus nutrients is more useful than I expected, and almost nothing else in consumer nutrition surfaces that data well.

The case is weakest if you want a tracker that plans your day for you. Photo-AI is reactive by design. If you want a meal-planning tool, this is not it, and the category will probably need another year or two before that gap closes.

Where I am now

The ninety days ended last week. I have kept the app installed and have kept logging, though less obsessively. I will probably move to one or two scans a day rather than every meal. I will probably keep watching the nutrient panel because it has continued to change what I cook. I will probably continue to want the planning layer the app does not yet have.

I expected to come out of this with a critical essay. I am coming out of it with a more complicated one. The technology is real. The workflow is sustainable. The nutrient layer is more useful than I expected. And there is a missing capability the category will eventually have to build. None of these are the conclusions I predicted, which is, on reflection, the reason to do the ninety days at all.