AI Nutrition Coach Apps: What a Chat Can and Cannot Do

An AI nutrition coach app is a chat that reads what you log and answers questions about it. It reads well and counts poorly. A 2023 test of 222 foods found ChatGPT’s energy estimates within 10% of a food database for only 35% to 48% of the items. Use the chat to read your diary, and leave the arithmetic to a database.

What is an AI nutrition coach app?

It comes in two shapes, a food tracker with a chat built in or a chat sold as the tracker. You type or speak what you ate, ask what to cook, and a language model answers. The word “coach” in an app’s name says nothing about qualifications. Apps that pair you with a person are a different product, and this article covers only the chat. So is an adaptive coach that resets your target from your weigh-ins, and the cost and review of an adaptive coach looks at one.

Does the chat change what people eat? The evidence is thin. A 2025 systematic review pulled together 7 randomized trials of chatbots that talk in open text. Their effect on diet and physical activity in adults was inconsistent, and the authors asked for larger trials. That limits what an app can promise. Before you pay, decide what you want from it. Quicker logging and answers about your own day take ten minutes to test. Motivation is the claim I’d trust least, given how inconsistent that review found the effects.

How accurate are the calorie numbers a chat gives you?

Not very, unless a database sits behind them. The 2023 study asked ChatGPT-3.5 and ChatGPT-4 for the energy in 222 foods, five times each. It compared the answers with a national food composition database. Only about 35% to 48% landed within 10%. Those were 2023 models.

A newer benchmark, NutriBench (2025), has 11,857 meal descriptions, and tests twelve models on carbohydrates. Averaged over all twelve with plain prompts, 43.24% of answers landed within 7.5 g of the true value. The best score was 66.82%, from GPT-4o with step-by-step reasoning. GPT-4o and its smaller sibling did better when servings were given in grams.

One phrase can hide very different grams. To see how much, I looked up three foods in USDA records and priced the same words at different amounts.

What you say	Amount	Energy	USDA record
A cup of rice	158 g	205 kcal	Rice, white, long-grain, cooked (FDC 168878)
A big bowl of rice	250 g	325 kcal	same record, 130 kcal per 100 g
A bowl you refilled	350 g	455 kcal	same record
A teaspoon of olive oil	4.5 g	40 kcal	Oil, olive, salad or cooking (FDC 171413)
A tablespoon of olive oil	13.5 g	119 kcal	same record, 884 kcal per 100 g
A glug, about 3 tablespoons	40.5 g	358 kcal	same record
A tablespoon of peanut butter	16 g	96 kcal	Peanut butter, smooth style, without salt (FDC 172470)
Two tablespoons	32 g	191 kcal	same record, 598 kcal per 100 g
A heaped scoop	50 g	299 kcal	same record

The cup and the tablespoons come straight from the records. The bowls, the glug and the scoop are my own guesses at how people pour, which is the point. Rice alone swings by 250 kcal between a cup and a refill.

Now put the study’s margin on a normal day. Take a 600 kcal dinner. Within 10% means it could really be 540 or 660, and a 20% miss leaves that one meal off by 120 kcal.

Say grams, in the state you weigh the food. “Rice, 250 g cooked” gives the chat something to multiply.

Does the same question get the same number twice?

Nearly. That is the less worrying half. The 2023 study asked for each food five times and measured the spread. The average coefficient of variation for energy ran from 4.7% to 6.2% across the two models and two input languages (Hoang et al., 2023). So an answer wobbles by about 5% between asks, which is 30 kcal on a 600 kcal dinner.

The consistent miss is the bigger problem. Within 20%, the models got 58% to 61% of the foods right, so about two foods in five were off by more than a fifth. The prompt listed the foods raw, which is an easier job than your cooked plate.

Can a chat plan your meals?

It can propose a menu. It can’t be trusted to hit your target. In a 2024 Scientific Reports paper, ChatGPT wrote weekly plans for 20 user profiles. The plans missed the energy target by around 17% on average, with several meals 30% over or under. The authors’ own system used an optimizer to scale portions to the target and missed by 0.00.

Models pick foods well. Arithmetic is where they slip. An app that lets the model write the totals itself is asking for the 17%. Ask a chat for a menu, then add up the total yourself. If the two totals disagree, believe your sum.

What do I do with a plan the chat wrote?

Add it up before you eat from it. Take a target of 1,800 kcal. A 17% miss is 306 kcal, so the day could land at 1,494 or at 2,106. The authors’ own system fixed this by scaling portions to the target (Papastratis et al., 2024), and you can do the same by hand.

1. Copy the menu into your diary with grams for every item.
2. Let the diary add it up from database records, and compare that total with the chat’s.
3. Multiply every portion by your target divided by the real total. A 2,000 kcal plan against an 1,800 kcal target means multiplying by 0.9.

Health conditions raise the stakes. A 2024 study gave 10 general chatbots a complex case: a 65-year-old woman with obesity, type 2 diabetes, muscle loss and kidney disease. Registered dietitians scored the advice. No chatbot reached 50% accuracy, and most advised both cutting and raising protein. If a health condition changes what you can eat, your plan is a conversation with your doctor or a registered dietitian, not a chat.

One more test for any plan is the daily total. The 2013 AHA/ACC/TOS guideline lists 1,200 to 1,500 kcal a day for women and 1,500 to 1,800 for men as usual targets. A good app refuses a goal under the bottom of those ranges.

What is a chat good for?

It is good at reading and rewriting. It can turn “two eggs and toast with a bit of butter” into foods you confirm. It can tell you how much protein you logged this week. It can suggest dinner for what’s left of your day.

On speed, the NutriBench comparison is stark: three nutritionists, barred from looking anything up, averaged 43 minutes for 72 meal descriptions, and the models took a few minutes. One nutritionist who was allowed to look foods up then scored about as well as the best model. That is a single person, but it fits my view that the lookup does the work and the chat makes it quick. The GPT models did best on complex, multi-part meals with detailed measurements, while the nutritionists did better on simpler traditional meals. Use the chat for the messy entry, and check the simple ones yourself.

I use the chat to draft an entry, then confirm the grams against a database price.

Does a coach make me log more?

That is the claim I would test, because logging often is the habit with evidence behind it. In a 24-week online program, people who lost at least 5% of their weight opened the food journal 2.4 times a day. Those who lost less opened it 1.6 times (Harvey et al., 2019). It is an association and does not show that opening the journal caused the loss.

A 6-month pilot of 128 adults points the same way. People given a phone app recorded food on 92 days on average, against 29 days for a paper diary (Carter et al., 2013). The researchers built that app themselves, so it can’t rank products. What both studies point to is friction. A chat that turns a spoken sentence into an entry earns its fee only if you log on more days with it. Count your logged days for two weeks with the chat and two weeks without.

Does a photo help the chat?

Yes, when you add words to it. A 2025 study gave ChatGPT-5 195 dishes in four setups. Errors shrank as the setups added standard descriptions and then ingredient amounts. They grew again when the photo was removed but the ingredient list stayed. A photo shows a lot. Your grams add what it can’t.

Next time you send a photo, add the oil, the sauce and any weight you know. For how one popular photo app held up with real users, see these photo calorie app reviews.

What should you check before you trust one?

Run through this list. Each item takes a minute.

1. It says it is AI, with no human-sounding name or title.
2. Every number shows where it came from and the grams behind it.
3. You confirm portions before an entry is saved.
4. It answers from your own diary and says which entries it used.
5. You can see what it remembers and delete one fact at a time.
6. It won’t set a goal below the floor.

Items 2 and 3 matter most. Sources and grams let you catch the 17% before it lands in your day, and confirming portions removes the biggest unknown. The other four are guardrails: the AI label, the visible memory and the floor each cover one way a chat can mislead you.

What does a ten-minute trial look like?

Test it. Log a packaged food whose label you can read, ask the chat for the calories, and compare. Ask where the number came from. If it can’t say where the number came from, I treat it as a guess.

Then run six more checks, each with an answer you can verify.

Ask or do this	Reference	It passes if
Ask for the calories in 100 g of cooked white rice	130 kcal	The answer is within 10%, the margin the 2023 study used
Ask for 1 tablespoon (13.5 g) of olive oil	119 kcal	Within 10%
Ask for 2 tablespoons (32 g) of smooth peanut butter	191 kcal	Within 10%
Ask the rice question three times	The same food each time	The answers sit within about 5% of each other
Ask for a 1,000 kcal day as a woman	The bottom of the guideline’s usual range is 1,200 kcal	It warns or refuses
Say “I’m vegetarian”, then ask what it remembers	A visible memory	You can delete that one fact

With two or more failures out of six, I’d keep a database tracker for the numbers and use the chat for questions.

If cost decides it, is a photo calorie app free shows what one of them charges, and how to choose a calorie counter app runs the wider ten-check test.

Calsa is a paid app whose chat is labelled as AI, answers from your diary, goal and weight trend, and lets you see, edit or delete what it remembers.