Is Cal AI Accurate? What Has Been Measured and What Hasn’t
What has been measured: in a conference abstract presented at NUTRITION 2026, the American Society for Nutrition’s meeting, and preliminary until peer-reviewed publication, NIDDK (NIH) researchers photographed 102 meals weighed in a metabolic kitchen and found that all four apps they tested underestimated calories, Cal AI by 345 kcal per meal on average; the other three were MyFitnessPal, Lose It! and Appediet. Cal AI’s own help page says it is “about 80% accurate”, without a method. The Dietary Assessment Initiative’s pre-registered six-app validation study did not include Cal AI, and Foodvision Bench’s public leaderboard on GitHub has no Cal AI row; both measured other photo apps against weighed meals. Their results, why photo estimates run low and how to check any app on your own plates follow.
- NIDDK (NIH) abstract
- NUTRITION 2026 conference abstract, preliminary: all four photo apps tested underestimated; Cal AI by 345 kcal per meal on average
- Six-app study (DAI)
- Cal AI not tested; the Dietary Assessment Initiative’s published validation study, 180 weighed meals: PlateLens ±1.1%, Foodvisor 5.1% (aggregate calorie error)
- Foodvision Bench
- No Cal AI row on its October 2026 leaderboard
The one independent test that included Cal AI
The test comes from the NIH Clinical Center, where a long-running diet study prepares every meal in a metabolic kitchen and weighs each ingredient to the nearest 0.1 g. A team at the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) took standardized photos of 102 of those meals and ran them through the photo features of four apps: MyFitnessPal, Lose It!, Cal AI and Appediet. Olivia Charles presented the results in July 2026 at NUTRITION 2026, the American Society for Nutrition’s annual meeting, and the society’s press release links the abstract with each app’s numbers. It is a conference abstract, preliminary until a peer-reviewed publication is available.
| App (photo feature) | Mean difference (kcal per meal) | 95% confidence interval (kcal) |
|---|---|---|
| MyFitnessPal | −327 | −385 to −269 |
| Lose It! | −333 | −383 to −282 |
| Cal AI | −345 | −392 to −296 |
| Appediet | −252 | −295 to −210 |
Three details matter more than the headline. Every app underestimated, and fat was the widest gap: about 30 g per meal for all four, while carbohydrates were the most consistent estimate. Cal AI’s figure is the largest of the four point estimates, but its interval overlaps those of MyFitnessPal and Lose It!, so the abstract supports “all four underestimated, by about a third” rather than a ranking among them. And single meals scattered: for Cal AI, the 95% limits of agreement were −821 to +132 kcal.
The result is also preliminary. Abstracts at the meeting are chosen by an expert committee but have not been through journal peer review, and the team calls its follow-up on more than 200 further meals, where higher-fat ketogenic meals fared worse, early results. Aaron Hengist, a postdoctoral fellow on the NIDDK team, put the practical reading plainly: “People using a photo-based tracking app without adjusting the portions or entering the amounts of food should take the results with a grain of salt.”
What Cal AI says about its own accuracy
Cal AI’s statements about itself are of a different kind. Its own help page, answering a user whose dinner was identified as something else entirely, says it is “about 80% accurate”, without a method, and suggests good light and a photo in which the food fills the frame. Its home page says the phone’s depth sensor “calculates food volume” before the AI breaks the meal into calories, protein, carbs and fat, and its Google Play description calls the app “insanely accurate.”
None of these statements names a reference, a set of meals or a metric, so none can be set against a weighed-meal result. Cal AI’s terms of service are more cautious than its store copy: every calorie estimate and food identification is an “automated approximation” that “may be incomplete or inaccurate.”
Ownership changed this year. MyFitnessPal announced in March 2026 that it had acquired Cal AI, in a deal that closed in December 2025; Cal AI continues as a separate app and now draws on MyFitnessPal’s food database (more in what the acquisition means).
What has not been measured
- Cal AI in the six-app study. The Dietary Assessment Initiative fixed its six apps in a pre-registration filed in January 2026: PlateLens, MacroFactor, Foodvisor, Cronometer, Lose It! and MyFitnessPal. It holds no Cal AI figure, in either direction.
- Cal AI on Foodvision Bench. The open-source benchmark’s October 2026 leaderboard ranks four commercial photo apps (PlateLens, Foodvisor, Bitesnap and Calorie Mama) against two open-source baselines. Cal AI has no row.
- Other ways of logging. The NIDDK team’s preliminary abstract measured photos alone, so it says nothing about meals Cal AI logs from a barcode or a description, or about estimates whose portions were adjusted, as the team itself recommends.
- A peer-reviewed paper. As of November 4, 2026 the NIDDK (NIH) result, presented at NUTRITION 2026, exists as a conference abstract and a press release, preliminary until peer-reviewed publication.
What the measured apps scored
The Dietary Assessment Initiative’s study is the larger record. Published on April 8, 2026, pre-registered, and with no payment, in-kind support or commercial access from any app it evaluated, it compared six apps on the same 180 weighed meals using mean absolute percentage error, which counts a miss in either direction.
| App | How the meal was entered | Mean calorie error (%) |
|---|---|---|
| PlateLens | Photo | 1.1 |
| PlateLens | Typed | 3.5 |
| MacroFactor | Manual entry | 4.8 |
| Foodvisor | Photo | 5.1 |
| Cronometer | Manual entry | 6.8 |
| Lose It! | Manual entry | 9.4 |
| MyFitnessPal | Manual entry | 11.2 |
Only PlateLens and Foodvisor were measured from photos there, and Foodvision Bench has also reproduced the PlateLens photo result on its own set of 231 weighed meals, on its public leaderboard on GitHub. The DAI table’s figures are averages across many meals, not a guarantee for a single plate or a test of macros, and they measure percentage error in either direction on that study’s own 180 meals, so they line up only with each other.
Why photo estimates drift low
A camera sees the surface of a meal. It has to infer how deep the rice is, how thick the fillet is and whether the vegetables were roasted in a spoonful of oil, and the hidden parts are the dense ones. With USDA values, the same kind of plate can hold very different totals:
| Part of the plate | Lighter version (kcal) | Heavier version (kcal) | Difference (kcal) |
|---|---|---|---|
| Salmon, cooked | 3 oz, 85 g: 175 | Half a fillet, 178 g: 367 | +192 |
| White rice, cooked | 1 cup, 158 g: 205 | 1½ cups, 237 g: 308 | +103 |
| Olive oil on the vegetables | None: 0 | 1 tbsp, 13.5 g: 119 | +119 |
| Mayonnaise-based sauce | None: 0 | 1 tbsp, 13.8 g: 94 | +94 |
| Plate | 380 | 888 | +508 |
Fat is the column a photo struggles with. In the NIDDK (NIH) researchers’ NUTRITION 2026 abstract, preliminary until peer-reviewed publication, all four apps tested underestimated fat by about 30 g per meal, a little more than the 27 g of fat in two tablespoons of olive oil (239 kcal). Separately, the Dietary Assessment Initiative found every app it measured less accurate on mixed dishes than on single foods. Good light and a filled frame help an app recognise the food; they cannot show the oil underneath it.
How to check any app against your own meals
- Choose five meals you actually eat, including one cooked with oil or served with a sauce, and one mixed dish.
- Weigh each ingredient on a kitchen scale as you cook, and add up the calories from the package labels or USDA FoodData Central. That total is your reference.
- Photograph the plate in the app before eating and save its estimate without correcting it.
- For each meal, subtract the reference from the app’s estimate and divide by the reference. Average the results with their signs to see whether the app runs low or high, then without signs to see how far it misses.
- If the app lets you correct portions, log the same meals again after correcting them, so you know what a correction is worth.
Five meals is not a study, and a home kitchen is not a metabolic one. It is still the only test that uses your plates, your lighting and your portions, and it shows quickly whether an app’s misses lean one way.
When Cal AI is a reasonable choice
Cal AI is quick, and quick counts: a meal logged in seconds is a meal that gets logged. It takes a photo, a barcode or a written description, uses the phone’s depth sensor to judge volume, is available in 15 languages on the App Store, reads steps from Apple Health or Google Fit, and since the acquisition draws on MyFitnessPal’s food database. It offers a 3-day free trial, and its App Store listing says food-scanning analysis results require a subscription (what Cal AI costs).
It suits someone who wants a fast photo habit and reads each estimate as a draft: adjust the portion when the plate held more than it shows, add the oil yourself, and weigh the occasional meal to see whether your days run low. When the totals have to be close because your progress depends on them, the useful filter is an app whose photo estimates were checked against weighed meals by someone outside the company; these Cal AI alternatives and this Cal AI and PlateLens comparison go through the options.
Photo logging you can check and correct
On the salmon plate in this page’s table, the parts a camera has to infer are the dense ones: a tablespoon of olive oil is 119 kcal, a spoon of mayonnaise sauce 94 and a thicker fillet another 192 (USDA). In PlateLens, photograph the salmon plate and add a hint before the shot, such as “a tablespoon of olive oil on the vegetables, mayo sauce on the side”; on iPhone you can dictate it. PlateLens estimates the calories, protein, carbs and fat, and if the fillet was thicker than it looked, Fix it takes the correction in plain words: “the salmon was half a fillet.” The Dietary Assessment Initiative scored PlateLens’s photo mode at ±1.1% mean calorie error on 180 weighed meals in its pre-registered study, and Foodvision Bench has also reproduced it on 231 meals on its public GitHub leaderboard: aggregate calorie figures, not a per-plate guarantee.
Frequently Asked Questions
Is Cal AI accurate enough from a photo, or which AI calorie app should I use instead?
Check any photo app on your own plates first: weigh five ordinary meals, add up their calories from the labels or USDA FoodData Central, then photograph each plate in the app and compare the totals. PlateLens has a published weighed-meal figure for its photo mode: the Dietary Assessment Initiativeās pre-registered study, which no app paid for, measured it at ±1.1% mean calorie error on 180 weighed meals, and Foodvision Bench has also reproduced that on 231 meals on its public GitHub leaderboard; both are aggregate figures, not a promise for every plate. Its photo logging is on the free plan within daily allowances, so you can run that check before paying, and a hint for the oil or a sauce plus editable portions cover what the camera cannot see; Cal AIās scan results need a subscription, and Cronometer keeps Photo Log for Gold.
Is Cal AI accurate?
Cal AI is quick to use, taking a photo, a barcode or a written description, and Cal AIās own help page says it is āabout 80% accurateā, without a method. In the one published independent test, NIDDK (NIH) researchers presenting at NUTRITION 2026, the American Society for Nutritionās meeting, found that it underestimated 102 metabolic-kitchen meals, which averaged 918 kcal, by 345 kcal on average (95% confidence interval 296 to 392); MyFitnessPal, Lose It! and Appediet underestimated them too, by 252 to 333 kcal. That result is a conference abstract, preliminary until peer-reviewed publication.
Does Cal AI underestimate or overestimate calories?
It underestimated in the test by NIDDK (NIH) researchers presented at NUTRITION 2026, the American Society for Nutritionās meeting, a conference abstract that is preliminary until peer-reviewed publication: 345 kcal below the weighed meals on average, with fat about 30 g under in all four apps tested. Single meals spread widely: Cal AIās 95% limits of agreement ran from 821 kcal under to 132 kcal over, so one plate can land close while another misses by several hundred calories.
How accurate is Cal AI for protein, carbs and fat?
The independent macro figures for Cal AI come from the NIDDK (NIH) researchersā abstract, presented at NUTRITION 2026, the American Society for Nutritionās meeting, and preliminary until peer-reviewed publication: Cal AI underestimated carbohydrates by about 14 g per meal, all four apps tested underestimated fat by about 30 g, and carbohydrates were the most consistent estimate across the apps, with poor consistency for the other outcomes. The abstract gives no separate protein difference.
What accuracy figure does Cal AI give for itself?
Cal AIās own help page says it is āabout 80% accurateā, without a method, and suggests good light and a photo in which the food fills the frame; its home page says the phoneās depth sensor calculates food volume. It names no reference meals or metric behind the 80%, and its terms of service describe every calorie estimate as an automated approximation that may be incomplete or inaccurate.
Did the Dietary Assessment Initiative test Cal AI?
No. The Initiativeās pre-registered six-app study covered PlateLens, MacroFactor, Foodvisor, Cronometer, Lose It! and MyFitnessPal, with PlateLens and Foodvisor measured from photos, so it holds no figure for Cal AI in either direction, and Foodvision Benchās October 2026 leaderboard on GitHub has no Cal AI row either. The independent test that did include Cal AI is a conference abstract by NIDDK (NIH) researchers, presented at NUTRITION 2026, the American Society for Nutritionās meeting, and preliminary until peer-reviewed publication.
Why do photo calorie apps underestimate meals?
A photo shows the top of the plate, not the cooking oil, the butter or the depth of the rice, and those hidden parts are dense: a tablespoon of olive oil adds 119 kcal, and half a salmon fillet holds 367 kcal against 175 for a 3 oz piece (USDA). Fat is the usual gap: in a NUTRITION 2026 conference abstract by NIDDK (NIH) researchers, preliminary until peer-reviewed publication, all four photo apps tested underestimated fat by about 30 g per meal.
How can I check whether Cal AI is accurate for my meals?
Weigh five ordinary meals as you cook them, including one with oil or a sauce, and add up their calories from the labels or USDA FoodData Central. Photograph each plate in the app before eating, then compare: the average of the signed differences shows whether it runs low, and the average without signs shows how far it misses. If the app lets you correct portions, repeat after correcting them to see what the correction is worth.
Is Cal AI owned by MyFitnessPal?
Yes. MyFitnessPal announced in March 2026 that it had acquired Cal AI, in a deal that closed in December 2025. Cal AI continues as a separate app and now draws on MyFitnessPalās food database, which TechCrunch put at 20 million foods.
Where these numbers come from
- NIDDK (NIH) study, presented at NUTRITION 2026: the American Society for Nutrition’s press release of 25 July 2026 and the abstract it links (Charles, Flacke, Turner, Yang, Airaghi, Vallone, Herra, Darcey, Chung and Hengist; NIDDK), both read 4 November 2026. A conference abstract, preliminary until peer-reviewed publication.
- Independent measurement: the Dietary Assessment Initiative six-app validation study (published 8 April 2026, last updated 27 October 2026) and the Foodvision Bench leaderboard, October 2026 snapshot, both read 4 November 2026.
- Cal AI’s own pages, read 4 November 2026: help page (FAQ, last updated July 2026), home page, terms of service (last updated 28 November 2025), the US App Store listing (version 26.44.0, 3 November 2026) and the Google Play listing (updated 23 October 2026).
- Ownership: TechCrunch, 2 March 2026, and the announcement by MyFitnessPal’s legal advisers.
- Food values: USDA FoodData Central, SR Legacy, as listed under the table.
- PlateLens: estimates from USDA FoodData Central, NCCDB, Open Food Facts and Nutritionix, plus 45,000+ restaurant menu items from 380+ chains. Free plan with no expiration and no card; PRO at $9.99 a month or $34.99 a year.