How accurate is Cal AI, by its own account and in research

In brief

Cal AI's FAQ says "about 80% accurate", while its blog claims 90% on visible foods, with no measurement method published for either figure on the pages checked October 6, 2026. A PubMed search that day found no peer-reviewed study testing Cal AI by name. A conference abstract from NIH researchers tested 102 weighed meals in four photo apps, including Cal AI, and found that all four counted calories and fat low. That abstract is not peer-reviewed and the release gives no Cal AI-only number.

What Cal AI says about its own accuracy

"About 80% accurate" is the wording in both Cal AI's FAQ and the home page FAQ, checked October 6, 2026. Neither answer defines what the percentage measures.

The blog's 90% claim has a stated scope. It covers calories, protein, carbs and fat for visible foods. The same post says hidden ingredients make smoothies and soup harder to scan, and suggests describing those foods in words.

Across the home page, FAQ, terms, citations page and blog posts checked October 6, 2026, Cal AI publishes no measurement method, reference or sample size for either percentage. It does not explain the difference between 80% and 90%.

For scan problems, the FAQ recommends bright lighting, a focused image and an updated app. A blog post says the AI learns from feedback, and another describes a "fix results" button for explaining an error.

The terms call the estimates automated approximations that may be incomplete or inaccurate. Cal AI disclaims liability for wrong calorie counts or food identification and gives no warranty of accuracy. Its citations page, dated August 2026, supports goals and nutrition targets rather than food values or scan accuracy.

On Google Play, the description says "Insanely accurate" without a percentage or method. That listing was also checked October 6, 2026.

Where Cal AI's numbers come from

On its product page, Cal AI describes using the phone's depth sensor to estimate food volume from a photo. AI then breaks down the meal and estimates its calories and macros. The same page advertises a database of over 1 million foods, searchable by food name, brand or barcode.

The FAQ lists manual search alongside photo and barcode scans. Cal AI's blog adds label recognition and typed descriptions. None of the checked home, FAQ, terms, citations or blog pages names the nutrition databases behind Cal AI's food values. The terms also allow uploaded content to be used for AI training.

Cal AI's FAQ says starting a food log opens the paywall. Its App Store description requires a subscription to see scan analysis results, while calai.app advertises a 3-day free trial.

Studies that tested Cal AI by name

A PubMed search on October 6, 2026 found no peer-reviewed study that tested Cal AI's accuracy by name. The search covered Cal AI, Cal-AI and CalAI spellings, along with broader searches for calorie apps, food photos and accuracy.

The NIH conference abstract is not peer-reviewed

NIH researchers presented a NUTRITION 2026 conference abstract using photos of 102 meals. A metabolic research kitchen prepared the meals with ingredients weighed to the nearest 0.1 gram. Researchers submitted the photos to MyFitnessPal, LoseIt!, CalAI and Appediet.

All four apps counted calories low by about 250 to 345 calories per meal on average, and fat by about 30 grams. Carbohydrate estimates were more consistent. The release reports that range across the four apps and supplies no separate Cal AI figure.

The release says these results are preliminary until a peer-reviewed paper is available. A conference abstract does not establish a published accuracy percentage for Cal AI.

Sreedevi et al.'s 2026 protocol in Trials plans to use CalAI for plate-photo calorie assessments but reports no accuracy test, as described in the trial protocol.

What research on photo calorie estimates found

None of these three studies tested Cal AI. They assessed other apps or general chatbots.

Li et al. measured food recognition and energy estimates in 2024

In Nutrients, Li et al. compared nutrition outputs with food records for Western, Asian and Recommended diets. Their photo test used 22 images containing 39 food components and 7 mixed dishes, submitted to 7 apps. The published study did not test Cal AI.

MyFitnessPal identified 97% of food components correctly and Fastic identified 92%. Those percentages measure food naming. The authors found automatic calorie estimates inaccurate, with large differences for mixed dishes even when foods were identified correctly. These were Australian store versions from 2024 and a small set of images, so the findings do not describe every food or current app version.

O'Hara et al. asked ChatGPT to estimate weighed meals in 2025

O'Hara et al.'s Nutrients study gave ChatGPT-4 114 meal photographs and compared its estimates for 16 nutrients with known meal values. Seven dietitians also estimated calories, protein and carbs for 38 photos. The paper tested a general chatbot and did not test Cal AI.

Food identification precision was 93.0%. Estimated weight agreed well for small meals, but poorly for medium and large meals. The authors found poor agreement for 10 of 16 nutrients. Differences exceeded 10% for 13 nutrients, and ChatGPT underestimated 11. They also noted that ChatGPT did not directly access nutrition databases when answering.

Fridolfsson et al. compared three chatbots in 2025

Fridolfsson et al., in Current Developments in Nutrition, used 52 standardized food photos with small, medium and large portions. They compared ChatGPT-4o, Claude 3.5 Sonnet and Gemini 1.5 Pro against direct weighing and Dietist NET nutrition database calculations. This comparison of general chatbots did not test Cal AI.

Mean absolute percentage error for weight was 36.3% for ChatGPT and 37.3% for Claude. Both had 35.8% error for energy. Gemini's errors across nutrients were 64.2% to 109.9%. All models underestimated more as portions grew. The authors pointed to large-portion errors and variable macro estimates as limits on uses requiring precise nutrition amounts.

Fridolfsson et al. judged ChatGPT and Claude comparable with traditional self-reported diet methods, while O'Hara et al. found ChatGPT poor at nutrient amounts. Across these studies, naming a food was easier than estimating how much of it was on the plate.

What a plate photo can miss

Li et al. describe the difficulty of detecting butter, oils, salt and sauces added during cooking. In their eggs on toast with butter example, MyFitnessPal omitted butter and underestimated calories by 35%. Foodvisor's estimate for that image was 73% low. These results concern those tested apps.

Li et al. also reported missed components in mixed dishes and errors in portion size or volume. Fridolfsson et al. describe the difficulty of estimating volume from a flat image. In O'Hara et al.'s photos, some foods were hidden under others in soups, sandwiches and salads.

The NIH conference release reports more difficulty with high-fat meals, whose fat was consistently underestimated. Cal AI itself acknowledges hidden ingredients in smoothies and soup as harder cases for scanning.

Checking one logged meal

1. For packaged food, compare the entry with its Nutrition Facts label. The FDA explains that nutrient amounts, including calories, refer to one serving.
2. For a food without a label, look it up in USDA FoodData Central, the USDA's food composition data resource.
3. Weigh the portion with a food scale and compare it with the logged amount. Portion estimation is one error Li et al. identified.
4. Edit the entry when its food or amount differs. Cal AI describes a "fix results" button for telling the app what was wrong.

How other apps describe their nutrition sources

Cronometer names NCCDB and USDA SR28 among its data sources and says submitted foods undergo verification, checked October 6, 2026.

MacroFactor says a large part of its common-food research data comes from NCC. A legacy database owned by another company remains a backup for barcode scans and supplies common foods for AI describe, checked October 6, 2026.

MyFitnessPal uses a check mark for foods it reviewed or added, but warns that marked foods can still contain mistakes. Its help page mentions tools and resources without describing the review method, checked October 6, 2026.

TrueCal's item receipts carry labels such as USDA, label and menu. Its product facts say nutrition estimates can be wrong, checked October 6, 2026. The guide to apps with source information covers the details a logger can inspect.

Foodvisor describes recognizing foods and estimating quantities from a photo. Its home page names no nutrition data source or accuracy figure, checked October 6, 2026. Read the Foodvisor review for a closer look.

Who Cal AI fits, and other picks

Cal AI fits someone who wants to start a food log with a photo and accepts the approximations described in its terms.

For more about the app, open the Cal AI review. To consider other logging options, read alternatives to Cal AI. The Cal AI and MyFitnessPal comparison puts those apps side by side, while the MyFitnessPal accuracy guide follows the published research on that app.

TrueCal offers free manual calorie entries, with macros and logging by typed description, voice, photo or barcode in Pro. The Pro subscription is $29.99 a year or $9.99 a month. Eligible customers can try Pro free for 14 days in the iPhone or Android app, through a trial run by the App Store or Google Play.

Cal AI accuracy, asked and answered