AI Meal Management 'VITL' Discloses Where Photo Analysis Missed — Most Common Correction Was 'Portion Size Error' at 43% of 93 Cases, 190 User Surveys Average 3.92 Points

ZeroGym369 Inc. (headquartered in Fukuoka City, Fukuoka Prefecture, CEO: Yusuke Yasukochi, hereinafter "the company"), developer and operator of the AI meal management service "VITL (Vital)," has published usage statistics on its official media. VITL is a Japanese AI meal management service that estimates calories, protein, fat, and carbohydrates from food photos, scores each meal, and provides personalized suggestions for "what to eat next."

The data released includes: (1) breakdown of user corrections to AI analysis from over 3,100 photo-based meal logs (93 corrections), (2) actual PFC balance observed from 1,262 days of meal records (83 users), and (3) aggregated results from 190 user surveys. No numerical adjustments or selective filtering of responses were made.

What is VITL (Vital)?

VITL is a browser-based Japanese AI meal management service (no app store download required) that analyzes food photos to estimate nutritional content, scores meals instantly, and provides feedback. Unlike simple tracking tools, VITL is designed as a coaching service that tells users "what to do next," inspired by the concept of hiring a personal AI trainer.

The pricing model is a 21-day program for ¥49,800 (tax included), with users who wish to continue beyond day 22 moving to a monthly subscription of ¥4,980 (tax included). The service is operated by ZeroGym369 Inc. in Fukuoka City.

Figure ①: Over 3,100 Photo-Based Meal Logs — How Users Corrected AI Analysis (Breakdown of 93 Corrections)

Nutritional estimation from photos works by "inferring from visible content," so errors can occur. The company allows users to add a note to correct AI analysis, and over 3,100 such corrected meal logs exist (as of August 5, 2026). From two time windows (July 22, 2026, and August 4–5, 2026), 85 logs with 93 corrections were analyzed and categorized.

- Portion size error: 43% - Misidentification of ingredients or dishes: 28% - Cooking method or seasoning (oil, flavoring): 19% - Missed items: 12% - Over-detection or leftovers: 10% - Additional intake: 8% - Correction based on nutrition labels: 5% - Date or meal timing correction: 2%

(n=93. Margin of error for each percentage is approximately ±10 points. Total exceeds 100% as one correction may fall into multiple categories.)

62% of corrections (43% portion size + 19% cooking method/seasoning) involved information "not determinable from photos alone." The amount taken from a shared dish or the quantity of oil used cannot be confirmed even by human dietitians from photos. The company's design allows users to supplement such ambiguous information with a brief note.

⚠️ These figures represent "what types of errors users corrected," not "how often AI is wrong." Accuracy or error rates cannot be derived from this data (undetected errors are not recorded).

Figure ②: Real Eating Habits Revealed from 1,262 Days of Meal Logs

PFC balance was analyzed from 1,262 days of meal records from 83 users who logged three or more meals per day for at least five days in June 2026 (median 15 days per user; 2 internal test accounts excluded).

Eating patterns were polarized: 37 users followed a "low-carb diet" (carbohydrates <40%), 37 followed a "balanced diet" (carbohydrates ≥50%), and 9 were in between.

Only 16 out of 83 users (19%) met all three national dietary guidelines (ages 18–64: protein 13–20%, fat 20–30%, carbohydrates 50–65% of energy), and all were in the balanced diet group.

No user fell below the minimum protein guideline (13% energy); median daily protein intake was 83.9g.

Only 213 out of 1,262 days (17%) met all three guidelines. Even within individuals, the median difference between daily maximum and minimum protein intake was 62g, indicating high day-to-day variability.

⚠️ These figures do not represent the average Japanese population but reflect meals recorded by 83 users of a diet management service in June 2026. Quantities are AI estimates from photos, not weighed measurements. Nutritional balance varies by age, body type, activity level, and health status. Consult a physician or dietitian for personal advice.

Figure ③: 190 User Surveys — Full Breakdown of Feedback and Improvement Requests

User responses to an in-app survey (April 5 to September 7, 2026; 190 valid responses, 6 questions) were aggregated without filtering or numerical adjustment.

Overall satisfaction (5-point scale):

- 5 (Very satisfied): 56 responses (29.5%) - 4: 75 responses (39.5%) - 3: 48 responses (25.3%) - 2: 10 responses (5.3%) - 1 (Dissatisfied): 1 response (0.5%)

Average: 3.92 points (5-point scale, n=190, in-app survey by ZeroGym369 Inc.). 131 responses (68.9%) rated 4 or higher.

The most common improvement request was "accuracy of ingredient and portion detection from photos" (28 of 156 responses up to August 1, 2026), but most cases involved requests for information not visible in photos, consistent with Figure ①.

The second most common, "previous inputs not reflected in future advice" (17 responses), was a bug fixed on August 1, 2026. In 30 responses after August 8, requests for these two issues dropped to 1 and 0, respectively, while requests like "make advice more specific and context-aware" and "remember past interactions" rose in frequency.

Average satisfaction before improvements (July 17, 2026 onward): 3.97 (121 responses); after: 3.84 (69 responses). The post-improvement score is lower, but this may reflect response timing (concentrated 7–8 days after onboarding), making it difficult to isolate the improvement effect.

⚠️ These figures cannot be directly compared to app store ratings due to different collection methods. The 190 respondents do not represent all VITL users. Individual responses are not disclosed (only aggregated, anonymized data is published per company privacy policy).

Why Disclose Even Unfavorable Data?

Prospective users of AI meal management services most want to know: "How accurate is it really?" and "How do actual users eat and rate it?" VITL launched in 2026 and is still in the early phase before third-party reviews accumulate. The most reliable way to provide decision-making information at this stage is to disclose internal data without correction.

Full reports are available on the official media site:

Breakdown of how users corrected AI analysis: https://media.vitl.jp/articles/ai-meal-photo-accuracy/

PFC balance from 1,262 days of meal logs: https://media.vitl.jp/articles/p