Big data, AI and personalized nutrition: what actually works

“Big data”, “precision nutrition”, “AI-designed diets”: the promise is that an algorithm, by crossing your data with data from thousands of other people, will know what you should eat better than any general guideline. Part of that promise has serious studies behind it and part of it is marketing. Here you will see what trials have actually measured for five technologies: calories from a photo, algorithms that predict your blood sugar, commercial personalized diet programs, genetic tests, and glucose sensors in people without diabetes.

Calories from a photo with AI: what the studies measure

This is the most widespread use of AI in nutrition, and there are now independent studies that compare it with food weighed on a scale. At the University of Gothenburg (Fridolfsson and colleagues, Current Developments in Nutrition, 2025), researchers gave 52 standardized photos, of single foods and complete meals in three portion sizes, to ChatGPT-4o, Claude 3.5 Sonnet and Gemini 1.5 Pro. The first two were off by 36-37% on average for weight and 36% for calories; Gemini was off by between 64% and 110% depending on the nutrient. And all three underestimated more as the portion got bigger. In Ireland (O’Hara and colleagues, Nutrients, 2025), ChatGPT-4 analyzed 114 photos of real meals: it recognized the foods correctly 93% of the time and estimated the weight of small portions reasonably well, but not that of medium and large ones. In Japan (Isobe and colleagues, Nutrients, 2026), with 15 hospital meals, the best models estimated energy well, but every model overestimated fat by more than 20%.

So is it useless? Compare it with logging from memory

A 36% error sounds like a lot until you compare it with the usual alternative. In a classic study (Lichtman and colleagues, New England Journal of Medicine, 1992), 10 people with obesity who said they could not lose weight on under 1,200 kcal a day underreported what they ate by 47% on average, as measured in the lab. The Gothenburg authors conclude that ChatGPT and Claude perform at a level comparable to self-reporting, without the effort, but are not yet good enough where clinical precision is needed. A photo works for logging consistently, as long as you check the quantity:

• Take the photo from above, with the whole plate and a fork in view as a size reference.
• If the plate was big, raise the amount: the models tend to fall short.
• Add what you cannot see: oil, butter or sauce.
• For packaged food, use the barcode, which takes the data from the label.
• If you are unsure about a food, look it up in the food calorie table.

How Renzy does it (and what we do not promise)

Renzy, the app behind this blog, uses AI to estimate calories and macros from a photo of your plate, and it also reads barcodes. It is an estimate that you can correct: swap a food that was recognized wrongly or adjust the amount, which is the part most worth checking. We have not published an accuracy study of our own, so we will not give you an accuracy percentage; be wary of any app that advertises one without a published study behind it. Renzy offers a 7-day trial that requires a card, and there is no free plan.

Zeevi 2015: what the 800-person study really did

Glucose-personalized dieting was born at the Weizmann Institute in Israel (Zeevi and colleagues, Cell, 2015). The team tracked the blood sugar of 800 people for a week and their responses to 46,898 meals: the same meal triggered very different rises from one person to the next. With that data they trained an algorithm that combined blood tests, dietary habits, body measurements, physical activity and the gut microbiome to predict the glucose rise of each meal. They validated it in another 100 people and, in a blinded randomized trial, the diets it designed reduced blood sugar rises after eating. It did not measure whether that led to weight loss or less diabetes. Later (Ben-Yacov and colleagues, Diabetes Care, 2021), 225 adults with prediabetes followed either the algorithm’s diet or a Mediterranean diet for 6 months: on the algorithm’s diet, daily time above 140 mg/dL fell by 1.3 hours versus 0.3, and HbA1c fell by 0.16 points versus 0.08. A real improvement, but a small one in HbA1c.

PREDICT 1: even identical twins respond differently

PREDICT 1 (Berry and colleagues, Nature Medicine, 2020), run by King’s College London with the company ZOE, measured 1,002 healthy adults in the United Kingdom over two weeks and validated the results in 100 more in the United States. Faced with identical meals, the coefficient of variation between people was 68% for glucose, 59% for insulin and 103% for triglycerides. In 230 pairs of twins (183 of them identical), genes explained 30% of the variation in glucose, 9% in insulin and 4% in triglycerides; the authors consider an algorithm based on DNA alone unlikely to work. A nuance that usually gets left out: for glucose, the composition of the meal mattered more (15.4% of the variation) than the gut microbiome (6.0%). Several authors worked for ZOE, which sells a program based on this data.

Personalized diet versus standard advice: the trials

The fact that we respond differently does not prove that a personalized plan works better. That is what trials comparing the two can tell you:

• Food4Me (Celis-Morales and colleagues, International Journal of Epidemiology, 2017): adults from 7 European countries received general or personalized advice online; 1,269 completed 6 months. With personalized advice they ate less red meat (about 5.5 g), salt (0.65 g) and saturated fat (1.1 points less of total energy). But adding blood tests, or those plus 5 genetic variants, did not improve the result.
• ZOE METHOD (Bermingham and colleagues, Nature Medicine, 2024): 347 US adults aged 41 to 70 (86% women) followed for 18 weeks either ZOE’s program in its app or the official dietary guidelines, delivered with a leaflet and videos. Triglycerides fell slightly more on the program (0.13 mmol/L) and LDL cholesterol did not change. The ZOE group lost 2.46 kg more and cut its intake by about 162 kcal a day more than the control group.

Genetic tests to choose a diet: the trial that checked

Nutrigenetics promises that your DNA tells you whether you will do better cutting fat or cutting carbs. Stanford’s DIETFITS trial (Gardner and colleagues, JAMA, 2018) put it to the test with 609 adults with overweight or obesity, who followed a healthy low-fat or a healthy low-carb diet for 12 months after being classified by a pattern of 3 genetic variants that supposedly showed which one suited them. They lost 5.3 kg on the low-fat diet and 6.0 kg on the low-carb diet, a 0.7 kg difference that was not significant, and genotype did not predict who lost more on each. Together with PREDICT and Food4Me, the conclusion is clear: today there is no basis for choosing your diet with a commercial genetic test.

Glucose sensors without diabetes: what is normal

Continuous glucose monitors (CGMs) were designed for diabetes and are now sold to anyone. In a multicenter study (Shah and colleagues, Journal of Clinical Endocrinology & Metabolism, 2019), 153 healthy people aged 7 to 80 spent a median of 96% of the time between 70 and 140 mg/dL and about 30 minutes a day above 140: going above 140 after some meals is normal. Sensors also do not always agree. At the US National Institutes of Health (Howard and colleagues, American Journal of Clinical Nutrition, 2020), 16 adults without diabetes wore two sensors from different brands at the same time over 760 meals, and when the meals were ranked by glucose rise, the sensors agreed poorly (a correlation of 0.43). The authors consider personalizing meals with a sensor premature. In short:

• With diabetes, or with prediabetes if your doctor recommends it, it is a useful clinical tool.
• Without diabetes, a short experiment can teach you something; a single spike is not a diagnosis.
• Do not decide that a healthy food “does not agree with you” from a single reading.

What you can personalize today without paying for tests

What has the most support needs neither DNA nor sensors: adjusting your plan to what you eat and to how your weight responds. In Food4Me, personalizing based on diet was enough. Concrete steps:

• Work out a starting point with the TDEE calculator and treat it as an estimate.
• Log what you eat for 2-3 weeks and weigh yourself several times a week.
• If your average weight does not move as expected, adjust your calories to your real trend.
• If you have diabetes or kidney disease, or take medication, personalize your plan with your doctor or a registered dietitian.

Summary: what works today and what does not

What the studies cited here say:

AI already makes the most important part easier, knowing what you eat, although the quantity still needs checking. Biological personalization has serious science behind it, but its benefits over good general advice are modest, and genetic tests have not helped people choose a diet. Start with what works: log, compare with your real weight, and adjust.