AI is being tested in nutrition mainly as a way to predict how a particular person's body responds to a meal, not as a source of general eating advice. Two research programs show how far that work has come: a 2015 Israeli study of 800 people that predicted blood sugar responses with a machine learning model, and a United States National Institutes of Health (NIH) program that registered a much larger trial to build similar models for glucose, insulin, triglycerides, and GLP-1. The second program is still recruiting as of October 2026.
Key takeaways
- A 2015 study in Cell measured responses to 46,898 meals in 800 people and found large differences in blood sugar after identical meals.
- The NIH Nutrition for Precision Health program is registered with 8,000 people in its first module, 1,200 in a home-based diet crossover, and 150 in a supervised residential study.
- The registry lists primary completion as estimated for February 25, 2027, so final results were not available at the last check.
- The program also tests whether tools such as food photos and wearable sensors measure intake accurately.
- Predicting a blood marker is a different claim from showing that people who follow the predictions stay healthier for years.
The 2015 study that started the idea
A team at the Weizmann Institute of Science in Israel, led by Eran Segal and Eran Elinav, published its work in Cell in November 2015. Participants wore continuous glucose monitors for a week, and the researchers recorded responses to 46,898 meals across an 800-person cohort. The central finding was variability: identical meals produced very different blood sugar responses in different people, which the authors took as a sign that one-size-fits-all dietary recommendations may have limited use.
The team then built a machine learning algorithm that combined blood parameters, dietary habits, body measurements, physical activity, and gut microbiome data. It predicted responses to real-life meals, and the predictions held in a separate 100-person cohort. A blinded randomized dietary intervention based on the algorithm produced significantly lower post-meal responses. The study measured blood sugar after meals, not diagnoses or long-term outcomes, and the participants came from one country.
The NIH program built on that question
Nutrition for Precision Health (NPH) runs inside the NIH All of Us Research Program and is managed by the NIH Common Fund. NIH announced awards of USD 170 million over five years, pending the availability of funds. The study record on US federal trial registry, with RTI International as lead sponsor, states the goal as developing AI and machine learning algorithms that predict individual responses to diet patterns from behavioral, social, environmental, clinical, and molecular data.
The design has three layers. Module 1 is a phenotyping cohort of 8,000 people. A subset forms Module 2, where 1,200 participants eat three distinct diets in a 14-day crossover sequence at home, with at least 14 days of washout between diets. Another subset forms Module 3, where 150 participants receive the same three diets under full-time supervision in a residential setting. After each diet, participants take a standardized meal test, and the primary outcomes are the zero- to four-hour response curves for glucose, insulin, triglycerides, and GLP-1.
Comparing the two studies
| Feature | Weizmann study (2015) | NIH NPH program (registry, October 2026) |
|---|---|---|
| Size | 800 people, plus 100 for validation | 8,000 (Module 1), 1,200 (Module 2), 150 (Module 3) |
| Outcome predicted | Blood glucose after meals | Glucose, insulin, triglycerides, and GLP-1 after meals |
| Diet control | Real-life meals, then a dietary intervention | Three assigned diets, at home and under supervision |
| Status | Published in Cell, 2015 | Recruiting; primary completion estimated February 25, 2027 |
Can a phone or a sensor measure what was eaten?
Predictions are only as good as the diet data behind them. The NPH registry therefore lists a secondary objective: testing the validity and acceptability of technology-based dietary assessment tools. These include the Automated Self-Administered 24-hour recall (ASA24), the Automatic Ingestion Monitor-2 (AIM-2), and the mobile food record (mFR). Total energy, macronutrient, and fiber intake reported by each tool is compared with measures of actual intake, which is possible because the supervised module controls the food served. Acceptability is judged from participant feedback surveys. Results for these tools were not published at the last check, so claims that a photo app can reliably count calories should be treated as unproven until such comparisons appear.
What these studies do not show
Neither program shows that following an algorithm's meal plan prevents diabetes or heart disease. The 2015 trial measured short-term blood sugar, and the NIH design measures blood responses over hours after a test meal. Both cohorts also have limits in who is represented. The Weizmann group drew from a single country, and the NIH program recruits from All of Us participants, so findings need testing in other populations, including Canadian ones. Commercial apps that advertise personalized nutrition rarely publish comparable data, and this article does not evaluate any of them.
This article offers general information about research and is not medical or dietary advice. Questions about diet, diabetes, or medication belong with a physician or a registered dietitian.
Questions to ask about an AI nutrition claim
- What was predicted: a blood marker after a meal, weight, or a disease outcome?
- How many people were studied, and were predictions tested on people the model had not seen?
- Was food intake measured by a controlled method or by self-report?
- Were the results peer reviewed or registered in a public trial database?
- Who funded the work, and does the publisher sell the product being described?
FAQ
Does AI already give accurate personalized diet plans?
Only in research settings, and only for narrow outcomes. The 2015 Cell study predicted post-meal blood sugar in a defined cohort, and the NIH program is still testing whether broader models work.
What is the NIH Nutrition for Precision Health program?
It is a research program within All of Us that feeds data from three modules of participants into machine learning models. Its registry record says the models target post-meal curves for glucose, insulin, triglycerides, and GLP-1.
Why do identical meals affect people differently?
The 2015 study linked differences to factors such as blood parameters, habits, activity, and gut microbiota. The research does not give a complete explanation for any one person.
Are results from these studies available for Canadians?
Neither study recruited specifically in Canada. Readers in Canada who want dietary guidance can use national guidance from Health Canada and individual advice from a registered dietitian.
Related reading
- How AI changes the work of data science
- The Deloitte AI Institute and industry applications
- AI in everyday communication
- AI and the future of work
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Zeevi D et al., Personalized Nutrition by Prediction of Glycemic Responses, Cell 163(5), 2015
- Nutrition for Precision Health, Powered by the All of Us Research Program (NCT05701657), record verified October 2026
- NIH awards $170 million for precision nutrition study, National Institutes of Health, news release