Algorithm-Based Meal Recommendations

Trial-tested algorithms

A model built from clinical and microbiome features to predict glucose responses guided a diet that, in 225 adults with prediabetes, reduced time above 140 mg/dL and HbA1c somewhat more than a Mediterranean diet. [1] In 23 adults with newly diagnosed type 2 diabetes, a two-week crossover pilot showed lower glucose measures on that diet. [2] The studies are summarized on Personalized Glycemic Predictions. Label: promising but limited, for those populations, using glucose measures rather than clinical outcomes.

A personalized program that scored foods using glucose and triglyceride responses, microbiomes and health history lowered triglycerides more than standard advice in 347 adults, with no significant LDL cholesterol difference. [3] Another trial of cluster-based personalized plans, 100 adults over 10 weeks, found no difference in fat mass compared with a generic healthy diet. [4] Label: mixed.

A review of randomized trials in healthy adults concluded that personalized advice improved dietary intake more than generalized advice, and called for better-designed trials comparing types of personalization. [5] Food4Me found improvements in diet from internet-delivered personalized advice, with no extra gain from adding blood or genetic information. [6] Both measure what people eat, not disease.

Chatbot-written meal plans

• In a study of ChatGPT and Bard plans (108 plans) across omnivorous, vegetarian and vegan patterns, plans on average had less energy and carbohydrate than reference intakes, vitamin D and fluoride fell short in every plan, and only vegan plans lacked enough vitamin B12. [7]
• A comparison of three chatbots for weight-loss plans found good overall diet-quality scores, but macronutrient and fatty-acid balance scores were lowest, and one chatbot missed calorie targets by more than 20% in over half its plans. [8]
• In a 2026 study, six language models wrote 1,800 kcal plans for type 2 diabetes; most systematically deviated from a dietitian-designed reference, with lower energy, carbohydrate and fiber, and adherence varied widely. [9] The authors concluded the plans are not interchangeable with dietitian-guided care and may be risky without oversight.
• Chatbot estimates of nutrition-label values for eight ready-to-eat meals were roughly 70 to 90% accurate for calories and macronutrients, with sodium and saturated fat underestimated. [10]
• In an exploratory test, two chatbots wrote gluten-free plans for a hypothetical woman with celiac disease; one showed calcium and iron shortfalls, and outputs varied when repeated a week later. [11]

These are cross-sectional tests of plan content with hypothetical prompts. None followed people eating the plans, so health effects are unknown. Label: insufficient evidence for health benefit; evidence of inaccuracy for some nutrients in specific tests.

How to judge a claim

Ask what the algorithm was tested on, who the participants were, whether a trial compared it with ordinary advice, and which outcome was measured. A feature list or testimonial is a marketing claim. See Research Hype and Press Releases.