AI diet plans match on quality but differ on water footprint, study finds

A new study comparing AI-generated diets found that the plans scored similarly on diet quality but differed in environmental costs, and that the AI rarely gave the same answer twice. 

When researchers asked ChatGPT (5.5 Instant), Claude (Opus 4.8), and Gemini (3 Flash) to produce meal plans aligned with the EAT-Lancet diet, energy-adjusted diet quality scores were comparable across platforms, while water footprints of the suggested foods differed by up to 15%. 

Nutrition Insight speaks with the paper’s co-author, Başak Öney, an associate professor in the Department of Nutrition and Dietetics at Bezmialem Vakıf University in Türkiye, who says that a single AI-generated plan should not be treated as a prescription but as a draft. She calls for sustainability to be integrated into AI systems rather than left up to users to request.

In the Frontiers in Sustainable Food Systems publication, the researchers call for validation, consistency testing, and clinical safety checks before digital tools are used to provide dietary advice.

The researchers generated 45 meal plans in total — three per platform for each of five clinical scenarios (hypertension, high cholesterol, type 2 diabetes, obesity, and general healthy eating) — with each prompt submitted in a separate session.

Water footprint gap drivers

All three platforms scored comparably on the energy-adjusted EAT-Lancet Diet Index 2.0, the study reveals, but Öney explains what drove the gap in water footprints.

“Gemini had the highest mean score at 37.9 out of 45, followed by Claude at 36.8 and ChatGPT at 36.3,” she says. “Water footprints, however, differed clearly.”

She details that water footprints were approximately 2.94 m3/day for ChatGPT, 2.67 m3/day for Claude, and 2.52 m3/day for Gemini. The contrast between ChatGPT and Gemini is explained by two factors:

“The first was quantity: ChatGPT’s plans came closest to the 2,000 kcal target, while Gemini’s were about 12% below it, and more food generally means a larger footprint. The second was food choice, as the gap persisted even after adjusting for energy,” says Öney.

“Almonds illustrate this well. They have a low carbon footprint but a very high water footprint and appeared 33 times in ChatGPT’s plans, 20 times in Claude’s, and 13 times in Gemini’s — a pattern that mirrored the water footprint ranking. Gemini combined the highest diet index score with the lowest footprints, but this was partly because it provided less energy than requested.”

Navigating variations in answers

Öney notes that the variation between AI answers to the same prompt was substantial.

Başak Öney says any AI-generated diet plan should be treated as a draft, not a prescription.“When the same prompt was used three times, diet scores varied by an average of 2.4 to 3.1 points between runs on a 45-point scale, with differences between two runs reaching 5 to 8 points,” she details.

“Carbon footprints varied by about 0.4 to 0.5 kg CO2-eq per day, and water footprints by 0.25 to 0.33 m3. These run-to-run differences were often as large as those between platforms.”

For consumers, she says this means that one AI-generated plan is just an example from a range of possible answers.

“You cannot know whether you received a stronger or weaker version, so I would treat any single output as a draft, not a prescription. For researchers, studies based on a single output can be misleading; repeated generation should be standard practice.”

What plans mean for health

Öney points out that the scenarios specified only the medical condition and energy level, since the study evaluated them based on their adherence to general dietary patterns. Individual patient safety was not assessed, she adds. 

“Plans for hypertension and high cholesterol achieved the highest nutritional scores and the lowest carbon footprints, whereas those for type 2 diabetes, obesity, and general healthy eating yielded lower scores and larger carbon footprints.” 

“Most notably, in the hypertension scenario, the plans generated by ChatGPT contained significantly higher sodium levels than those generated by Claude or Gemini. However, as this finding is based on only three plans per platform, it warrants further verification — though similar observations have been reported in other studies.”

Better AI models

According to Öney, AI tools should account for sustainability by default, since even prompts asking for EAT-Lancet-aligned plans produced varied environmental footprints.

According to Öney, AI tools should account for sustainability by default.“The diet score was not even associated with the water footprint. A plan that looks healthy and sustainable on paper may therefore not be the lowest-impact option. If sustainability depends on users remembering to ask for it, most will never get it,” she says.

“Consumers should explicitly request a plant-forward plan aligned with the EAT-Lancet diet. Check whether it actually meets your requested calorie target and how much red meat, cheese, and nuts it includes. Dietitians should treat AI output as a trainee’s draft.”

Öney advises dietitians to check portions, energy, sodium, and clinical targets, while estimating the environmental footprint where possible. She also recommends generating several plans and comparing the versions.

What must be addressed?

At a minimum, Öney calls for meeting five requirements before AI models are used in diet planning.

“First, validation against national dietary guidelines and food composition databases — in Türkiye, Tüber, and TürKomp. Second, accurate energy content, since even a 10% shortfall can alter the entire plan. Third, consistency: the same input should generate clinically equivalent plans, verified through repeated testing.” 

“Fourth, safety checks for conditions in which errors have significant consequences, including sodium in hypertension, carbohydrates in diabetes, and phosphorus and potassium in kidney disease, with both shortfalls and excesses flagged. Fifth, transparency about model versions, as findings from one version may not apply to the next,” she details. 

Öney also calls for environmental indicators to be reported alongside nutritional ones, with independent evaluation and dietitians involved in development and supervision

“Users should always be informed that the plan does not replace an individual assessment by a health professional,” she concludes.