ChatGPT peut-il calculer combien de calories je dois manger ?

Il donne un chiffre, mais pas une mesure. Sur 291 réponses, aucune IA n'a demandé la masse grasse (une seule a demandé le nombre de pas), et 97 % des réponses au prompt simple reposent sur une grille d'activité déclarative (sédentaire, modéré, actif). Le résultat est une moyenne statistique qui ignore ta composition corporelle et ton activité réelle, avec jusqu'à 1 260 kcal d'écart pour la même personne.

Pourquoi les IA donnent-elles des chiffres si différents pour le même profil ?

Parce qu'elles n'ont pas de mesure. Elles appliquent une formule de population (Mifflin-St Jeor 1990 le plus souvent) puis un multiplicateur d'activité choisi au jugé. Deux hommes de 78 kg, l'un à 12 % de masse grasse et 12 000 pas, l'autre à 28 % et 4 000 pas, reçoivent la même réponse au prompt simple, alors que leur dépense réelle diffère de près de 1 000 kcal.

Est-ce que donner plus de détails à ChatGPT corrige le problème ?

En partie seulement. Avec la masse grasse, les pas et les séances dans le prompt, l'écart entre IA reste de 650 kcal sur un même profil, la masse grasse n'est réellement utilisée dans le calcul que dans 40 % des réponses et les pas dans 42 %. Et surtout, la réponse est figée : elle ne sait pas ce que tu as marché ni mangé aujourd'hui, ni comment ton métabolisme s'adapte au fil des semaines.

Qu'est-ce que l'adaptation métabolique et pourquoi ça change la réponse ?

En déficit prolongé, le métabolisme de base baisse au-delà de ce que la perte de poids explique. Un chiffre juste aujourd'hui devient trop haut au bout de quelques semaines. Seules 14 % des réponses détaillées l'ont mentionné, et aucune ne la calcule. C'est la cause la plus fréquente des plateaux.

Comment calculer correctement ses calories pour maigrir ?

En reconstruisant la dépense brique par brique, TDEE = BMR + NEAT + EAT + TEF, avec le métabolisme de base calculé sur la masse maigre réelle, la NEAT sur les pas mesurés, l'EAT sur les séances réellement faites, le TEF sur les macros ingérés, puis en suivant l'évolution jour après jour et en corrigeant quand le corps s'adapte. C'est un suivi, pas une réponse en une phrase.

L'IA est-elle inutile pour la nutrition ?

Non. L'IA est précieuse quand elle mesure : estimer une masse grasse sur une photo, reconnaître un plat, lire un code-barres. Elle est faible quand elle devine à partir de trois chiffres. La différence est entre une IA branchée sur des mesures mises à jour en continu et un chatbot qui répond de mémoire à un inconnu.

Puis-je reproduire l'étude ?

Oui. Les 12 prompts, les réponses brutes horodatées, l'extraction structurée et les scripts sont publiés en données ouvertes (lien en bas de page). Chaque appel est une conversation neuve, sans instruction système, avec les paramètres par défaut.

Study · Losing weight with AI

Does losing weight with ChatGPT work? We asked 5 AIs 291 times

Same person, same question, up to 1,260 kcal of difference. Here is what the AIs answer, and what they cannot do.

Independent study by The Lean Team, the app that recomputes your expenditure every day from your measurements · September 9, 2026 · Open data

1,260 kcalof difference for the same profile, simple prompt
0 %ask for body fat
97 %answer with a “sedentary / active” grid
14 %mention metabolic adaptation

Summary

The study in 30 seconds

1291 answers from ChatGPT, Gemini and Claude to the question “how many calories to lose weight?”, for 6 identical profiles.
2Simple prompt (age, weight, height): the same person is told from 1,500 to 2,700 kcal depending on the AI and the moment.
30 % of answers ask for body fat. 97 % return a “sedentary / moderate / active” grid, the 1990 formula.
4Full prompt (body fat, steps, sessions): still 650 kcal of difference, and body fat is only used in 40 % of the calculations.
5No AI follows up the next day: steps of the day, meals, thermic effect, metabolic adaptation (14 % mention it). The number is frozen on day 1.

Verdict. It can work by luck, not by method: without measurement or daily updates, the answer is a population average frozen on the first day.

01 · Method

What we did

6identical profiles, from the athletic woman to the sedentary man
2prompt levels: simple, then full
5consumer AIs, fresh conversation every time
291timestamped answers, published as open data

ChatGPT (GPT-5.5 and GPT-5.4 mini), Gemini (3.7 Flash and 3.1 Pro), Claude Opus. French and English, 2 to 3 repetitions, no instructions at all. The black line on the charts: the target computed component by component on measured data (BMR on lean mass, steps, sessions, TEF, 500 kcal deficit).

02 · Result 1

Simple prompt: from 1,500 to 2,700 kcal for the same person

“I'm a 32-year-old man, 78 kg, 180 cm, I want to lose weight, how many calories a day?” The question as millions of people ask it.

P1 · Man, 32, 78 kg, 180 cm, 12% BF, 12,000 steps, 4 strength sessions/wk

1,000 kcal3,000 kcal

ChatGPT (GPT-5.5)

1,500 to 2,700

ChatGPT (GPT-5.4 mini)

1,500 to 2,450

Gemini 3.7 Flash

1,600 to 2,650

Gemini 3.1 Pro

1,600 to 2,700

Claude Opus

1,500 to 2,600

All AIs combined: 1,500 to 2,700 kcal, i.e. 1,200 kcal spread (28 answers). Black line: target computed component by component on measured data (2,360 kcal).

P2 · Man, 32, 78 kg, 180 cm, 28% BF, 4,000 steps, no sport

1,000 kcal3,000 kcal

ChatGPT (GPT-5.5)

1,600 to 2,700

ChatGPT (GPT-5.4 mini)

1,600 to 2,760

Gemini 3.7 Flash

1,600 to 2,600

Gemini 3.1 Pro

1,600 to 2,700

Claude Opus

1,500 to 2,500

All AIs combined: 1,500 to 2,760 kcal, i.e. 1,260 kcal spread (28 answers). Black line: target computed component by component on measured data (1,390 kcal).

P3 · Woman, 28, 62 kg, 165 cm, 22% BF, 9,000 steps, 3 fitness sessions/wk

1,000 kcal3,000 kcal

ChatGPT (GPT-5.5)

1,200 to 2,000

ChatGPT (GPT-5.4 mini)

1,130 to 1,820

Gemini 3.7 Flash

1,300 to 1,750

Gemini 3.1 Pro

1,200 to 1,800

Claude Opus

1,200 to 1,900

All AIs combined: 1,130 to 2,000 kcal, i.e. 870 kcal spread (25 answers). Black line: target computed component by component on measured data (1,499 kcal).

P4 · Woman, 45, 70 kg, 165 cm, 35% BF, 3,500 steps, no sport

1,000 kcal3,000 kcal

ChatGPT (GPT-5.5)

1,200 to 2,000

ChatGPT (GPT-5.4 mini)

1,100 to 1,850

Gemini 3.7 Flash

1,250 to 1,850

Claude Opus

1,200 to 1,800

All AIs combined: 1,100 to 2,000 kcal, i.e. 900 kcal spread (22 answers). Black line: target computed component by component on measured data (1,082 kcal).

P5 · Man, 50, 95 kg, 178 cm, 30% BF, 6,000 steps, 2 walks/wk

1,000 kcal3,000 kcal

ChatGPT (GPT-5.5)

1,500 to 2,400

ChatGPT (GPT-5.4 mini)

1,680 to 2,550

Gemini 3.7 Flash

1,650 to 2,300

Claude Opus

1,500 to 2,300

All AIs combined: 1,500 to 2,550 kcal, i.e. 1,050 kcal spread (22 answers). Black line: target computed component by component on measured data (1,918 kcal).

P6 · Woman, 24, 55 kg, 170 cm, 18% BF, 14,000 steps, 5 runs/wk

1,000 kcal3,000 kcal

ChatGPT (GPT-5.5)

1,200 to 1,900

ChatGPT (GPT-5.4 mini)

1,200 to 1,850

Gemini 3.7 Flash

1,600 to 2,200

All AIs combined: 1,200 to 2,200 kcal, i.e. 1,000 kcal spread (16 answers). Black line: target computed component by component on measured data (1,735 kcal).

AI behaviour, simple prompt (147 answers)

0 %ask for body fat percentage
1 %ask for daily step count
97 %use a declarative activity grid
48 %cite Mifflin-St Jeor or Harris-Benedict (1990, 1919)
0 %mention the thermic effect of food

P1 and P2 asked exactly the same question (same age, weight, height) and got the same grid. Yet their real expenditure differs by nearly 1,000 kcal a day: 12% body fat and 12,000 steps on one side, 28% and 4,000 steps on the other.

03 · Result 2

Full prompt: still a 650 kcal gap

We give everything: body fat, steps per day, sessions, and “give me a precise number”. All comply. A precise number is not a correct number.

P1 · Man, 32, 78 kg, 180 cm, 12% BF, 12,000 steps, 4 strength sessions/wk

1,000 kcal3,000 kcal

ChatGPT (GPT-5.5)

2,350 to 2,750

ChatGPT (GPT-5.4 mini)

2,300 to 2,500

Gemini 3.7 Flash

2,100 to 2,400

Gemini 3.1 Pro

2,300 to 2,700

Claude Opus

2,250 to 2,450

All AIs combined: 2,100 to 2,750 kcal, i.e. 650 kcal spread (28 answers). Black line: target computed component by component on measured data (2,360 kcal).

P2 · Man, 32, 78 kg, 180 cm, 28% BF, 4,000 steps, no sport

1,000 kcal3,000 kcal

ChatGPT (GPT-5.5)

1,600 to 1,800

ChatGPT (GPT-5.4 mini)

1,750 to 2,000

Gemini 3.7 Flash

1,550 to 1,750

Gemini 3.1 Pro

1,400 to 1,800

Claude Opus

1,650 to 1,950

All AIs combined: 1,400 to 2,000 kcal, i.e. 600 kcal spread (28 answers). Black line: target computed component by component on measured data (1,390 kcal).

P3 · Woman, 28, 62 kg, 165 cm, 22% BF, 9,000 steps, 3 fitness sessions/wk

1,000 kcal3,000 kcal

ChatGPT (GPT-5.5)

1,600 to 1,850

ChatGPT (GPT-5.4 mini)

1,700 to 1,850

Gemini 3.7 Flash

1,500 to 1,750

Claude Opus

1,450 to 1,900

All AIs combined: 1,450 to 1,900 kcal, i.e. 450 kcal spread (22 answers). Black line: target computed component by component on measured data (1,499 kcal).

P4 · Woman, 45, 70 kg, 165 cm, 35% BF, 3,500 steps, no sport

1,000 kcal3,000 kcal

ChatGPT (GPT-5.5)

1,200 to 1,450

ChatGPT (GPT-5.4 mini)

1,250 to 1,500

Gemini 3.7 Flash

1,300 to 1,400

Claude Opus

1,200 to 1,600

All AIs combined: 1,200 to 1,600 kcal, i.e. 400 kcal spread (22 answers). Black line: target computed component by component on measured data (1,082 kcal).

P5 · Man, 50, 95 kg, 178 cm, 30% BF, 6,000 steps, 2 walks/wk

1,000 kcal3,000 kcal

ChatGPT (GPT-5.5)

1,850 to 2,200

ChatGPT (GPT-5.4 mini)

1,900 to 2,050

Gemini 3.7 Flash

1,850 to 1,980

Claude Opus

1,600 to 2,100

All AIs combined: 1,600 to 2,200 kcal, i.e. 600 kcal spread (22 answers). Black line: target computed component by component on measured data (1,918 kcal).

P6 · Woman, 24, 55 kg, 170 cm, 18% BF, 14,000 steps, 5 runs/wk

1,000 kcal3,000 kcal

ChatGPT (GPT-5.5)

1,900 to 2,300

ChatGPT (GPT-5.4 mini)

1,850 to 2,200

Gemini 3.7 Flash

2,300 to 2,500

Claude Opus

1,900 to 2,400

All AIs combined: 1,850 to 2,500 kcal, i.e. 650 kcal spread (19 answers). Black line: target computed component by component on measured data (1,735 kcal).

AI behaviour, full prompt (144 answers)

40 %actually use the provided body fat
42 %convert the provided steps into calories
14 %mention metabolic adaptation
3 %mention the thermic effect of food
97 %give a single number, as requested

The P1 vs P2 test: same weight, different bodies

699 kcalof difference between P1 and P2 according to the AIs (average)
970 kcalof difference when computing component by component
+324 kcaltoo much for the sedentary profile at 28% body fat

The AIs see the difference between the two men, but underestimate it: they adjust a population formula instead of measuring. The sedentary profile, the one who “does everything right”, gets a target that is too high and loses nothing.

04 · The real issue

The problem is not day 1, it is the 89 days that follow

78% of detailed answers advise to “weigh yourself and adjust”. None does it for you.

Day 1The chatbot receives a weight and answers once.
Day 2You walked 4,000 steps instead of 11,000. It does not know.
Week 3You did 6 sessions and changed your macros. It does not know.
Day 90You lost 4 kg, your metabolism dropped then adapted. It still does not know.

What the chatbot does not know

✕Your steps of the day
✕Your meals of the day and their thermic effect
✕Your real body fat, and how it changes
✕How your metabolism adapts over the weeks
✕No curve, no chart of your expenditure

What real tracking requires

✓Expenditure recomputed every day from measurements
✓A target that moves with steps and sessions
✓Body fat re-measured regularly
✓Automatic correction of metabolic adaptation
✓A readable curve of your TDEE and its evolution

Verdict: does losing weight with ChatGPT work? By luck, sometimes. By method, no: without measurement or daily updates, the answer remains a population estimate frozen on the first day. The test can be redone in five minutes with your own profile.

05 · And Lean?

The AI that measures versus the AI that guesses

Lean also uses AI, but to measure: BodyScan for body fat, photo scan for the plate, pedometer for steps. Then it recomputes the five components of your expenditure every day and plots the curve.

BMR on measured lean massNEAT on real stepsEAT per sessionTEF from macrosMetabolic adaptation

Lean measures, recomputes and tracks, every dayFirst app to compute metabolic adaptation. Free download, 7-day free trial on the annual subscription. 4.7/5 on the App Store, more than 10,000 users.

Data

The data, profile by profile

P1 · Man, 32, 78 kg, 180 cm, 12% BF, 12,000 steps, 4 strength sessions/wk
Simple prompt: 1,500 to 2,700 kcal (28 answers) · Full prompt: 2,100 to 2,750 kcal (28 answers) · Component-by-component target: 2,360 kcal (BMR 1,853, NEAT 504, EAT 267, TEF 236, TDEE 2,860)
P2 · Man, 32, 78 kg, 180 cm, 28% BF, 4,000 steps, no sport
Simple prompt: 1,500 to 2,760 kcal (28 answers) · Full prompt: 1,400 to 2,000 kcal (28 answers) · Component-by-component target: 1,390 kcal (BMR 1,583, NEAT 168, EAT 0, TEF 139, TDEE 1,890)
P3 · Woman, 28, 62 kg, 165 cm, 22% BF, 9,000 steps, 3 fitness sessions/wk
Simple prompt: 1,130 to 2,000 kcal (25 answers) · Full prompt: 1,450 to 1,900 kcal (22 answers) · Component-by-component target: 1,499 kcal (BMR 1,415, NEAT 275, EAT 159, TEF 150, TDEE 1,999)
P4 · Woman, 45, 70 kg, 165 cm, 35% BF, 3,500 steps, no sport
Simple prompt: 1,100 to 2,000 kcal (22 answers) · Full prompt: 1,200 to 1,600 kcal (22 answers) · Component-by-component target: 1,082 kcal (BMR 1,353, NEAT 121, EAT 0, TEF 108, TDEE 1,582)
P5 · Man, 50, 95 kg, 178 cm, 30% BF, 6,000 steps, 2 walks/wk
Simple prompt: 1,500 to 2,550 kcal (22 answers) · Full prompt: 1,600 to 2,200 kcal (22 answers) · Component-by-component target: 1,918 kcal (BMR 1,806, NEAT 303, EAT 117, TEF 192, TDEE 2,418)
P6 · Woman, 24, 55 kg, 170 cm, 18% BF, 14,000 steps, 5 runs/wk
Simple prompt: 1,200 to 2,200 kcal (16 answers) · Full prompt: 1,850 to 2,500 kcal (19 answers) · Component-by-component target: 1,735 kcal (BMR 1,344, NEAT 391, EAT 326, TEF 174, TDEE 2,235)

Variability of the same AI on the same repeated prompt (full prompt, average gap between repetitions): Claude Opus 73 kcal, Gemini 3.1 Pro 175 kcal, Gemini 3.7 Flash 57 kcal, ChatGPT (GPT-5.4 mini) 98 kcal, ChatGPT (GPT-5.5) 96 kcal.

Full dataset (prompts, 291 raw answers, extraction, scripts): Zenodo, DOI 10.5281/zenodo.22662006, CC BY 4.0 licence.

Detailed method and limitations

Models queried by API with default parameters and no system instruction: GPT-5.5 and GPT-5.4 mini (OpenAI), Gemini 3.7 Flash and Gemini 3.1 Pro (Google). Claude Opus (Anthropic) queried through the official command-line interface with a neutral system prompt and no context. Six profiles, two prompt levels, French and English, 2 to 3 repetitions per combination, 291 answers collected on September 8, 2026 (Gemini 3.1 Pro: 27 answers, API rate limited; other models: 48 to 72).

Reading the answers: an automatic extractor (GPT-5.4 mini, temperature 0, strict JSON schema) identifies the recommended intake target (min and max when a range or a grid is given) and the behaviours (question asked before answering, activity grid, actual use of the provided body fat and steps, mention of metabolic adaptation and of the thermic effect). The extraction is published with the raw answers, line by line.

Limitations: fictional profiles; public reference formulas (Katch-McArdle on lean mass, Ainsworth 2011 compendium, TEF at 10%) that serve as a benchmark, not ground truth; models evolve and answers change from one week to the next, which is part of the finding. Lean publishes this study and sells a tracking app: method and data are open so that anyone can redo it.

Scientific sources

Mifflin MD et al. A new predictive equation for resting energy expenditure in healthy individuals. Am J Clin Nutr. 1990;51(2):241-7. PubMed 2305711
Ainsworth BE et al. 2011 Compendium of Physical Activities. Med Sci Sports Exerc. 2011;43(8):1575-81. PubMed 21681120
Levine JA et al. Interindividual variation in posture allocation. Science. 2005;307(5709):584-6. PubMed 15681386
Trexler ET et al. Metabolic adaptation to weight loss: implications for the athlete. J Int Soc Sports Nutr. 2014;11:7. PubMed 24571926
Westerterp KR. Diet induced thermogenesis. Nutr Metab. 2004;1:5. PubMed 15507147

Study published by The Lean Team. Press contact: franckbarriere@lean-app.com. Not medical advice.

Frequently asked questions

Can ChatGPT calculate how many calories I should eat?

It gives a number, not a measurement. Out of 291 answers, no AI asked for body fat (a single one asked for the step count), and 97% of answers to the simple prompt rely on a declarative activity grid (sedentary, moderate, active). The result is a statistical average that ignores your body composition and your real activity, with up to 1,260 kcal of difference for the same person.

Why do AIs give such different numbers for the same profile?

Because they have no measurement. They apply a population formula (most often Mifflin-St Jeor 1990) then an activity multiplier picked by guesswork. Two 78 kg men, one at 12% body fat and 12,000 steps, the other at 28% and 4,000 steps, get the same answer to the simple prompt, while their real expenditure differs by nearly 1,000 kcal.

Does giving ChatGPT more details fix the problem?

Only partly. With body fat, steps and sessions in the prompt, the gap between AIs remains 650 kcal on the same profile, body fat is actually used in the calculation in only 40% of answers and steps in 42%. Above all, the answer is frozen: it does not know how much you walked or ate today, nor how your metabolism adapts over the weeks.

What is metabolic adaptation and why does it change the answer?

In a prolonged deficit, basal metabolic rate drops beyond what the weight loss explains. A number that is right today becomes too high after a few weeks. Only 14% of detailed answers mentioned it, and none computes it. It is the most common cause of plateaus.

How do you correctly calculate your calories to lose weight?

By rebuilding expenditure component by component, TDEE = BMR + NEAT + EAT + TEF, with basal metabolic rate computed on real lean mass, NEAT on measured steps, EAT on sessions actually done, TEF on macros eaten, then tracking the evolution day after day and correcting when the body adapts. It is tracking, not a one-sentence answer.

Is AI useless for nutrition?

No. AI is valuable when it measures: estimating body fat from a photo, recognising a meal, reading a barcode. It is weak when it guesses from three numbers. The difference is between an AI plugged into continuously updated measurements and a chatbot answering a stranger from memory.

Can I reproduce the study?

Yes. The 12 prompts, the timestamped raw answers, the structured extraction and the scripts are published as open data (link at the bottom of the page). Each call is a fresh conversation, with no system instruction, with default parameters.

Does losing weight with ChatGPT work? We asked 5 AIs 291 times

The study in 30 seconds

What we did

Simple prompt: from 1,500 to 2,700 kcal for the same person

Full prompt: still a 650 kcal gap

The problem is not day 1, it is the 89 days that follow

The AI that measures versus the AI that guesses

The data, profile by profile

Frequently asked questions

Read also