# Weigh it, or just say it: the BurnWeek scale and logging from ChatGPT

September 14, 2026 · From the Lab · 9 min read · https://burnweek.fit/blog/smart-scale-and-chatgpt-logging/

> Dietitians keeping weighed records still came in 223 kcal a day short. Weighing kills one error and leaves the other, so the scale has a button you talk to.

**Key takeaways**

- The standalone BurnWeek scale (August 2026) has no camera: hold the button, say the food, and the measured grams become the portion; the phone gets a notification.
- Portion misjudgement is systematic: >20 percent errors on 10 of 17 dishes (Faggiano 1992), small over and large under (Nelson 1996), beverages 30 to 46 percent low (Almiron-Roig 2013).
- Weighing does not fix omission: dietitians keeping 7-day weighed records still underreported by 223 kcal/day against doubly labelled water, non-dietitians by 429 (Champagne 2002).
- A 7-day estimated diary came closest to 16 days of weighed records in 160 women, with no significant difference in average intake (Bingham 1994): item errors partly cancel across a week.
- The BurnWeek ChatGPT app is published in the ChatGPT app directory (version 1.0.0), and any MCP-capable assistant can connect to the same server. The assistant only displays the range BurnWeek returns.

## Portion size is the error a sentence cannot fix, and a scale can

Two ways of getting food into BurnWeek arrived this summer at opposite ends of the precision scale. At one end is a small standalone kitchen scale with a push-to-talk button: put the food on it, hold the button, say what it is, and the meal is logged with the measured grams as the portion, the words as the identity, and a notification on your phone to prove it happened. At the other end is a connector that lets ChatGPT, or any assistant that speaks the Model Context Protocol, log a meal to your account from a sentence typed in a chat you were already having. The scale is for the portions where the number matters. The assistant is for everywhere else.

The reason to have both is that self-reported intake fails in two separable ways, and the fixes do not overlap. One failure is omission: the snack that never gets written down. The other is misjudging the amount of what did get written down. The first is a habit problem, and the [lock-screen logging](/blog/log-from-lock-screen) post is about it. The second is a measurement problem, and it is the one this post is about, because weighing is the only method that removes it outright.

## How badly people judge the amount on the plate

The studies on portion estimation are old, small, and remarkably consistent. In 1992, 103 Italian volunteers were served a dinner of 17 standard dishes, every portion weighed as they chose it, and interviewed the next day with a set of photographs. They overestimated by more than 20 percent for six of the foods and underestimated by more than 20 percent for four, with a pattern the authors named the "flat slope syndrome": people who took small portions guessed high, people who took large portions guessed low ([Faggiano et al., 1992](https://pubmed.ncbi.nlm.nih.gov/1637903/)).

Four years later a London group ran a tighter version. 136 adults served themselves four to six foods at a real meal, had the portions weighed, and were shown eight-step photograph series within five minutes of finishing. Individual estimates varied widely, small portions were overestimated and large ones underestimated again, and butter and margarine were substantially overestimated. Averaged across a meal, and excluding the spreads, the estimated nutrient content came within about ±7 percent of the weighed truth, but adults over 65 overestimated energy and fat by 15 to 20 percent ([Nelson et al., 1996](https://pubmed.ncbi.nlm.nih.gov/8774215/)). The [eyeballing portions](/blog/eyeballing-portions-accuracy) article walks through that study's slope in detail.

The error is not only about size. In a 2013 Cambridge study, 32 healthy-weight adults judged how many "portions" 33 foods and drinks represented. They underestimated overall, made larger errors on single-unit foods and items labelled as a meal or a drink than on multi-unit snacks, and underestimated beverages and medium-energy-density foods by 30 to 46 percent while overestimating high-energy-density items ([Almiron-Roig et al., 2013](https://pubmed.ncbi.nlm.nih.gov/23932948/)). The food itself sets the error. A glass of juice and a slice of cake are misjudged in opposite directions, and no amount of care with the words fixes that.

| Study | What was estimated | Typical error |
| --- | --- | --- |
| Faggiano 1992, n=103 | 17 dishes at a dinner, recalled next day from photos | >20% over on 6 foods, >20% under on 4 |
| Nelson 1996, n=136 | 4–6 self-served foods, photos within 5 minutes | Meal average within ±7%; over-65s 15–20% high |
| Almiron-Roig 2013, n=32 | Number of portions in 33 foods | Beverages and medium-density foods 30–46% low |

A weighed portion has none of these errors. It has others, covered below, but the flat slope, the beverage undershoot, and the age effect all vanish when the number comes from a load cell rather than an eye.

## What the scale does when you hold the button

The scale is a BurnWeek input device, not a separate food log. It has no camera. Voice is the only thing that names the food, and the weight is an accuracy upgrade layered on top of the same estimate a spoken sentence in the app would get.

Pairing takes a six-digit code from Profile → BurnWeek scale → Add scale; the code is single-use and expires in ten minutes. After that, each time the scale wakes up it opens a new meal session, and every hold of the button adds one component to that session's single meal:

- Put the food on the platform. If a stable weight of more than 5 grams has landed since the last hold, that delta is the portion.
- Hold the button and say what it is. Live captions appear on the scale's own screen as you speak, so you can see what it heard before you let go.
- Release. The transcript and the weight are combined, so the estimate is for "180 g of grilled chicken thigh" rather than "some chicken", so identity and cooking state come from your words and the amount comes from the scale.
- Your phone gets a notification with the entry: the food, the portion, the calorie range, the protein. Undo and Edit are on the card.

A hold with nothing on the platform is a valid voice-only log; the portion is then the estimator's guess and the entry says so. Taking food off the scale never changes anything already logged. A hold can be up to thirty seconds long. If a caption is wrong, a retake replaces the previous component rather than adding a second one, and the meal's totals are recomputed. In the app, a scale-logged meal shows a distinct scale source mark, and a weighed component inside a mixed meal carries a small "weighed N g on scale" chip, so a meal that started as a voice log and had a weighed item appended to it still says which number was measured.

The first real weigh-in taught us something. A 225 gram apple came back as 564 calories, a figure no apple can have, because the words had not come through and the weight alone was all there was to go on. Since August 25 your words always decide what the food is, and a result that is physically implausible for any food is rejected: the scale shows that the entry was not logged rather than showing a wrong number.

## Weighing removes one error and leaves the other

It would be convenient to say that a weighed log is an accurate log. The evidence says something more specific. In a 2002 study at Pennington, ten registered dietitians and ten women of similar age and weight kept seven-day weighed food records while their energy expenditure was measured by doubly labelled water. The non-dietitians underreported by 429 kcal a day. The dietitians, who weigh food for a living, still came in 223 kcal a day low, a gap that did not reach significance but was not zero ([Champagne et al., 2002](https://pubmed.ncbi.nlm.nih.gov/12396160/)). Weighing what you record does not make you record everything.

And the reverse is also true: a diligent estimated diary gets closer to weighed records than the portion literature might suggest. In a 1994 British comparison, 160 women weighed their food for 16 days across a year using electronic scales, and among seven other methods, a seven-day open-ended estimated diary was the one whose individual values came closest to the weighed record, with no significant difference in average intake ([Bingham et al., 1994](https://pubmed.ncbi.nlm.nih.gov/7986792/)). Estimation errors on individual foods partly cancel across a week; omissions do not.

So the scale earns its place on the foods where the per-item error is large and systematic: the beverages Almiron-Roig's participants undershot by a third, the spreads Nelson's overestimated, the single dense food that is most of a meal's calories. A range around a weighed 180 grams of chicken is narrow because the only thing left to estimate is the chicken. A range around "a bowl of pasta" is wide because the bowl is the question. The [why calorie counts are ranges](/blog/why-calorie-counts-are-ranges) piece explains why the app keeps showing the width even when the input is precise.

## Logging from ChatGPT, or from anything that speaks MCP

The other release is a connector. BurnWeek runs a Model Context Protocol server, and the ChatGPT app built on it is published in the ChatGPT app directory as version 1.0.0, so it is available in ChatGPT today. Any assistant that can connect to an MCP server can use the same connector too. Sign-in is a six-digit code sent to your email; the first sign-in creates your account, and meals logged from the assistant appear in the phone app.

The assistant's tools cover logging a meal from a sentence, showing today or the week, undoing the last meal, adjusting a component's grams, and working with recipes and plans. What none of them do is estimate calories: every meal logged this way goes to the same estimator the app uses, and the assistant renders the result as a card with the calorie range, the protein, and the day's total, plus Undo, Edit grams, and See today. The assistant is a keyboard and a display; the numbers come from one place.

The reason this belongs in the same post as the scale is the same reason the scale exists. Portion error is a property of the input, not of the channel. "A chicken burrito and a coke" typed into a chat gets the same wide-ish range as the same words spoken into the phone, and the same 30 to 46 percent beverage undershoot risk that the Cambridge participants showed applies to the coke. The connector removes the app switch, which is a habit cost. It does not, and does not claim to, remove the portion uncertainty. If you want that number narrow, put the burrito on the scale. Either way the log is the same one-sentence act of attention the [mindful eating with numbers](/blog/mindful-eating-with-numbers) pillar argues for; the scale just tells you how much of the sentence was measured.

## FAQ

### Does the BurnWeek scale need a photo of the food?

No. The scale has no camera. You hold the button and say what the food is; the scale measures how much of it there is. Identity and cooking state come from your words, the portion from the weight.

### What happens if I hold the scale's button with nothing on it?

It logs a voice-only entry. The portion is then estimated from your words, exactly as an in-app voice log would be, and the entry is marked as estimated rather than weighed.

### Can ChatGPT estimate my calories on its own?

Not through this connector. ChatGPT sends the text of your meal to the app's own estimator, which returns a range; the assistant only displays it. That is true of the published ChatGPT app and of any other MCP-capable assistant on the same connector.

### If I weigh everything, will my calorie log finally be exact?

No. Weighing removes the portion-size error, but it does not catch the meals you never log. Even dietitians keeping seven-day weighed records came in a couple of hundred calories a day short against measured expenditure. A weighed entry gets a narrower range, not a single number.

## Sources

- [Faggiano F, et al. Validation of a method for the estimation of food portion size. Epidemiology. 1992.](https://pubmed.ncbi.nlm.nih.gov/1637903/)
- [Nelson M, Atkinson M, Darbyshire S. Food photography II: use of food photographs for estimating portion size and the nutrient content of meals. Br J Nutr. 1996.](https://pubmed.ncbi.nlm.nih.gov/8774215/)
- [Almiron-Roig E, Solis-Trapala I, Dodd J, Jebb SA. Estimating food portions. Influence of unit number, meal type and energy density. Appetite. 2013.](https://pubmed.ncbi.nlm.nih.gov/23932948/)
- [Champagne CM, et al. Energy intake and energy expenditure: a controlled study comparing dietitians and non-dietitians. J Am Diet Assoc. 2002.](https://pubmed.ncbi.nlm.nih.gov/12396160/)
- [Bingham SA, et al. Comparison of dietary assessment methods in nutritional epidemiology: weighed records v. 24 h recalls, food-frequency questionnaires and estimated-diet records. Br J Nutr. 1994.](https://pubmed.ncbi.nlm.nih.gov/7986792/)

Source: BurnWeek — "Weigh it, or just say it: the BurnWeek scale and logging from ChatGPT", https://burnweek.fit/blog/smart-scale-and-chatgpt-logging/. Licensed CC BY 4.0: free to quote or reuse with a link to this page.