Meal Intelligence (Photo Carb Estimates)

Meal Intelligence lets you take a photo of a meal and get an AI estimate of its carbohydrate content as a range (for example, "≈ 40–55 g carbs"), with a confidence signal. You can correct the estimate, save foods you eat often, and the AI can reference your logged meals in chat and daily briefs.

Photo carb estimates are AI guesses, frequently wrong, and a rough starting point only — never calculate an insulin dose or bolus from them. The AI looks at an image and guesses; it regularly misjudges portions and sometimes misidentifies the food entirely. Treat every number as a ballpark to sanity-check against your own carb counting. Always verify carbs yourself before dosing, and consult your healthcare provider about your diabetes management.

This is an experimental feature, on by default, that you can turn off (or back on) at any time from Settings. It is not FDA-cleared and is not a carb-counting authority.

What it does

1. Estimate — you take or pick a meal photo; the AI returns a carb range and a confidence level (low / medium / high).
2. Correct — if the estimate is off, you correct it. Your correction becomes the truth the app remembers — it never silently overwrites your number with a guess.
3. Save — foods you eat often can be saved so a re-photographed meal recognizes them ("you've logged this before") instead of re-guessing.
4. Aware — when you've logged meals, chat and daily briefs can reflect them back to you ("you logged a high-carb dinner — how did that sit with you?"). They never tell you how much insulin to take.

How it's built to be safe

The feature is designed so a guess can never quietly turn into a dose:

• A range, not a single confident number. Real plates are uncertain; a single integer invites you to dose off it.
• Confidence comes from disagreement, not the model's self-rating. The same photo is sampled several times; the spread between answers drives the confidence. (A model's own "I'm confident" score does not track accuracy and is never shown as a safety signal.) A wide spread is shown plainly — "this could be 40 g or 90 g, we're not sure."
• You confirm what the food is before the app grounds it against nutrition data, so a misidentified food can't be "certified" with an authoritative-looking citation. For branded restaurant items this can cite the chain's own published nutrition — see Restaurant Nutrition Grounding.
• Carb estimates never flow into insulin-on-board, safety limits, or any dosing math. They are descriptive notes, structurally isolated from the dosing engine.
• Every screen that shows a carb estimate carries a persistent reminder — "Rough estimate — an AI guess that's often wrong. Never use it to calculate an insulin dose or bolus."

Known limitations

• Misidentification is the most common error — look-alike foods (a Linzer torte read as a Bakewell tart, crème catalana as crème brûlée) can be confidently wrong.
• Run-to-run variance — the same photo can yield different numbers; the confidence range is meant to surface that, not hide it.
• Accuracy is honest, not perfect — on a labeled test set the estimate was within a useful band most of the time, but tail cases can be far off. This is why it stays experimental and why you must verify.

Turning it on or off

Meal Intelligence is on by default. You control it per-account from Settings → Meal Intelligence in both the web app and the phone app — turn it off to hide the meal surfaces (the "Log a meal" button, the Meals area, and the photo-capture flow), or back on to restore them. There's no environment variable or operator step: it's an ordinary in-app preference, like your glucose display unit, and your choice syncs across your devices.

To actually produce estimates the feature needs a vision-capable AI provider configured — with it on but no vision provider, the surfaces appear but a photo estimate returns a clear "not available" error rather than a guess. Because it sends a food photo to your configured AI provider, the same BYOAI data-handling rules apply as the rest of the app — review your provider's policy.

If you run a local model, note that meal photos have a higher quality bar than text chat: an unverified local model is refused for photo estimates (with a clear message) rather than allowed to produce a low-quality guess. See Local AI Vision for which models clear the bar and why cloud is the verified path for photos today.

The bottom line

A carb estimate from a photo is a guess about a picture, not a measurement and never a dose. Use it to get a baseline, correct it when it's wrong, and count your carbs and decide your insulin the way you and your healthcare provider do today.