Key takeaways
- Independent testing puts AI photo estimates in a wide band — roughly 62–99% accuracy depending on app and food.
- Portion size, not food identification, is the weakest step in the pipeline.
- Single foods score far better than mixed dishes, where oils and sauces are invisible to a camera.
- For weight change, consistency beats precision — a repeatable error gets corrected by the scale.
We build an AI food scanner, so this is a question we get asked constantly — usually by people who suspect the honest answer is worse than the marketing. It is a fair suspicion. Here is what the accuracy research actually shows, where the errors come from, and how much it matters for the thing you are trying to do.
The numbers
Published comparisons of consumer AI calorie trackers report accuracy across a broad range, commonly summarised as 62% to 99%, with typical errors falling between 10% and 40%. Vendor-independent benchmarking has found average error rates differing by more than an order of magnitude between the best and worst apps tested.
85–90%
Single, distinct foods
An apple, a chicken breast, a bowl of rice — clear boundaries, familiar shapes.
65–75%
Mixed dishes
Burrito bowls, casseroles, curries — ingredients hidden below the surface.
~39%
Portion estimation, worst case
The weakest link in the chain, and the reason two identical plates can score differently.
Read those together and a pattern emerges. Identifying what the food is has become reliable. Estimating how much of it there is remains genuinely hard, and no amount of model improvement fully solves it, because the information simply is not in a single photograph.
Why portions are the hard part
A photo is a two-dimensional projection of a three-dimensional plate. To turn it into grams, a model has to infer depth, density, and what is underneath the visible top layer — three things a flat image only hints at.
Depth is invisible
Density varies
Fats hide
Preparation is unknowable
The comparison people forget
Human estimates are not a clean baseline either. Research using doubly labelled water has consistently found people under-report their own intake by a wide margin — often 20–40% — when logging manually. The relevant question is not “is AI perfect?” but “is AI better than the eyeball estimate you would otherwise have entered?”
Accuracy versus consistency
Per entry, careful manual logging with a food scale beats a photo scan — one analysis put manual database entry roughly 12% more accurate on average. Across a month, that advantage often evaporates, because manual logging is tedious enough that most people stop.
This is the trade-off that actually decides outcomes. A log that is 15% off but complete tells you far more than a perfect log covering four days in March. And here is the part that surprises people: a consistent error barely matters.
Why systematic error self-corrects
Suppose your scans read 12% low every day. Your log says 2,000 kcal; you really ate 2,270. If you set a target from your TDEE and your weight does not move for three weeks, you cut the logged target by 200 and progress resumes. The absolute number was wrong the whole time — the direction and size of the adjustment were still right, because the error stayed put.
What genuinely hurts is random error: scanning some meals and eyeballing others, logging weekdays but not weekends. That noise cannot be calibrated away. The eight most common tracking mistakes are almost all of this type.
Scan, then correct — the accuracy lives in the edit
2BIB returns a per-item breakdown rather than a single number, so you can adjust the one portion the AI misjudged instead of accepting or rejecting the whole meal. Barcode and nutrition-label scans are exact, and unlimited on every plan.
How to make your scans meaningfully better
- 1Shoot at roughly 45 degrees, not straight down. An angled shot gives the model depth cues a top-down photo destroys.
- 2Get the whole plate in frame, in good light. Shadows, harsh backlighting, and cropped edges all degrade the estimate.
- 3Include a size reference. A fork, a standard mug, or your hand at the edge of frame gives scale. Without one, plate diameter is a guess.
- 4Scan before you start eating. A half-eaten plate is a much harder estimation problem than a full one.
- 5Separate mixed dishes when you can. Photographing components before they are combined turns one 70%-accurate estimate into three 88%-accurate ones.
- 6Correct the portion when it looks off. You know whether that was one cup of rice or two. Two seconds of editing removes the single largest source of error.
- 7Use the barcode or nutrition label whenever there is one. Those are not estimates — they are the manufacturer’s own figures.
When to use which method
| Situation | Best method | Why |
|---|---|---|
| Packaged food | Barcode scan | Exact manufacturer data, no estimation involved |
| Home cooking, known recipe | Manual or saved meal | You know the ingredients and weights already |
| Restaurant meal | AI photo scan | No label, no recipe — an estimate beats a shrug |
| A meal you eat weekly | Log once, reuse | Recurring entries remove both effort and variance |
| Prepared food with a label | Label scan (OCR) | Reads the printed panel exactly as written |
The best trackers are not the ones that insist on a single method. Barcodes for packages, label OCR for prepared food, photo scans for everything else, and manual entry when you know better than any of them — that combination gets you a complete log, which is the thing that actually predicts results.
The honest summary
- AI photo scanning is good enough to steer by and not precise enough to treat as measurement.
- Expect it to be close on simple foods and rougher on layered, saucy, restaurant-style dishes.
- The scale over two to three weeks is your calibration instrument — it tells you whether your logged target is set too high or too low, whatever the underlying error.
- A complete log with modest error will outperform a precise log you abandon. That is the entire case for scanning.
If you want the underlying maths behind the target you are steering toward, start with what TDEE is and run your own numbers through the free calculator. For how 2BIB’s scanner works in practice, see the AI food scanner page.