How accurate is AI at counting calories?

Short answer: AI is good at recognizing what’s on a plate and much shakier at quantifying it. Across peer-reviewed studies, AI calorie estimates range from roughly a few percent off to meaningfully off in some cases — commonly cited calorie (energy) estimation error bands span from around 0.1% up to ~38%, with individual items off by as much as ~80%, and portion sizes are the single biggest source of error. Simple, single foods are estimated reasonably well; mixed meals are where accuracy suffers most (commonly ~25–40% off). No app has “solved” calorie counting — and the honest takeaway is to treat any number, from any tool, as an estimate you should be able to see and adjust.

This post synthesizes published research. We’re not presenting our own lab study here, and we’re not going to pretend the technology is better than the evidence says it is.

What the peer-reviewed studies actually found

Two lines of research are worth knowing.

A 2025 study of ChatGPT-4’s nutrient estimates from food images (PMC11858203) found the model was systematically off on macros — for example, a fat estimation error around 14.8% — and, tellingly, it underestimated portion size in about 76% of the photos it was shown. The model could usually name the food; it struggled to say how much was there.

A systematic review of AI/image-based dietary assessment (PMC10836267) pulled together many studies and reported energy-estimation errors spanning roughly 0.10% to 38.3%, with per-item errors reaching as high as ~79.6%. The review’s consistent theme: accuracy is decent for simple, well-separated foods and degrades sharply for mixed or composite meals — think a stir-fry, a curry, or a loaded sandwich, where ingredients and quantities are hard to disentangle.

Here’s the shape of it:

What’s being estimated	Typical accuracy picture	Source
Single, simple foods (an apple, a plain chicken breast)	Best case — often within ~10%	PMC10836267 (systematic review)
Macros from a photo (fat, carbs, protein)	Meaningful error; e.g. ~14.8% fat error	PMC11858203 (ChatGPT-4 study)
Portion / serving size	The weakest link — underestimated in ~76% of photos	PMC11858203
Mixed / composite meals	Worst case — commonly ~25–40% off, up to ~80% per item	PMC10836267

The pattern is consistent across the literature: recognition is a solved-ish problem; quantification is not. And it’s not unique to AI — humans famously misjudge portions too, and food databases carry their own margins. AI didn’t invent the accuracy problem; it inherited it.

A note on “±1.1% accurate” and “independently validated”

If you shop around, you’ll see calorie apps advertise eye-popping precision — “±1.1% accurate,” “97% accurate,” “independently validated.” Treat those with healthy skepticism. Many such figures are self-published marketing claims, not peer-reviewed studies, and some circulate through affiliate and review networks that trace back to the vendor. We’re not going to name-and-shame specific numbers we can’t verify — but we will say plainly: if an accuracy claim isn’t backed by a published, independent study you can read, it’s a marketing number, not a scientific one. The genuine peer-reviewed evidence above tells a much more modest story than “±1.1%.”

So is AI calorie counting useless? No.

Here’s the encouraging part. The research says AI estimates are imperfect, not worthless. For most people, the goal isn’t a lab-grade measurement — it’s a consistent, directionally-correct trend you can actually maintain. Weight change is driven by patterns over weeks, not by whether Tuesday’s lunch was 540 or 590 calories. And the biggest predictor of success with any tracker is depressingly simple: do you keep using it? Most people quit because logging is tedious, not because the numbers were 15% off.

So the useful question isn’t “is AI perfectly accurate?” (nothing is). It’s “does this help me log consistently and stay honest about the numbers?”

How Forkmate handles this — honestly

We’ll be direct: Forkmate does not claim to be more accurate than the studies above. We use the same kinds of public food data everyone does (USDA FoodData Central and Open Food Facts), and any estimate we produce carries the same real-world uncertainty.

What we do differently is refuse to hide the number. When you log a meal by talking to your AI assistant, Forkmate shows you the calories and macros before they’re saved, so you can confirm or correct them. A photo-scan app that silently books “620 calories” gives you no chance to catch a bad portion guess. Forkmate’s mechanic is the opposite: you see the number, and you fix it if it’s wrong, every time. That doesn’t make the underlying estimate more precise — it makes you the check on it, which is exactly where the research says the weak link is (portions and mixed meals).

It also means the number in your diary is one you agreed to, which — for a habit you’re trying to keep for months — matters more than a decimal point of theoretical precision.

If you want to try that confirm-every-number flow, here’s how to connect Forkmate to your AI assistant. Coming from another app? You can bring your MyFitnessPal history over for free — and we won’t re-estimate the numbers you already logged. Not sure what your daily target should even be? Get a calorie goal in about 60 seconds.

The honest bottom line

• AI is strong at identifying food, weak at estimating portions — and weakest on mixed meals (commonly ~25–40% off, sometimes far more).
• Peer-reviewed calorie (energy) estimation errors span roughly 0.1–38%, with outliers near 80% per item (PMC10836267, PMC11858203).
• Ultra-precise marketing claims (“±1.1% accurate”) are usually self-published, not independent science.
• No tracker is exact — so the winning move is a tool that’s easy enough to keep using and honest enough to show you the number so you can correct it.

Forkmate is a general-wellness tool, not medical advice. If you need clinical-grade accuracy, weigh your food and consult your physician or a registered dietitian.

Sources

• ChatGPT-4 nutrient estimation from food images (2025): PMC11858203
• Systematic review of AI/image-based dietary assessment: PMC10836267