Random errors cancel out. Systematic bias gets corrected. Here's the math.
The first question people ask about AI-powered calorie tracking is: how accurate is it?
It's a fair question. We take it seriously enough to publish our full accuracy numbers — 743 lab-verified dishes, no cherry-picking. The headline: a mean absolute error of 152 kcal per meal, with about a third of estimates landing within ±20% of the true value.
Those numbers aren't perfect. No vision-based food estimation system is. But here's the thing most people miss: for the purpose of hitting your weight goals, individual meal accuracy barely matters.
What matters is what happens over days and weeks. And that's where the math gets interesting.
Imagine the AI estimates your lunch at 650 kcal when it's really 550 kcal (overestimate by 100). Then at dinner, it estimates 480 kcal when it's really 580 kcal (underestimate by 100). Your daily total is dead on.
This isn't just a convenient example — it's how random errors behave mathematically. If the AI's estimation errors are distributed around the true value (sometimes over, sometimes under), the errors tend to cancel when you sum across multiple meals.
With 3 to 4 meals a day, you log 21 to 28 data points per week. The law of large numbers tells us that the average error shrinks in proportion to 1/√N. After a week of tracking, the noise in your daily total is substantially smaller than the error on any single meal.
Your weekly calorie average can be close to the truth even when individual meal estimates are off by 20–30%.
This is why logging consistently matters so much more than logging perfectly. A week of imprecise estimates gives a better picture than three days of carefully weighed portions followed by four days of nothing.
Random error cancellation only works if the errors are roughly balanced — sometimes high, sometimes low. What if the AI consistently overestimates, or consistently underestimates?
This is a real possibility. Our benchmark data shows a small positive bias on calories (+13 kcal on average) and larger biases in some calorie ranges. For small meals under 200 kcal, the AI tends to overestimate by about 40 kcal. For large meals over 900 kcal, it underestimates by over 300 kcal.
If this were a static calorie counter — one that simply sets a target and hopes you hit it — systematic bias would be a problem. You'd consistently eat more or less than you think, and your weight goals would drift.
But Clawrie isn't a static calculator. It runs an adaptive algorithm that self-corrects.
TDEE stands for Total Daily Energy Expenditure — the number of calories your body burns each day. Most apps estimate this once from a formula and never update it. Clawrie does something different: it continuously recalculates your TDEE from real data.
Every week, Clawrie looks at two things:
From these two signals, it computes your implied TDEE using a fundamental energy balance equation:
The constant 7,700 represents the approximate number of kilocalories stored in one kilogram of body fat. If you ate an average of 2,000 kcal/day and lost 0.5 kg that week, the math implies your true TDEE is about 2,550 kcal/day — you were in a 550 kcal daily deficit.
Clawrie smooths these weekly estimates using an exponentially weighted moving average (EWMA) with a ~4-day effective window, which filters out day-to-day noise from water retention and glycogen fluctuations while still responding quickly to real metabolic changes. Weight readings are separately smoothed with a ~6-day window before being used in the calculation.
Here's where bias gets neutralised. Let's walk through a concrete scenario:
The same mechanism works in reverse. If the AI consistently underestimates, you gain weight faster (or lose slower) than expected, the algorithm adjusts TDEE upward, and the target rises to compensate.
The system converges to the calorie target that produces your desired weight change — regardless of whether individual meal estimates are perfectly accurate.
This is a closed-loop feedback system, not an open-loop calculator. The output (your weight trend) feeds back into the input (your calorie target). Any consistent estimation error gets absorbed into the TDEE estimate and automatically compensated for.
The adaptive TDEE needs a minimum of 7 days of data to produce its first estimate — one full sliding window with at least 2 weight entries and at least 4 days of calorie logs. Before that, Clawrie falls back to a standard formula-based estimate (Mifflin-St Jeor BMR × activity multiplier).
After the first week, the EWMA smoothing means the algorithm gives more weight to recent data while still considering history. In practice, you'll see significant correction within 2 to 3 weeks of consistent tracking. By week 4, the adaptive TDEE has typically converged close to your true expenditure — even if the AI has been systematically over- or underestimating every meal by 15–20%.
The key requirement is consistency. The algorithm can correct for biased estimates, but it can't correct for missing data. If you only log 2 out of 7 days, the sliding window doesn't have enough signal to compute a reliable implied TDEE.
Given how the system works, here's what moves the needle for your results:
The one real failure mode isn't inaccuracy — it's inconsistency. A food tracker with 30% error that you use every day will get you to your goal. A food tracker with 5% error that you abandon after two weeks won't.
AI calorie estimation isn't perfect. We know this — we publish the exact numbers to prove it. But perfection isn't the point.
Clawrie isn't a nutrition database that looks up numbers and hopes they're right. It's a feedback control system that continuously learns from your body's real response. That's what makes it robust.