Are AI Nutrition Plans Safe? The Danger of LLM Coaching
Why KEXBI uses deterministic nodes for safety-critical numbers like energy availability and hydration ceilings, reserving language models for creative synthesis only.
~6 MIN · KEXBI RESEARCH
Why KEXBI uses deterministic nodes for safety-critical numbers like energy availability and hydration ceilings, reserving language models for creative synthesis only.
The new wave of chatbot-first AI nutrition tools is handing safety-critical numbers to a language model. Calorie floors. Hydration ceilings. Sodium targets. KEXBI's architecture refuses that choice. Safety-critical calculations are fully deterministic, bounded by peer-reviewed physiology. AI is used for creative synthesis only. Here is exactly where the line is drawn, and why it is the only line that matters.
The AI-nutrition market is not one category. It is three.
There are the established tracking apps (MyFitnessPal, Cronometer, Lifesum) whose core logic is deterministic and which have begun layering AI features on top. There are data-first coaches (Zoe, Levels, MacroFactor) built on specifically trained models over microbiome data, continuous glucose, or adaptive expenditure. And then there is a third category, the chatbot-first cohort, where the nutrition plan itself is generated by a frontier language model on demand, reasoned-to rather than computed.
It is this third category we are arguing against.
The specific problem inside that cohort is not what most people worry about. It is not that the advice is bland. It is not that the meal plans are generic. It is not that the onboarding is slow.
The problem is that the numbers themselves are being generated by a language model. The calorie floor. The sodium target. The hydration ceiling. The protein dose.
The model has no concept of what happens to an endurance athlete whose energy availability drops below a physiological floor. It has no model of gastric emptying. It has no concept of sleep-adjusted insulin sensitivity. It produces plausible-sounding text that is statistically close to the training corpus, and then it stops.
That is fine when the answer is “a recipe for roast aubergine.” It is not fine when the answer is “how many calories you can cut this week before your cycle stops arriving,” or “how much sodium to consume across a 3-hour ride in 28°C heat.” These questions have real physiological floors and ceilings. Cross them, and people get hurt.
KEXBI was built on one engineering conviction, stated plainly: the numbers are never hallucinated. Only the language is modelled. Here is exactly how that split is implemented, and why every peer-reviewed threshold it rests on is one we can name.
Inside the KEXBI pipeline, every calculation that can harm an athlete if it is wrong runs through a deterministic node. Every calculation that can be creative without consequence runs through a language model. The two layers are not mixed. The outputs of the synthesis layer are bounded at runtime by the outputs of the deterministic layer.
If the language model tries to plate a 420 kcal breakfast on a day where the deterministic engine has calculated a 620 kcal target, the Bouncer rejects the plate. The meal gets rebuilt until the constraint is satisfied. The athlete never sees the rejected version.
The four calculations below are representative, not exhaustive. Each one sits on a peer-reviewed physiological threshold. Each one would be unsafe to delegate to a language model. Each one is, in KEXBI, a closed-form equation or rule that runs the same way every time and can be audited line-by-line.
Energy availability (EA) is the single most important number in endurance nutrition. It is also the one most likely to be wrong in a chatbot-generated plan.
EA is defined as dietary energy intake minus the energy cost of exercise, divided by fat-free mass. When it drops below about 30 kcal per kilogram of fat-free mass per day, the body enters the cluster of symptoms now formalised as Relative Energy Deficiency in Sport, or RED-S.
The consequences are not subtle: suppressed reproductive function, thyroid downregulation, bone-mineral-density loss, impaired immunity, and elevated cortisol [1][2][3].
The 30 kcal/kg FFM threshold is the most robust number in the entire field. It is the point below which luteal-phase suppression, T3 suppression, and bone turnover markers shift in controlled feeding studies. The 45 kcal/kg figure is the eumenorrheic optimum. KEXBI's engine refuses to build a day below the hard floor. A language model, by contrast, has no reason to refuse. It will cheerfully output a 1,600 kcal target for a 65 kg runner burning 900 kcal a day, because nothing in its weights objects.
Cites: Loucks & Thuma 2003 [1]; Mountjoy et al. 2014/2018/2023 IOC consensus [2]; De Souza et al. 2014 [3].
Hydration is the other place a hallucinated number becomes a medical problem.
Under-drink in the heat and you cook. Over-drink and you dilute plasma sodium into exercise-associated hyponatraemia, which is a hospitalisable event. The corridor is not a matter of taste. It is bounded by three numbers with decades of peer-reviewed characterisation: gastric emptying, intestinal absorption, and sweat sodium.
Gastric emptying is rate-limited. The stomach cannot hand off fluid faster than about 1.5 to 2.4 L/hr in most trained adults, with individual variation [4][5]. The 150% post-exercise replacement figure is the Shirreffs & Maughan finding that has stood for two decades: drink only what you lost and you are still in negative water balance two hours later [6]. The carbohydrate ceilings come from Jeukendrup's multi-transportable carb work [7]. An LLM asked to “plan hydration for a 3-hour ride” has no guarantee it will respect any of these. KEXBI's Hydration Architect does, because the ceilings are hard-coded, not prompted.
Cites: Rehrer et al. 1990 [4]; Jeukendrup & Moseley 2010 [5]; Shirreffs & Maughan 1998 [6]; Jeukendrup 2014 [7].
Sleep is not a wellness variable. It is a metabolic variable, and KEXBI treats it as one.
When sleep efficiency drops below about 85%, insulin signalling is measurably impaired the following day. The KEXBI internal model applies a 23% penalty, consistent with the Spiegel and Donga literature showing 25 to 30% drops in peripheral insulin sensitivity after a single night of restricted sleep [8][9]. Resting metabolic rate drops by roughly 3%. Free T3 is suppressed by about 5% [10]. Cortisol rises. Hunger-regulating hormones (leptin, ghrelin) swing sharply against the athlete.
Handing these adjustments to a language model and asking it to “modify today's plan for a bad night's sleep” is not engineering. It is vibes.
Every coefficient here is a closed-form multiplier derived from the controlled-sleep-restriction literature. They are not prompted. They are applied. The downstream effect propagates deterministically through the day's blueprint: shifted macro timing, a softer session cap, a tighter evening carbohydrate bias. A user can audit the adjustment. A regulator can audit the adjustment. An LLM-only competitor cannot, because no such fixed chain exists in their system.
Cites: Spiegel et al. 1999 [8]; Donga et al. 2010 [9]; Nedeltcheva et al. 2010 [10].
The least glamorous deterministic layer is also the one most often botched by LLM nutrition tools: the basic arithmetic of calories and macros.
KEXBI uses Mifflin-St Jeor for resting metabolic rate. It derives a physical activity factor (PAL) from the training-block detector. It applies a protein floor from the Helms and Phillips consensus on athletic populations, 1.6 to 2.2 g per kilogram of body mass per day, with the upper end reserved for energy-restricted or hypertrophy blocks [11][12][13][14].
None of this is novel. That is the point. These are the same equations a registered sports dietitian has used for decades, implemented exactly, every time, with no drift. An LLM will approximate them. It will occasionally forget to subtract protein calories before computing carbohydrate. It will round in ways that move protein under the floor. KEXBI does not approximate. It calculates once, and hands the fixed numbers to the synthesis layer to build meals within.
Cites: Mifflin et al. 1990 [11]; FAO/WHO/UNU PAL tables [12]; Helms et al. 2014 [13]; Phillips & Van Loon 2011 [14].
None of the above means KEXBI is anti-AI. The opposite is true. AI is indispensable exactly where deterministic logic fails: in the creative, culturally-rich, palate-sensitive act of building a meal an athlete will actually eat.
Once the deterministic engine has produced a constraint set (“this meal must be 620 kcal, 45 g protein, 80 g carbohydrate, ≤18 g fat, draws only from the Shadow Pantry, uses FlavorKit 04, lands between 13:30 and 14:15”), a language model does exactly what it is good at. It assembles a plausible, appealing, cohesive dish. It names it. It writes the three-sentence coach narration. It explains why this plate fits the biological window the athlete is in.
Then the output passes through Bouncer, the final deterministic check. Bouncer re-verifies macro totals, FlavorKit integrity, pantry availability, and safety-sentinel flags. Anything that drifted during the synthesis step is rebuilt. The athlete only sees plates that survive both layers.
In practice, the synthesis layer is where personality, cuisine, and variety live. The deterministic layer is the reason that personality cannot kill anyone.
In the next eighteen months, the number of nutrition apps claiming “AI coaching” will continue to multiply. A significant share of them will be a single API call to a frontier model, a system prompt, and a subscription page.
The margin of error on those products is exactly the margin of error on the underlying language model. For the calculations that matter, that margin is not bounded by anything physiological. One hallucinated sodium target in a heat-wave training block is one product-ending event.
KEXBI's position is not that AI is dangerous. It is that AI is incorrectly scoped in the chatbot-first segment of the market.
Safety-critical calculations belong in deterministic nodes with peer-reviewed thresholds that can be named, cited, and audited. Creative synthesis belongs in a language model with its outputs gated. The moat is not a secret algorithm. It is the discipline of having drawn that line in the first place, and of having built the twenty-six-node pipeline that enforces it.
The numbers are never hallucinated. Only the language is modelled. That is the entire argument. And it is the only argument that survives regulatory contact.
All physiological thresholds and coefficients in this article are drawn from peer-reviewed primary literature and major consensus statements. KEXBI's internal implementation uses these same sources as its ground truth. The numerical coefficients shown in the calculation boxes (0.77, 0.97, 0.95, 1.13) are the KEXBI metabolic-brake multipliers, calibrated against the cited sleep-restriction literature and subject to ongoing refinement as new data arrives.