Why every AI nutrition plan we generate gets a human signature
An AI model can produce a seven-day meal plan that respects a calorie target, hits a macro split, avoids peanuts, stays vegetarian, and reads beautifully. It can do that in about four seconds, and it will be right most of the time.
Most of the time is not the standard you want for food someone eats every day.
What the model is actually good at
Worth being fair about this, because the useful parts are genuinely useful:
- Constraint juggling. Vegan, gluten-free, no nuts, 1,900 calories, 150g
protein, hates mushrooms, cooks twice a week. Satisfying all of that at once is tedious for a human and trivial for a model.
- Variety. Left to a template, most plans repeat five meals. A model will
cheerfully generate thirty.
- Rewriting to fit a life. "Same plan but nothing that takes more than fifteen
minutes" is a one-line request and a complete redraft.
That is real work, and it is the boring part of a nutritionist's job. Handing it over is a good trade.
What it still gets wrong
It does not know what it was not told. A member with early kidney disease needs a protein ceiling, not a protein target. If that condition is not in the intake — and members routinely omit things they do not think are relevant to food — the model will confidently push protein in exactly the wrong direction. It has no way to know it is missing something.
It optimizes the numbers it can see. Iron and B12 are the classic case in plant-based plans: the macros land perfectly and the micronutrients quietly do not. A model asked about protein will solve for protein.
Drug and food interactions are not a macro problem. Warfarin and vitamin K. MAOIs and aged cheese. Grapefruit and a long list of medications. These live in the member's medication list, not their calorie target, and they matter more than the macro split.
Confidence reads as competence. This is the one that should worry anyone shipping this. The wrong plan is written in the same assured, well-organised prose as the right one. There is no tremor in the output to warn a member that this particular plan is the one with a problem in it. A human expert says "hold on, what medication are you on?" — the model does not, unless something makes it.
So a human signs it
Every AI-generated plan in Conexus Thrive is created in a pending_review state. A member cannot see it as final. A board-certified nutritionist opens it, reads it against the member's clinical profile and medication list, edits whatever needs editing, and approves it. The approval is recorded against that person's name and the time they gave it.
That last detail matters more than it sounds. An approval attached to a specific professional is accountability. An approval that any staff account can rubber-stamp is theatre.
We do not call the AI a nutritionist
There is a marketing temptation here that we have decided against permanently. It would be easy to write "your AI nutritionist" on a landing page. It converts well. It is also a claim about credentials that no model holds, and in a health product that is not a grey area.
So: the AI drafts. A qualified human is accountable for what a member receives. We say which is which, in the product and in the marketing, and the plan a member reads carries the name of the person who approved it.
The result is a nutritionist who spends their time on judgement instead of on arithmetic — which is what you were paying them for anyway.
This article is general wellness education and not medical advice. Talk to a qualified health provider about your own circumstances, particularly if you take medication or have a diagnosed condition.