If AI generates a competent meal plan for free, is the registered dietitian's value the clinical-risk catch rather than the plan?

The shift

Producing a personalized meal plan — calories, macros, meal timing, swaps, a shopping list tailored to stated preferences and goals — goes from scarce (an RD's chargeable hours) to abundant, instant, and near-free. The generation itself, the thing most RDs still price and sell as the deliverable, is exactly the token-production task a model does well.

The axioms

• The personalized meal plan is the dietitian's product — the artifact the client pays for.
• Translating dietary goals into a concrete, followable plan is skilled work that takes trained time.
• Someone with a license and legal standing must own a nutrition recommendation when a client has a real medical condition.
• Catching the dangerous edge case — an allergy, a comorbidity, a drug-nutrient interaction, a disordered-eating red flag — is what separates a safe plan from a harmful one.
• Nutrition change happens through a relationship over time, not through a document handed over once.
• The client can't reliably tell a safe plan from a plausible-but-dangerous one; the RD's judgment is the safeguard against that gap.

Invalid axioms

1. The personalized meal plan is the dietitian's product. Plan generation was chargeable because turning preferences and targets into a structured, followable document took trained time. A model does it instantly for free. The habit-trap: RDs, apps, and clinics still price and market "a custom meal plan" as the headline deliverable — the one thing that's now the commodity, not the value.
2. Translating goals into a concrete followable plan is skilled scarce work. The mechanical act — portioning, sequencing meals, generating swaps and a grocery list — is pattern-matching against everything ever written about food, which is precisely what's now abundant. The habit-trap: junior RD time and service tiers are still structured around plan-building hours as the unit of work, when that's the part that collapsed.

Unchanged axioms

1. A licensed, accountable professional must own a nutrition recommendation for a client with a medical condition. Liability doesn't transfer to a model — when a plan harms a diabetic, a renal patient, or someone on interacting medication, a person with credentials and a name answers for it. This is a legal and institutional fact, not a capability gap, so it doesn't move as models improve.
2. Catching the dangerous edge case is what separates a safe plan from a harmful one. Recommending almond milk in a nut-free plan, missing that a "healthy high-potassium" push is dangerous in kidney disease, or overlooking a drug-nutrient interaction — these are the failures that cause direct physical harm. A confident-wrong plan here isn't a redo; it's an ER visit. This raises the bar on verification rather than lowering it, and it's exactly where pattern-matching against the average case breaks on the specific patient.
3. Spotting a disordered-eating red flag is a clinical judgment, not a plan output. A request phrased as a diet goal can be the presenting edge of an eating disorder, where generating any restrictive plan is the harm. Reading that from what a client says and doesn't say, under real stakes, is judgment on novel ambiguity — and a model optimizing for a "competent meal plan" will cheerfully produce the dangerous thing.
4. Nutrition change runs on a relationship over time, not a document. Adherence, honest disclosure of what someone actually eats, working through relapse and shame — these depend on trust with an invested person, not on the quality of the plan file. The plan was never the mechanism of change; it was the artifact people mistook for it.

New axioms

1. Clients now act on confident-wrong AI plans before any professional sees them. The asymmetry that used to protect people — they couldn't generate a plausible, authoritative-sounding plan on their own — is gone. Someone with a condition can follow a dangerous plan for weeks before an RD is ever in the loop, if one ever is.
2. Liability is unsettled when an AI plan harms someone with a condition. Malpractice and scope-of-practice frameworks assume a licensed human authored the recommendation. When the plan came from an app, was lightly reviewed by an RD, or was never reviewed at all, who answers is undefined — and the incentive is to disclaim rather than to own.
3. The RD's job is shifting from authoring plans to auditing them for danger, and nothing is built for that. If the deliverable is free, the scarce act becomes screening AI-generated plans against a specific patient's comorbidities, meds, and history at the volume patients now generate them. That's a different skill, a different workflow, and a different billing unit than "write a meal plan" — none of which the profession has stood up.
4. The screening step gets skipped precisely when it matters most. A free, competent-looking plan feels finished, so the client with the complex condition — the one who most needs the edge-case catch — is the least likely to think they need to pay for a review. The people at highest risk self-select out of the safeguard.

Where it breaks

"The meal plan is the dietitian's product" (invalid) collides with "the RD's job is now auditing AI plans for danger" (new): the profession still prices, markets, and staffs around producing the plan — the free part — while the scarce, value-bearing act has moved to catching what makes a specific plan dangerous for a specific patient. An RD selling plan generation is selling the commodity and giving away the safeguard, and reimbursement has no line item for "reviewed an AI plan and caught the contraindication."

A second collision: "the client can't tell a safe plan from a dangerous one" (still holds) meets "clients act on confident-wrong AI plans before a professional sees them" (new). The safeguard that judgment provides only works if the RD is in the loop — but a free, fluent plan is most convincing exactly to the client who can't evaluate it, and it reaches them before any accountable person is involved. Whether this call holds depends on how fast models get at flagging their own high-risk cases and refusing to generate for them; that reliability is moving fast and is worth watching, but "usually refuses" is not the same safeguard as an accountable human, and the failure mode is silent.

ai-axioms.vercel.app/axioms/healthcare/dietitian-ai-axioms