Evaluating an AI Health Platform: Eight Questions
Every platform in this category describes itself the same way. Eight questions separate the ones doing something from the ones with a questionnaire and a catalogue.
The Short Answer
Every product in this category claims artificial intelligence, evidence-based recommendations and personalisation, which means the descriptions carry no information. The differences are real and they show up in eight specific behaviours rather than in marketing language. Most of the eight cost the provider something, which is what makes them informative, and a platform failing four or more is a questionnaire with a catalogue attached.
Questions One to Three: Does It Show Its Reasoning
One: does each recommendation state its evidence and its strength? A recommendation resting on twenty randomised trials and one resting on a cell-culture study should not be presented identically. Look for an explicit grading, and for a stated reason rather than a mechanism narrative.
Two: does it distinguish mechanism from outcome? The most common failure. A compound with abundant in vitro pharmacology and no human data at an achievable dose should be marked as such, and platforms frequently present the mechanism as though it were the evidence.
Three: is the dose specified, and does it match the trials? A recommendation for omega-3 without stating combined EPA and DHA content, or for curcumin without specifying a bioavailability-enhanced formulation, is not recommending the studied intervention. Precision theatre is the opposite failure: a dose of 847 mg implies a resolution the evidence does not support.
These three are checkable in minutes by reading a single recommendation carefully, and they eliminate a substantial share of the category.
Questions Four and Five: Does It Ever Say No
| Behaviour | What it indicates |
|---|---|
| Recommends removing an item | The system is not purely additive |
| Concludes that no change is needed | It is evaluating rather than generating output |
| Publishes what it recommends against | A filter is being applied at all |
| States uncertainty where evidence is thin | Calibration rather than confidence |
| Revises when evidence changes | The recommendation tracks the literature rather than the launch date |
Four: does it ever recommend nothing, or removal? The sharpest single test. The correct answer for a well-nourished person with an adequate stack is frequently no change, and a system that always finds something to suggest is optimising something other than the user's health.
Five: does it have an exclusion list? Applied honestly, an evidence hierarchy produces a long set of negative conclusions: ginkgo for cognitive decline, glucosamine for osteoarthritis, B vitamins for cardiovascular prevention, tribulus for testosterone. A platform presenting these as unproven rather than as tested and negative is misrepresenting the evidence in a commercially convenient direction.
These two are the questions with the highest discriminating power, because both cost revenue.
Questions Six and Seven: Safety and Incentive
Six: does it ask about medications, and check them? Interaction checking is the highest-value and least visible function any such platform performs, since documented pairs run to thousands across cytochrome P450 enzymes, transporters, absorption competition and pharmacodynamic overlap. A platform that does not ask what you take is skipping the part that prevents harm.
Follow-up worth asking: does it re-run the check when you add something, and does it handle a serious interaction as an exclusion rather than a warning? In a weighted model a strong efficacy case can outweigh a safety concern, which is the wrong structure.
Seven: who profits from the recommendation? A platform recommending from its own product line has a structural incentive, which does not invalidate the output and does mean it should be weighted accordingly and disclosed rather than discovered.
The related question: does the retest and review cadence match the biology, or the subscription? Most blood markers have a three-month floor, and quarterly panels of slow markers serve retention rather than the member.
Question Eight: Does It Learn From Your Data
This is the question that separates a static recommender from something whose value compounds.
Does it record what you tried, at what dose, for how long, and what happened? Including the null results, which are the most valuable and least recorded category. Without this, the same experiment gets repeated under different branding.
Does it enforce one change at a time with an observation window? Attribution requires it, and a platform recommending five simultaneous additions has made evaluation impossible.
Does it re-measure and conclude? A window that opens and never closes is not an experiment.
Does it detect drift? Change of one or two per cent per year is below perceptual threshold and trivially detectable in a kept sequence, and this is something software genuinely does better than a person.
Can you export the record? If not, the asset belongs to the platform, and the record is the thing that becomes valuable over a decade.
A platform strong on this question and mediocre on the others is more useful over ten years than the reverse, because the temporal information cannot be computed and the reasoning largely can.
What No Platform Can Do
Worth stating so that the eight questions are not read as a search for a complete solution.
Generate evidence that does not exist. Where a compound has thin trial data, sophistication does not resolve it, and the honest output is a calibrated uncertainty.
Predict your outcome. Population associations are wide at the individual level, and no amount of personal data overcomes that.
Verify slow outcomes. Whether bone loading or apoB reduction changed your trajectory is not observable in any personal record.
Replace clinical care. Where a marker is markedly abnormal, a symptom is progressive, or medication is indicated, a platform's role is preparation and routing.
Produce adherence. The largest determinant of outcomes, and no software makes the training happen. This is the biggest limit and the one least often stated.
A platform claiming otherwise on any of these five is overclaiming in a way the eight questions will not catch, which is why they are worth reading alongside them.
The AEONNN Perspective
These are the questions AEONNN should be judged by, so it is worth stating where the platform stands. Evidence Levels A, B and C appear on every recommendation, doses are specified against the trial form, and the Journal's negative conclusions are the exclusion list this article says to look for.
Two of the eight cost revenue and are therefore the most informative. The platform recommends removal and concludes that no change is needed where that is the answer, which is why removal trials sit at the centre of its governance logic rather than at the edge.
Question six is the one AEONNN weights most heavily internally, and its structure matters: the Safety layer operates as a veto rather than an input, because in a weighted model a strong efficacy case can outweigh a serious interaction. And question eight is the one the platform is actually built around. Response history, drift detection and enforced one-change-at-a-time observation are where the value compounds, because the temporal information cannot be computed while the reasoning largely can. The limit worth repeating is the last one: no platform produces adherence.
Pillar Matrix mapping
Database Matrix layers
- Evidence Layer (PubMed, Cochrane, ClinicalTrials.gov)
- Safety Layer (DrugBank, FAERS)
- Quality / Formulation Layer (ConsumerLab, Labdoor)
- Meta / Consensus Layer (JAMA, BMJ, specialty society positions)
Frequently Asked
How do I evaluate an AI health platform?
Eight questions: does it state evidence and strength, distinguish mechanism from outcome, specify doses matching trials, ever recommend nothing or removal, publish an exclusion list, check medications, disclose who profits, and learn from your data.
Which question discriminates most?
Whether it ever recommends nothing or removal. The correct answer for a well-nourished person with an adequate stack is frequently no change, and a system that always finds something is optimising something else.
What is precision theatre?
Returning a dose like 847 mg, which implies a resolution the underlying evidence does not support. Trials establishing dose ranges used round numbers.
Why does medication checking matter so much?
Documented interaction pairs run to thousands across enzymes, transporters, absorption competition and pharmacodynamic overlap. A platform that does not ask what you take is skipping the part that prevents harm.
Should a serious interaction be a warning or an exclusion?
An exclusion. In a weighted model a strong efficacy case can outweigh a safety concern, which is the wrong structure for a contraindication.
What makes a platform valuable over years?
Recording what you tried and what happened including null results, enforcing one change at a time with a window, closing that window with a re-measurement, detecting drift, and letting you export the record.
What can no platform do?
Generate evidence that does not exist, predict your individual outcome, verify slow outcomes, replace clinical care, or produce adherence. The last is the largest limit.
Evidence and review
Any dosage ranges cited here reflect the ranges used in published human trials, not personal recommendations. Evidence in this field moves, so this article is reviewed quarterly and carries its last-updated date above. Nothing here is intended as medical advice, and supplementation should be discussed with a qualified clinician, particularly alongside prescribed medication or an existing condition.