AI has become a marketing term as much as a technical term. Two apps can both say “AI-powered” while doing radically different things. One may use a simple model to adjust a training plan. Another may analyze images or physiological data. The label does not tell you whether the output is accurate, useful, or appropriate for your situation.
1. Ask what the AI actually does
Try to finish this sentence: “The model takes these inputs and produces this output to help with this decision.”
If the product description stays vague, it becomes difficult to evaluate. “Personalizes your health” is not a testable claim. “Uses recent workout performance and reported difficulty to adjust upcoming training volume” is much more specific.
2. Separate measurement from inference
A phone or wearable may measure one signal and infer another. For example, a sensor can detect movement or pulse-related information, while an algorithm estimates calories, sleep stage, readiness, stress, or recovery.
The FDA’s digital-health glossary is useful because it distinguishes sensors, software, analytics, and medical-device functions rather than treating them as one thing. That vocabulary helps you ask which part of the pipeline has actually been validated.
3. Look for evidence tied to the exact claim
An app may cite studies showing that heart-rate variability is related to recovery, then imply that its proprietary readiness score is validated. Those are not the same claim.
Useful evidence should connect the actual input, algorithm, output, and intended use. If a company has validation studies, look for details about sample size, population, comparison standard, error rates, and whether the study was independent.
4. Ask whether the validation population looks like you
A model evaluated in healthy young adults may behave differently in older adults, people with chronic conditions, highly trained athletes, or users taking medications that affect the measured signal.
This does not make the model useless outside the original sample. It means confidence should decrease as you move farther from the population in which performance was demonstrated.
5. Pay attention to the outcome the app optimizes
Personalization can only be as good as the objective. An exercise app might optimize adherence, progression, predicted performance, or user enjoyment. Those are different goals.
Similarly, a wellness app might optimize engagement rather than a health outcome. That is not inherently bad, but you should know which problem the product is actually designed to solve.
6. Look for uncertainty, not only a score
A model that outputs “73” can feel precise even if the underlying prediction is noisy. Better tools explain how much confidence to place in a result, what can distort it, and which decisions should not be based on it alone.
WHO’s AI guidance emphasizes transparency and intelligibility alongside safety and accountability. Those principles are especially relevant when a product is asking users to change behavior based on a proprietary score.
7. Check privacy and data use separately from accuracy
A useful model can still have privacy practices you dislike. An inaccurate model can still have excellent privacy. Review what information the app collects, whether it shares data, how long information is retained, and whether you can delete it.
8. Match scrutiny to consequences
The FDA notes that digital-health technologies range from general wellness products to technologies used as medical devices. That spectrum should affect your standard.
| If the app is wrong... | Reasonable evidence threshold |
|---|---|
| Your workout is slightly easier than ideal | Useful product testing may be enough |
| You spend significant money on a health product | Look for stronger independent evidence |
| You delay medical evaluation or change treatment | Use a much higher threshold and qualified human input |
9. Notice whether the company tells you what the product cannot do
Clear limitations are a positive signal. Every model has boundaries. A company that describes appropriate use, known weaknesses, and populations not studied is giving you more information than one that only lists benefits.
WHO’s guidance on large multimodal models similarly stresses the importance of governance, transparency, and managing risks rather than assuming capability from fluent output. The principle applies broadly: the more consequential the output, the more visible the boundaries should be.
A simple final test
Ask whether the app has earned the level of trust it is requesting. A recommendation that nudges a low-stakes habit may not need clinical-grade evidence. A tool asking you to interpret symptoms or make treatment decisions should face a much higher bar.
For a broader map of the field, see what digital health means. If the app specifically claims to personalize training, can AI personalize exercise programs? breaks down what real personalization requires.
