Future Innovations in Fitness
Linas JuozėnasShare
Future Innovations in Fitness
What artificial intelligence and genetic testing can offer—and what still needs to be demonstrated.
A training plan can feel personal because it knows your name, reads your watch or includes a DNA report. The more useful question is whether that information leads to better decisions for your goals, health and everyday life.
This guide explains the technologies, examines the evidence behind their claims, and offers a practical way to decide when they deserve your attention, trust and data.
Identify the decision you want help with before adding another measurement.
Accurate data, a plausible prediction and effective advice are different achievements.
A useful tool should help you understand, question and adapt the plan.
What personalisation should mean
Personalisation means adapting something to a particular person. In fitness, that might involve selecting an accessible exercise, arranging sessions around work, adjusting a programme after illness, or explaining an unfamiliar movement in a way the person understands.
None of these tasks necessarily requires genetic testing. Some need little technology at all. The starting information may be the person’s goals, experience, preferences, equipment, available time and response to previous training.
Define the problem before choosing the tool
“Help me make progress” is a reasonable ambition, but it leaves many questions unanswered. Progress towards what? What is getting in the way? What decision would become easier if you had more information?
Consider three people. One keeps missing sessions because the schedule is unrealistic. Another wants feedback on a movement they have just learned. A third has recurring symptoms during exercise. A calendar assistant, a qualified instructor and a clinical assessment address different needs. A product marketed as personalised fitness should not blur those distinctions.
Separate a description from a recommendation
A report can describe a pattern without establishing what to do about it. It might identify an association between a genetic variant and a trait, or between a recorded training load and a later outcome. The next step—choosing an intervention—requires additional justification.
The distinction between measuring a feature accurately and showing its relevance in a defined setting is central to the V3 framework for digital measurement tools. The framework also recognises that useful deployment involves issues beyond measurement, including usability and clinical utility.[1]
For readers, this suggests three practical questions: Is the information dependable? Does it answer the question I actually have? Would using it improve my decision? An impressive answer to the first question does not settle the other two.
Leave room for more than one good plan
A person may have several reasonable ways to pursue the same goal. Enjoyment, cost, transport and family responsibilities can help determine which plan is workable. A system that ignores those realities may produce an elaborate programme that rarely gets used.
The aim is not to discover a hidden, perfect programme encoded somewhere in a dashboard. It is to make informed choices, observe what happens and adapt with a clear reason. Technology earns its place when it makes that process easier or more effective.
What AI actually does in a fitness product
Artificial intelligence is an umbrella term. Two apps using the same label may work very differently, and a single product can combine several approaches. Knowing which part performs which task makes its claims easier to assess.
| Approach | What it may do | What to ask |
|---|---|---|
| Rules and ordinary automation | Apply a predefined instruction, such as offering a shorter session when you select “less time”. | What triggers the change, and can I inspect or override it? |
| Machine-learning prediction | Use patterns learned from data to estimate an outcome, classify an activity or rank suggestions. | Was it evaluated on new people in a setting relevant to me? |
| Generative AI | Produce conversational explanations, summaries or draft plans in response to instructions. | Are factual claims and calculations checked, and are my constraints respected? |
| Computer vision | Estimate body landmarks or movement features from images and video. | Which movements, camera conditions and users were included in validation? |
| Automatic equipment control | Change a machine’s resistance, speed or other settings; the control may or may not use AI. | What limits, stop controls and supervision does this particular system require? |
New inputs do not necessarily mean new learning
A product can respond to today’s information using a model that remains unchanged. Feeding a new workout into an existing algorithm is different from retraining that algorithm. Some systems update periodically; some use fixed rules; others adapt selected parameters. “It learns from every workout” needs a clear explanation of what actually changes.
More data can create more opportunities for analysis, but quantity alone does not establish quality. Repeated errors, missing context or a dataset that poorly represents the intended users can undermine the result. A new model version also needs evaluation; an update is not automatically an improvement for every person or task.
A fluent explanation can still be wrong
Generative systems can confidently produce false information, contradict instructions or invent supporting references. NIST’s guidance calls this problem confabulation. It also cautions against broad capability claims based on narrow or anecdotal demonstrations.[2]
For example, an assistant might describe a programme as suitable for a home session while including equipment you do not have. It might offer a reassuring explanation for a training increase without checking what happened in the previous sessions. Read the actual recommendation and its assumptions, rather than relying on the tone of the explanation.
Useful questions are ordinary ones: What information did this use? What is missing? Why did the plan change? What happens if I decline? You should not need a computing qualification to understand the answer.
Measurements, estimates and readiness scores
An AI system often receives a mixture of sensor signals, calculated metrics and information entered by the user. A completed-session log, an estimated calorie total and a note saying “felt unusually difficult” do not have the same meaning or the same sources of error.
Know what the device measured
A wrist sensor may detect changes in reflected light and use an algorithm to estimate heart rate. A device can perform differently during different activities. Research examining optical sensors found that accuracy depended on the device and conditions; a validation result for one setting should not be treated as a guarantee for every movement or wearer.[3]
Further calculations may turn available signals into estimates of energy expenditure, sleep stages, aerobic fitness or recovery. The presence of a precise number on the screen does not make all those quantities directly measured.
A comparison of seven older consumer sleep-tracking devices against polysomnography found varying performance, including limitations in sleep-stage classification. This illustrates why a sleep-stage display needs appropriate validation; it does not establish the accuracy of every current device.[4]
A combined score is a model’s interpretation
A readiness score usually combines selected inputs according to a provider’s method. Without knowing that method and its validation, you cannot assume that 80 on one platform means the same thing as 80 on another. Nor does a change of a few points necessarily represent a meaningful change in your capacity.
Ask whether a score reflects a measurement, a prediction or an instruction. Those are separate functions. A green indicator is not a clinical assessment, and it should not be used to dismiss new symptoms or override restrictions agreed with a treating professional.
Check context before changing the programme
Imagine that a wearable records much less activity than usual. The person may have been inactive, forgotten the device, charged it during a session or done an activity the system captured poorly. Each explanation has different implications. Automatically prescribing extra exercise assumes the first explanation without checking the others.
A short note can sometimes be more informative than another calculated metric: the device was not worn, the session was interrupted, the route was different, or the person was ill. A useful system should make it easy to correct those assumptions.
Avoid exact calorie compensation
Research comparing wrist-worn devices found that heart-rate and energy-expenditure accuracy differed substantially. The study tested older products, so its value here is the distinction between metrics, rather than a current device ranking.[5]
Do not treat an estimated exercise calorie total as an exact instruction to eat or remove the same amount of food. If food intake is the concern, review the wider pattern and the reason for making a change. Additional decimal places do not resolve uncertain inputs.
Where AI-assisted coaching can be useful
Some valuable uses of AI are modest and concrete. A tool may organise a training log, summarise an agreed plan, suggest questions for a coach, or help compare ways to fit familiar sessions into a changing week. These uses are worth judging on their own terms.
Reduce the work around the workout
Suppose someone already has a suitable programme but repeatedly loses track of which sessions they completed. A clear weekly summary could help them and their coach review attendance and identify scheduling problems. The benefit is better organisation. It does not require a claim that the software has discovered their biological optimum.
Similarly, an assistant might turn a long explanation into a short reminder, arrange equipment alternatives supplied by an instructor, or prepare a list of questions for the next appointment. The person can check the output against the original information.
Adapt to an explicit constraint
A useful request is specific: “I have twenty minutes available today. Help me review the shorter option already included in my programme.” That gives the system a defined task and a reference point. “Optimise everything about my fitness” invites assumptions that may remain invisible.
When a plan changes, look for a short account of the change and its reason. Was the adjustment made because time was limited, a session was missed, the person reported difficulty, or a metric crossed a threshold? Those explanations should be distinguishable.
Keep continuity as well as flexibility
A changing recommendation can feel attentive, but change itself is not evidence of better coaching. If exercises, targets and schedules are all replaced repeatedly, it becomes harder to understand what was actually attempted and how the person responded.
One sensible design goal is to preserve a coherent plan while allowing adjustments for relevant information. A busy week might require a simpler schedule. It does not necessarily require a new training philosophy. The reason for an adjustment should be more substantial than producing a fresh-looking screen.
Use explanations to support a conversation
Ask the tool to list assumptions, identify information it lacks and explain what would change its recommendation. Then inspect those answers. An AI-generated explanation can itself contain errors, so it is a starting point for review rather than an independent verification.
A coach or clinician should also assess the underlying information, rather than accepting the generated summary without checking it. Human involvement is useful when it adds relevant expertise, responsibility and understanding of the person’s circumstances.
An example of a bounded request
“Summarise these completed sessions and my notes. Keep recorded facts separate from possible explanations. Identify questions to discuss with my coach. Do not infer a diagnosis or invent missing measurements.”
This is an illustration of clearer task-setting, not a guarantee that a particular model will respond correctly. Check the summary against the original log before using it.
The outcome to look for might be time saved, fewer scheduling errors or a clearer conversation. Those are meaningful benefits when they address a real problem, even without a promise of superior athletic performance.
Movement feedback and injury prediction
Video analysis offers an appealing possibility: receive feedback without needing a laboratory visit for every movement. Its usefulness depends on what the system estimates, how it was tested and what conclusion is drawn from the result.
What a camera-based system can estimate
OpenCap is a research example combining videos from two or more smartphones, pose estimation and biomechanical modelling to estimate movement and related dynamics. Its validation supports that defined system and the tasks studied. It does not validate every single-camera exercise app. Estimated muscle forces are model outputs, rather than direct measurements from inside the body.[6]
Camera placement, lighting, visibility and the movement being performed matter to interpretation. If a joint is obscured or the system has not been evaluated for an assistive device, an apparently precise overlay may be misleading. Ask what an app does when it cannot obtain a reliable estimate.
A different movement is not automatically a faulty movement
A squat can look different with different body proportions, loads, stances and task instructions. Forward trunk inclination alone does not establish an error or explain pain. A coloured line on a video cannot replace an assessment of what the person is trying to do, how the movement feels and whether the feedback is relevant.
Feedback should identify the specific observation and the intended benefit of any proposed change. “Your knee is unsafe” is a much stronger claim than “the camera estimated a different knee position in these repetitions”. A responsible interpretation respects that difference.
Injury prediction is an active research area
A 2026 study used twelve months of prospective monitoring in 142 competitive endurance runners to develop injury-prediction models. This is useful research into possible predictors. It does not establish that the model works in a separate population or that providing its feedback prevents injuries.[7]
That distinction matters when a consumer product claims to predict an injury before it happens. A model can find patterns in recorded outcomes without proving that the recommended response improves what happens next.
False alarms and missed problems both matter
A system that labels many people “high risk” may create unnecessary worry or restriction. One that misses relevant problems may give false reassurance. Assessing a model requires more than hearing about successful predictions.
Calibration concerns whether predicted probabilities agree with observed frequencies in a relevant group. Ranking some people as more likely to experience an outcome than others is a different achievement from providing dependable absolute probabilities.[8]
For an individual, the practical question remains: what action follows, and what evidence supports that action? New or worsening pain, fainting, chest discomfort or other concerning symptoms should not be handled by asking an exercise app to repeatedly recalculate a score. Seek appropriate clinical help; severe or acute symptoms may require urgent care.
How to judge the evidence behind an AI tool
“Science-backed” can mean that a product cites general exercise research, that a sensor was compared with a reference method, or that the complete service was tested in users. Those are different kinds of evidence. Ask which claim each study supports.
| Question | Useful information | What remains unresolved |
|---|---|---|
| What was tested? | The actual feature, model version, measurement or complete coaching service. | Research on a component may not establish the effectiveness of the whole product. |
| Who was included? | Training experience, age range, relevant health conditions and other features of the study population. | Results may not transfer to users or settings absent from the evaluation. |
| What was it compared with? | A clearly described alternative, such as ordinary coaching or the same app without the new feature. | An improvement over time does not identify which component caused it. |
| What outcome improved? | An outcome connected to the claim: performance, participation, symptoms, injuries, burden or another defined result. | More app engagement does not by itself prove better health. |
| What went wrong? | Withdrawals, adverse events, missing information and cases where recommendations were inappropriate. | Reporting only successful users can hide important limitations. |
Test on information the model has not already seen
A model can appear unusually accurate when evaluation information leaks into its development. For example, if closely related records from one person appear in both sets, or information available only after an injury is used to predict it, the test may exaggerate real-world performance. Research on data leakage describes this as a wider problem across machine-learning-based science.[9]
Readers do not need to audit the code. They can ask whether evaluation used genuinely separate people or settings, and whether the input information would have been available at the time a real decision was made.
Compare the added feature fairly
If an AI service provides reminders, extra coaching contact and a new plan simultaneously, an improvement does not reveal which part helped. A fair comparison of the AI contribution should make the alternatives clear.
The same applies to a DNA add-on. Comparing a heavily supported genetic programme with a leaflet cannot isolate the value of genetic information. The question is whether adding that information improves an otherwise comparable service.
Read uncertainty as useful information
A small exploratory study may justify further research. It may not settle whether a tool improves outcomes over months or years. A wide range of plausible effects is different from a precise estimate, and the absence of a statistically clear benefit is not evidence that every possible future approach will fail.
Look for a provider willing to say what remains unknown. Clear limits help you choose an appropriate level of reliance. A feature used to arrange a calendar and one used to make a consequential health recommendation should not demand the same degree of trust.
Regulatory status has a defined scope
European guidance on medical-device software places importance on intended purpose. Whether software falls within medical-device rules depends on what it is intended to do and the applicable criteria. A CE mark should not be read as proof that every optional coaching feature produces an ideal workout.[10]
Ask which function and intended use a regulatory claim covers. Professional endorsements, a laboratory credential and a product’s popularity also answer different questions. None should be used as a substitute for evidence supporting the specific recommendation.
What a consumer DNA test actually measures
Consumer genetic testing commonly starts with a saliva sample or cheek swab, followed by laboratory analysis and a report. The convenience of that process should not obscure the more important questions: what was examined, how accurately, and what the result can justify.[11]
A test may examine selected variants
A single nucleotide polymorphism, or SNP, is a variation at a particular position in DNA. Many consumer tests use arrays that examine selected positions. Some services use sequencing, and methods differ. Testing selected variants is not equivalent to reading and interpreting every relevant part of someone’s genome.[12]
Ask which method the laboratory uses and which results are directly measured. If a service infers additional variants from measured ones, ask how that inference was validated. A large number of reported positions is not, by itself, evidence of useful fitness advice.
Three separate questions about quality
Genetic-test evaluation distinguishes analytical validity, clinical validity and clinical utility. In clinical testing these terms concern measurement, the relationship to a condition and usefulness for care. The same distinction helps expose gaps in a fitness claim.[13]
| Question | Meaning | Illustration in fitness |
|---|---|---|
| Analytical validity | Does the assay identify the relevant genetic variant accurately? | The laboratory correctly reports which variant was present in the sample. |
| Clinical validity or predictive relevance | How well does the result relate to the claimed trait or outcome in an appropriate population? | The variant meaningfully predicts the response the company claims to predict. |
| Clinical utility or decision benefit | Does using the result improve a consequential decision or outcome? | Choosing a programme from the report works better than a suitable alternative without it. |
The report includes another layer of judgement
The laboratory result and the advice may come from different organisations or methods. One service might label a result “power potential”; another might use it in a combined score. The label is an interpretation. Ask for the evidence connecting the measured variant to that exact conclusion.
Consumer reports often give probabilities and examine only a portion of the variants associated with a trait or health condition. A lower reported risk cannot exclude every relevant cause, and a higher reported risk does not establish that the outcome will occur.[14]
It is also possible for an interpretation to change while the underlying measured DNA result stays the same. A company may change its algorithm or incorporate new research. That change should be explained as a revised interpretation, rather than presented as a discovery that your biology suddenly became different.
Before buying a test, complete this sentence: “If the result is different from what I expect, I would change this decision.” Then ask whether evidence supports that change. If the answer is unclear, the report may be interesting without being useful for the purpose you had in mind.
Interpreting common fitness gene claims
Genes contribute to physical traits, but athletic performance involves many traits and a person’s lived circumstances. Research associations are often converted into much stronger consumer labels: “born sprinter”, “endurance type” or “poor recovery”. Those labels can conceal uncertainty.
ACTN3: a protein difference is not a training programme
ACTN3 provides instructions for alpha-actinin-3, a protein found predominantly in fast-twitch muscle fibres. People with the R577X XX genotype can lack functional alpha-actinin-3 protein; they do not simply lack the ACTN3 gene. Associations with athletic traits do not directly reveal an individual’s muscle-fibre proportions or ideal exercises.[15]
Being represented more often in an elite-athlete group is not the same as responding better to a particular programme. To justify a prescription, the relevant question is whether people with the result benefit from that prescription compared with a suitable alternative.
| Marker or label | What the research concerns | What the report does not establish |
|---|---|---|
| ACTN3 | A functional protein difference and associations with aspects of athletic performance. | A fixed sporting identity, exact muscle-fibre ratio or ideal repetition range. |
| ACE | An enzyme involved in physiological regulation, with variants studied in relation to athletic traits.[15] | A dependable selector between strength and endurance programmes. |
| FTO | Common-variant associations with body mass index and obesity risk.[16] | The calorie target or macronutrient split that will work best for this individual. |
| Variants near MC4R | Population associations with fat mass, weight and obesity risk.[17] | A diagnosis from a wellness label or proof that one particular diet is required. |
| Common MTHFR variants | Variation in an enzyme involved in folate processing. | A diagnosis of poor recovery, a need for high-dose supplements or an inability to use folic acid.[18] |
Why “genes are not destiny” is not a sufficient qualification
A report may acknowledge environmental influences and still make an unsupported recommendation. Saying that a result is only one part of the picture does not demonstrate that it adds useful information to the picture.
A 2015 sports consensus concluded that available genetic tests had no role in talent identification or individualised training prescription. That historical statement does not settle every future technology. It does establish why a current provider needs direct, relevant evidence rather than relying on the general fact that genes influence traits.[19]
Do not turn a label into a limit
If a report calls you an “endurance type”, that does not justify giving up an activity you enjoy or assuming that strength work cannot help you. Likewise, a “power” label is not permission to attempt unfamiliar explosive exercise without suitable preparation.
A more useful discussion begins with your goal, current abilities, instruction needs and response to training. Those observations can be reviewed directly. A genetic label should not displace them with a prediction that the provider cannot substantiate.
Be precise about supplements
CDC guidance states that people with common MTHFR variants can process folate, including folic acid. These variants are not a reason to avoid folic acid.[18] If a report recommends a supplement, ask what clinical or dietary need it addresses and how the dose was determined. The name of a gene is not a dosing instruction.
What nutrition studies actually show
Personalised nutrition can involve ordinary, valuable information: current food intake, allergies, preferences, cultural practices, available money, training demands and relevant health conditions. Adding genetic information is a separate intervention whose benefit needs to be tested.
DIETFITS: testing a genetic matching hypothesis
The DIETFITS randomised trial followed 609 adults without diabetes for twelve months while comparing healthy low-fat and healthy low-carbohydrate approaches. The tested genotype pattern did not identify who would lose more weight on one approach than the other.[20]
This result concerns the particular matching hypothesis, participants and outcome studied. It does not prove that all gene–diet interactions are irrelevant. It does show why a plausible genetic explanation should be tested before being sold as a dependable diet selector.
Food4Me: personalisation helped without an added genetic benefit
The six-month Food4Me trial involved adults from seven European countries. Personalised advice improved aspects of dietary behaviour compared with conventional advice, but adding phenotypic information and then genetic information did not enhance effectiveness in the strategy tested.[21]
The practical distinction is useful: advice may be personal and helpful because it responds to what a person actually eats and the changes they can make. The presence of DNA results is not what makes every form of personalisation valuable.
A predisposition is not a food-intolerance diagnosis
Variants associated with lactase persistence concern the ability to digest lactose after childhood. Lactose intolerance involves symptoms, and not everyone with lactose malabsorption has them. It also differs from milk allergy.[22] A genetic label alone therefore does not explain every digestive symptom or establish how much of a food someone can tolerate.
For coeliac disease, HLA-DQ2 or HLA-DQ8 can be relevant to clinical assessment. Their absence makes the condition very unlikely, but their presence alone does not diagnose it. Starting a gluten-free diet before indicated diagnostic testing can affect the results; discuss the testing process with a clinician first.[23]
These examples illustrate why “food sensitivity genes” is too vague a category for making major restrictions. Different conditions involve different mechanisms and diagnostic pathways.
AI does not remove uncertainty from food records
A meal photograph or incomplete food log leaves information unknown. Portions, ingredients, preparation and unrecorded meals can change the interpretation. An assistant can help organise the record or generate questions, but an apparently precise nutrient total may rest on several assumptions.
It should not diagnose a deficiency from those assumptions or convert a casual pattern into a firm cause. If later meals and poorer sleep occurred together, for example, work schedules, stress or other circumstances may also have changed.
An appropriate nutrition discussion asks what problem needs solving and what evidence could clarify it. Persistent digestive symptoms, unintended weight change, difficulty eating enough or concern about a deficiency deserve assessment in their own right. Buying a genetic add-on is not a prerequisite for receiving useful nutrition support.
Wellness reports and clinical genetic testing
A fitness report and a clinical genetic investigation may both involve DNA, but they can have very different purposes. Clinical testing may be used to investigate a suspected inherited condition, clarify a relevant family history or guide a particular aspect of care.
Start with the clinical question
Healthcare-led testing includes choosing an appropriate test and interpreting the result with the person’s history. Consumer results generally require additional clinical evaluation, and often confirmatory testing, before they are used to diagnose a condition or make healthcare decisions.[11]
That does not mean every consumer result is wrong. It means that a meaningful health decision needs the right test, interpretation and context. A wellness report may never have examined the variant or condition that actually matters.
What to do with a concerning result
Keep the original report, the laboratory details, the exact variant information if provided, and the date of analysis. Bring those to an appropriate healthcare professional or genetics service. Ask whether the finding is relevant, whether it needs confirmation and what further information is needed.
Avoid making major changes to medicines, restricting a child’s activities or abandoning exercise solely because an automated report used alarming language. Conversely, a reassuring wellness score should not close a clinical question raised by symptoms or family history.
Raw data are not a diagnosis
Some services allow raw-data downloads and third-party interpretation. A new website can attach additional labels to the same file without having collected a new sample or confirmed the underlying result. MedlinePlus advises assessing both the credibility and privacy practices of such interpretation services.[14]
Uploading the file repeatedly can multiply interpretations without resolving which one is correct. If the question has become medical, the next useful step is clinical clarification rather than accumulating more automated reports.
Common variants and rare disorders should not be conflated
A common variant associated with a small population difference is not interchangeable with a pathogenic variant interpreted in an appropriate clinical setting. Even when two reports mention the same gene, they may concern different variants, evidence and consequences.
This is why a gene name should never be treated as the whole diagnosis. Ask exactly what was found and what conclusion is justified. You can also ask what remains uncertain and whether the result changes any current decision.
Counselling includes the choice to know
Genetic information may raise questions about biological relatives and findings beyond the original fitness interest. A professional discussion can help clarify the scope of a proposed clinical test and the implications of possible results. It should also give the person space to express what they want to know and why.
Curiosity about fitness does not obligate someone to investigate every possible genetic risk. The scope of testing should be understandable before consent is given, with a clear distinction between optional wellness information and a clinical investigation.
Fairness, access and personal autonomy
A technology can work well for some people and poorly for others. It may also create practical barriers even when its calculations are accurate. Evaluating innovation therefore includes asking who can use it, who was represented in its development and who controls the resulting decisions.
Representation affects what a model can claim
Research on polygenic risk scores has documented differences in predictive performance associated with ancestry and the populations represented in discovery datasets. A major 2019 analysis highlighted the consequences of heavily European-ancestry data.[24]
Those findings are not a numerical performance estimate for every present-day fitness score. They are a reason to ask for validation in relevant populations rather than assuming that a large study automatically represents everyone.
For AI movement or coaching systems, ask similarly whether evaluation included people with relevant ages, training experience, disabilities and health conditions. A system developed around competitive athletes may need different evidence before being used to guide someone returning to activity after illness.
Preserve opportunity for young people
A youth sports programme should allow room to learn, enjoy different activities and develop over time. A consumer DNA label should not become a gate that decides who gets coaching, encouragement or a place on a team.
The sports-genetics consensus explicitly raised the issue of talent identification.[19] Beyond the scientific limitation, a useful question for any organisation is what happens to the child who declines testing. Participation should not depend on surrendering unnecessary genetic information.
Notice pressure disguised as choice
Consent can be difficult to give freely when the person requesting information controls employment, selection or access to support. EDPB guidance identifies the power imbalance in employment as a particular concern for consent-based processing.[25]
A coach or organisation should be able to explain why information is needed, who will see it and what alternatives exist. “Everyone else has uploaded their report” is not an explanation of necessity or benefit.
Consider the whole cost
The cost of a service includes subscriptions, compatible equipment, time spent entering data and the effort of interpreting alerts. It may also include access needs: captions, readable text, language options, usable controls or the ability to record an adapted activity.
Compare that total burden with the problem the tool is meant to solve. If the main difficulty is finding an accessible place to exercise, a detailed genetic report may add little. If a simple reminder or shared log already meets the need, greater complexity should earn its place.
A good outcome can include more confidence, clearer choices and less administrative work. A system that makes someone anxious about every meal or reluctant to move without permission from a score may be working against the purpose for which it was adopted.
Privacy, consent and the life of your data
A fitness service may hold location records, training history, sleep estimates, photographs and health information. Adding a genetic file expands the sensitivity of that collection. It may also reveal information relevant to biological relatives, so the implications can extend beyond the person opening the account.
The EU framework in plain language
Under the GDPR, genetic data and health data are special categories of personal data. Processing generally needs a lawful basis under Article 6 and an applicable condition under Article 9. Explicit consent is one possible route, rather than the only route in every setting. Purpose limitation, data minimisation and appropriate protection remain relevant.[26]
Encryption helps protect information against certain forms of access. It does not, by itself, establish a lawful purpose, valid consent or anonymity. Ask both how data are protected and what the organisation is allowed to do with them.
Follow the information beyond the upload screen
Before sharing a genetic file, find clear answers to these questions:
- Purpose: Which requested feature needs these data, and can I use the service without that feature?
- Recipients: Which organisations receive the sample, raw data, derived scores or reports?
- Additional uses: Are the information or samples used for research, advertising, product development or model training?
- Retention: How long are the physical sample, account records and backups kept?
- Control: How can I download information, change optional permissions, request deletion or close the account?
- Change of ownership: What does the policy say about a sale, merger or closure of the service?
These are questions to ask a provider, not a claim that every service uses data in all these ways. A concise, specific answer is more useful than a general assurance that privacy is taken seriously.
Understand what withdrawing consent changes
Where processing relies on consent, withdrawing it does not retrospectively make earlier lawful processing unlawful. Data-erasure rights also have conditions and exceptions. Account closure should not be assumed to remove every retained copy or reverse every completed use.[25][26]
Ask the provider to distinguish stopping future optional use, deleting identifiable records, destroying a stored sample and dealing with already-produced research or models. These may involve different processes. If the explanation is unclear, resolve it before uploading information you cannot meaningfully retrieve from every recipient.
Choose the smallest useful disclosure
If the task is to rearrange a training week, a schedule and a few constraints may be sufficient. If you want a coach to discuss one report, sharing that relevant result may be more appropriate than transferring an entire raw genetic file to an unrelated chat service.
Review location sharing and public profiles as well. A useful private training record does not need to become a public route map. Use available account protections and remove connections that no longer serve a purpose.
Do not assume universal rules about employers or insurers
Restrictions on requesting or using genetic information depend on the country, purpose and circumstances. An international article cannot establish the legal position for a particular employment or insurance decision. Seek advice relevant to that jurisdiction when the consequences matter.
The practical principle is to make the information flow understandable before agreeing to it. Privacy is part of whether the tool is suitable for you, alongside its usefulness and cost.
What future progress would need to demonstrate
AI and genetics can be combined technically. That fact alone does not establish that the combination improves fitness decisions. Adding a weak predictor to a sophisticated system does not turn it into strong evidence.
Test the value of the genetic addition
A useful research comparison would give one group a system using ordinary personal information and observed training response, and another an otherwise comparable system that also uses genetic data. Researchers would then examine whether the extra data improve outcomes enough to justify the cost, burden and potential harms.
That is a proposed evaluation approach, not a claim that every such trial has already been completed. It asks the question a prospective user actually faces: what do I gain by adding the test?
Keep different kinds of information in their proper roles
A measured inherited variant is generally stable, while available time, recent training, illness and other circumstances can change. A future system would need a sound reason for how each type of information influences a recommendation.
A genetic report should not be described as having measured someone’s current lactate threshold, protein requirement or recovery state unless the claimed measurement and interpretation are independently justified. The label “multimodal” simply means that several kinds of information are being used; it is not a certificate of accuracy.
Promising directions to investigate
Better measurement
More dependable movement analysis across settings and users, with clear uncertainty when a measurement cannot be trusted.
Better decisions
Recommendations evaluated against useful alternatives, with outcomes that matter to participants.
Better access
Tools that work with varied abilities, languages and resources while requiring only necessary data.
These are directions for progress, not promises that every new product already delivers them. Advances should be judged by the evidence for each task and population.
Protect the ability to question and correct
A well-designed future service should explain why it changed a recommendation, record relevant model updates, permit correction of inaccurate inputs and provide a way to escalate questions. It should also recognise when it lacks enough information to give a useful answer.
Users and professionals need to distinguish a recorded fact, an estimate and a proposed action. A polished conversational interface should make those distinctions easier to see.
Make success meaningful
Success might mean that a person can participate more consistently, understand a programme better, need less administrative help or reach a relevant performance goal. An increase in time spent inside the app is useful only if it serves the person’s purpose.
The strongest future claim would therefore be a specific one: a defined tool helped a defined group achieve a worthwhile outcome under stated conditions, with acceptable drawbacks. That claim is less dramatic than a universal biological blueprint, but far more useful for deciding what to trust.
A practical decision framework
You do not need to adopt every new tool to keep learning about fitness. Begin with one problem and judge a proposed service against it. The following framework is an original decision aid, rather than a validated clinical checklist.
-
Name the decision.
Describe what you want help with: arranging sessions, reviewing a log, understanding feedback or investigating a health concern. Choose an outcome you can recognise.
-
Ask what information is necessary.
List the facts that could change the decision. Consider whether the task can be handled with your existing plan and observations before adding sensors or genetic testing.
-
Match the claim to evidence.
Look for evaluation of the relevant feature, users and outcome. Ask what the comparison was and what remains uncertain. For a DNA add-on, ask what benefit came specifically from genetic information.
-
Check control and practical fit.
Review data use, cost, accessibility, cancellation, human support and the ability to reject or correct recommendations. Make sure the tool fits the way you actually live.
-
Review the result and decide again.
For an appropriate low-risk use, assess whether the tool solved the original problem. Continue, simplify or stop according to that result. Health concerns require the relevant clinical pathway.
Example: the schedule is the problem
A person with an established programme struggles when work hours change. They use an assistant to arrange the already-agreed sessions around the coming week, then check that the output preserved the instructions. They share only the schedule information needed for the task.
The review question is simple: did planning become easier and more accurate? A DNA test would need a separate, clearly justified purpose. It would not solve the scheduling problem merely by making the profile more detailed.
Example: a report recommends a different sport
A recreational cyclist receives a “power-oriented” label and wonders whether to stop longer rides. Before changing activities, they ask the provider what study shows that people with this exact result do better when following that advice.
If the answer is only an association with elite athletes, the proposed change remains unsupported. The person can discuss their actual goals and training response with a coach without treating the label as an obligation.
Example: the dashboard and symptoms disagree
A runner has worsening pain while an app reports high readiness. The relevant issue is the symptom and its context. Repeating the measurement until a different score appears does not address it. Appropriate assessment can determine what to do next; the app’s colour should not decide whether a concern deserves attention.
Record what would justify keeping the service
Write one sentence before starting: “This is useful if it helps me ___ without adding an unacceptable burden.” The blank might concern a clearer log, fewer scheduling mistakes or better access to an agreed programme.
Decide when to review that purpose. This is a usability review, not a home experiment capable of proving medical effectiveness. Avoid changing multiple important health behaviours simply to test a product’s marketing claim.
You may conclude that a simple tool is enough, that professional help better addresses the problem, or that a particular technology is worthwhile for a limited task. Each is a valid outcome of an informed decision.
Common questions
Can AI create a useful training plan?
It can generate a plan and help organise relevant information. Whether the plan is suitable depends on the person, the system, the information available and the quality of review. Check the exercises, progression, constraints and assumptions. A convincing explanation alone does not establish that the programme is appropriate.
Does an app learn every time I record a workout?
Not necessarily. It may apply a fixed model or predefined rule to new information. Ask whether the underlying model changes, what triggers an update and how the updated version is evaluated. A new output is not proof of continuous learning.
Can an ACTN3 result tell me which sport to choose?
It does not provide a validated instruction to choose or abandon a sport. An association with athletic traits differs from evidence that following a specific recommendation improves your outcome. Use your interests, opportunities, current abilities and experience of training when making that choice.
Is a laboratory-accurate DNA test automatically useful?
No. Correctly identifying a variant answers a measurement question. Whether it predicts the relevant outcome, and whether using it improves a decision, are additional questions. Look for evidence supporting the advice attached to the result.
Should an MTHFR result change my supplements?
A common MTHFR variant alone is not a supplement prescription or a reason to avoid folic acid. Ask what need the proposed supplement addresses and whether the recommendation is supported by an appropriate dietary or clinical assessment.
Does a low genetic risk result mean I am protected?
A consumer report may examine only selected variants and provide a probability estimate. It cannot exclude all relevant causes of a problem. Symptoms or a concerning family history should be assessed on their own merits, even when a wellness report appears reassuring.
Can an app predict and prevent my next injury?
Those are two different claims. A prediction model requires appropriate evaluation, and the action taken in response needs evidence that it helps. Do not treat a score as certainty or use it to dismiss pain and other concerning symptoms.
Should I upload raw DNA data into a general chatbot?
First ask what useful task requires the file and whether the service is suitable for that information. Review retention, additional uses and recipients. An ordinary planning task may need only a schedule and a few constraints. Consequential genetic interpretation belongs in an appropriate clinical discussion.
Will deleting my account remove every use of my data?
Do not assume so. Ask about retained samples, backups, recipients and already-completed uses, as well as the applicable conditions for deletion. Withdrawal of consent and account closure are related to data control but do not necessarily have identical effects.
Do I need AI or genetic testing to personalise fitness?
No. Personalisation can begin with goals, experience, preferences, access, health needs and observed response. Add technology when it answers a real question or reduces a real difficulty. More collected information is not itself a better outcome.
Sources and scope
This guide draws on primary research, methodological frameworks, public health information and European regulatory guidance. A study of a sensor, an association study and a trial of advice answer different questions. Findings are limited to the methods, populations and outcomes evaluated.
Older studies are included when they illustrate an enduring distinction, such as the difference between predicting an outcome and improving it. They should not be read as current product rankings. The 2015 sports-genetics consensus is identified as historical; present-day commercial claims still require relevant evidence of their own.
The planning examples, evaluation questions and five-step framework are original teaching illustrations. They are not clinical protocols, diagnostic tools or proof of effectiveness for a particular product. No consumer service is endorsed, and no genetic result is used here to prescribe an individual workout or diet.
Sources checked: 6 September 2026. Product capabilities, evidence and legal requirements can change. The linked sources and their scope notes allow future readers to trace the basis of the discussion.
- Goldsack JC, Coravos A, Bakker JP, et al. Verification, analytical validation, and clinical validation (V3): the foundation of determining fit-for-purpose for Biometric Monitoring Technologies (BioMeTs). npj Digital Medicine. 2020;3:55.A methodological framework separating sensor performance, measurement validation and clinical relevance. It does not validate an individual product. Return to text ↑
- Autio C, Schwartz R, Dunietz J, et al. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1. July 2024.Official cross-sector guidance on generative AI risks, including confabulation. Not a study of exercise outcomes. Return to text ↑
- Bent B, Goldstein BA, Kibbe WA, Dunn JP. Investigating sources of inaccuracy in wearable optical heart rate sensors. npj Digital Medicine (2020), 3:18.Fifty-three adults across skin-tone categories; device and activity effects, with a device-by-skin-tone interaction. No overall significant skin-tone effect does not establish universal equivalence. Return to text ↑
- Chinoy ED, et al. Performance of seven consumer sleep-tracking devices compared with polysomnography. Sleep (2021), 44(5):zsaa291.Thirty-four healthy young adults across three laboratory nights; older wearable and contactless devices varied in wake detection and stage classification. Return to text ↑
- Shcherbina A, Mattsson CM, Waggott D, et al. Accuracy in Wrist-Worn, Sensor-Based Measurements of Heart Rate and Energy Expenditure in a Diverse Cohort. Journal of Personalized Medicine. 2017;7(2):3.Historical comparison of seven devices showing why heart-rate and energy-expenditure accuracy must be evaluated separately. It is not a current product ranking. Return to text ↑
- Uhlrich SD, Falisse A, Kidziński Ł, et al. OpenCap: Human movement dynamics from smartphone videos. PLOS Computational Biology. 2023;19(10):e1011462.Validation and application of a defined multi-camera movement-analysis system. Findings do not transfer automatically to single-camera consumer apps or every clinical population. Return to text ↑
- Wu H, Brooke-Wavell K, Barnes MR, et al. Multidisciplinary prediction of running-related injuries using machine learning. npj Digital Medicine. 2026;9:213.Prediction modelling using 12-month prospective monitoring of 142 competitive endurance runners. No separate-cohort external validation or trial of injury reduction from model-informed feedback. Return to text ↑
- Van Calster B, McLernon DJ, van Smeden M, Wynants L, Steyerberg EW; STRATOS initiative. Calibration: the Achilles heel of predictive analytics. BMC Medicine. 2019;17:230.Methodological explanation of probability calibration. The article’s hypothetical prediction example is original, not data from this paper. Return to text ↑
- Kapoor S, Narayanan A. Leakage and the reproducibility crisis in machine-learning-based science. Patterns. 2023;4(9):100804.Research examining leakage and reproducibility across fields. The training-log examples in this guide illustrate the general issue. Return to text ↑
- Medical Device Coordination Group. Guidance on Qualification and Classification of Software in Regulation (EU) 2017/745—MDR and Regulation (EU) 2017/746—IVDR. MDCG 2019-11 rev. 1; June 2025.European guidance explaining the role of intended purpose in software qualification and classification. Guidance does not replace the applicable legislation. Return to text ↑
- MedlinePlus Genetics, National Library of Medicine. What is direct-to-consumer genetic testing?. Updated 2 June 2026.Explains consumer and clinical testing pathways; consequential health decisions may require confirmatory testing. Return to text ↑
- National Human Genome Research Institute. Direct-to-Consumer Genetic Testing FAQ for Healthcare Professionals. Provider education; accessed 6 September 2026.Explains SNP arrays, test scope and clinical interpretation. Assays vary; this is not a current catalogue of consumer services. Return to text ↑
- MedlinePlus Genetics, National Library of Medicine. How can I be sure a genetic test is valid and useful?. Public genetics education; accessed 6 September 2026.Distinguishes analytical validity, clinical validity and clinical utility. The fitness examples in the text apply that distinction. Return to text ↑
- MedlinePlus Genetics, National Library of Medicine. What do the results of direct-to-consumer genetic testing mean?. Public genetics education; accessed 6 September 2026.Covers probabilistic results, incomplete variant coverage and third-party interpretation of raw data. Return to text ↑
- MedlinePlus Genetics, National Library of Medicine. Is athletic performance determined by genetics?. Public genetics education; accessed 6 September 2026.Explains genetic and environmental contributions. Associations do not establish an individual training prescription. Return to text ↑
- Frayling TM, Timpson NJ, Weedon MN, et al. A common variant in the FTO gene is associated with body mass index and predisposes to childhood and adult obesity. Science. 2007;316(5826):889–894; doi:10.1126/science.1141634.Population association study, not a trial of genotype-matched macronutrient advice. Return to text ↑
- Loos RJF, Lindgren CM, Li S, et al. Common variants near MC4R are associated with fat mass, weight and risk of obesity. Nature Genetics. 2008;40:768–775; doi:10.1038/ng.140.Population associations do not establish an ideal diet for an individual or equate to a clinical diagnosis. Return to text ↑
- Centers for Disease Control and Prevention. MTHFR Gene Variant and Folic Acid Facts. Updated 16 July 2026.Common MTHFR variants do not prevent use of folic acid. This guide does not prescribe a supplement dose. Return to text ↑
- Webborn N, Williams A, McNamee M, et al. Direct-to-consumer genetic testing for predicting sports performance and talent identification: Consensus statement. British Journal of Sports Medicine. 2015;49(23):1486–1491; doi:10.1136/bjsports-2015-095343.Historical sports consensus opposing use for talent identification and individual training prescription. A present-day product still requires direct evidence for its claims. Return to text ↑
- Gardner CD, Trepanowski JF, Del Gobbo LC, et al. Effect of Low-Fat vs Low-Carbohydrate Diet on 12-Month Weight Loss: the DIETFITS Randomized Clinical Trial. JAMA, 2018;319:667–679; doi:10.1001/jama.2018.0245.Randomised trial in 609 adults without diabetes. Title shortened; findings concern the tested genetic matching approach and weight outcome. Return to text ↑
- Celis-Morales C, Livingstone KM, Marsaux CFM, et al. Effect of personalized nutrition on health-related behaviour change: evidence from the Food4Me European randomized controlled trial. International Journal of Epidemiology, 2017;46:578–588; doi:10.1093/ije/dyw186.Six-month European randomised trial. Added phenotypic and genetic information did not improve the advice's effectiveness; long-term disease prevention was not established. Return to text ↑
- National Institute of Diabetes and Digestive and Kidney Diseases. Definition & Facts for Lactose Intolerance. Patient information; reviewed February 2018.Distinguishes lactose malabsorption from symptomatic intolerance and milk allergy. Return to text ↑
- National Institute of Diabetes and Digestive and Kidney Diseases. Diagnosis of Celiac Disease. Patient information; reviewed October 2020.DQ2/DQ8 results have a limited role in clinical assessment. Positive genetic results alone do not diagnose coeliac disease. Return to text ↑
- Martin AR, Kanai M, Kamatani Y, et al. Clinical use of current polygenic risk scores may exacerbate health disparities. Nature Genetics. 2019;51:584–591; doi:10.1038/s41588-019-0379-x.Analysis of ancestry-related differences in polygenic prediction. Historical findings establish a validation concern, not a performance estimate for every current score. Return to text ↑
- European Data Protection Board. Guidelines 05/2020 on consent under Regulation 2016/679. Adopted 4 May 2020; version 1.1, 13 May 2020.Guidance on freely given consent, including the imbalance in employment relationships; not a determination that every workplace fitness scheme is unlawful. Return to text ↑
- European Parliament and Council of the European Union. Regulation (EU) 2016/679: General Data Protection Regulation. Regulation adopted 27 April 2016; Articles 5, 6, 9 and 17.EU legal framework. This article explains selected principles; legal rights and exceptions depend on the circumstances and applicable national provisions. Return to text ↑