Artificial Intelligence and Machine Learning in Sports
Linas JuozėnasShare
Artificial Intelligence & Fitness
AI can help turn a crowded training log into a clearer next step. The most useful systems make their evidence, assumptions and limits visible—and leave room for the person doing the training.
A watch may detect a change, a model may estimate a risk, and a chatbot may suggest a workout. These are different capabilities. Understanding how they fit together helps us use technology with curiosity, confidence and judgment.
This guide explores predictive analytics, adaptive coaching and movement analysis, with practical ways to assess their value in everyday exercise and sport.
A measured signal, an inferred state and a suggested action answer different questions.
Accurate answers or predictions do not by themselves establish better training outcomes.
Goals, symptoms, preferences and everyday circumstances belong in the decision.
What can artificial intelligence do in fitness?
Artificial intelligence is a broad term for systems that perform tasks such as recognising patterns, generating language or selecting actions. Machine learning refers to methods that learn patterns from data rather than requiring a person to specify every decision rule.
In fitness, these methods can help classify activities, organise training records, estimate aspects of movement or offer suggestions. Their value depends on the actual task and how well the system performs it. A sophisticated algorithm answering the wrong question is less useful than a simple tool addressing a real need.
Several technologies can sit inside the same app. Knowing which one produces a feature helps you ask the right questions about its reliability.
| Approach | What it does | A fitness example |
|---|---|---|
| Rules-based automation | Follows explicit instructions written by people; it need not use machine learning. | Suggests an easier session when a predefined condition is met. |
| Statistical prediction and machine learning | Fits relationships in data to estimate an outcome or classify an observation. | Estimates future performance from a defined set of training records. |
| Computer vision | Extracts information from images or video, often with learned models. | Estimates body landmarks or counts visible repetitions. |
| Generative AI | Produces text, images or other content from learned patterns and supplied context. | Explains a training term or drafts a schedule to review. |
| Recommendation and optimisation | Selects among options according to objectives and constraints; methods vary. | Chooses a session that fits available time and a planned training emphasis. |
“Personalised” also has several meanings. A programme might simply substitute exercises according to equipment, adjust rules using your recent results, or use a model trained on many people. These approaches can all be useful, but the label alone does not tell us which is happening.
Ask a concrete question: What work does this feature perform, and how will I know whether it helps? Saving time on a reliable summary is a legitimate benefit. It does not need to become a claim that the software understands every aspect of your body.
From recorded data to a training decision
Consider an app that changes tomorrow’s workout. Before the recommendation appears, several things must happen: data are recorded, those data are processed, the system interprets them, and a rule or model chooses an action.
Each step can introduce a different kind of error. A sensor might miss part of a session. The app might treat the missing period as inactivity. A readiness model might interpret that incomplete record as unusual recovery. The final recommendation can sound confident even though its starting point was uncertain.
Observation
“The recorded pace was slower than usual.” Check the route, weather, device and purpose of the session.
Interpretation
“Fatigue might explain the change.” This is one possible explanation, alongside others.
Decision
“Keep, reduce or change the session?” The answer also depends on symptoms, goals and consequences.
A model can find an association without identifying its cause. For example, if low training volume frequently occurs before a recorded injury, the model might use it as a warning signal. But an athlete may already have reduced training because discomfort had begun. That pattern does not show that low volume caused the injury.
Likewise, identifying a characteristic associated with injury does not prove that changing that characteristic will prevent it. Prediction research and intervention research answer different questions.
A useful output should say which observations informed it, what is missing, and what uncertainty remains. NIST’s AI Risk Management Framework treats validity, reliability, transparency and the context of use as central considerations.[1] For a reader, this becomes a simple habit: ask how the system got from the record to the recommendation.
There should also be a workable response when information is insufficient. “Please confirm this entry” or “I cannot assess that from this recording” can be more helpful than another precise-looking score.
What your training data can—and cannot—tell you
A training record can contain external measures, such as distance, duration, speed, repetitions and resistance, alongside internal responses, such as heart rate and perceived effort. The first describes aspects of the task; the second describes aspects of the person’s response.
Neither captures the whole experience. A steep walk with a backpack, a technical practice session and an easy cycle may have similar durations while placing different demands on the person. The context gives the numbers meaning.
| Input | Potential use | Important interpretation limit |
|---|---|---|
| Duration, distance and pace | Describe the amount and pattern of recorded activity. | Terrain, conditions, pauses and device processing affect comparisons. |
| Sets, repetitions and resistance | Track a particular strength exercise over time. | Technique, range of motion, rest and exercise setup also matter. |
| Heart rate | Describe cardiovascular response during suitable recording conditions. | Sensor error and context can change the reading; it is not a complete measure of training difficulty. |
| Heart-rate variability | Track a selected measure under a consistent protocol. | Device, recording method and physiological context matter; a single value does not diagnose recovery. |
| Sleep estimates | Help notice broad patterns in recorded sleep timing and duration. | Device estimates of wakefulness and sleep stages have measurement limitations. |
| Perceived effort and notes | Record how a session felt and why it changed. | Use consistent questions and scales; the numbers are subjective information, not physical units. |
| Video | Review visible movement and selected estimated features. | Camera view and model assumptions limit what can be inferred. |
Good inputs are more useful than a larger collection
In a study of wearable optical heart-rate sensors, accuracy varied with the device and activity.[2] This illustrates why a measurement needs validation under relevant conditions. It does not mean every current device performs like the older devices studied.
Sleep tracking also involves inference. A laboratory comparison of seven consumer devices found differences in detecting wakefulness and classifying sleep stages.[3] Use a sleep graph as one source of information, alongside your schedule, experience and daytime functioning.
Session-RPE is one simple monitoring method: session duration in minutes multiplied by a rating of overall perceived exertion using a consistent 0–10 scale. For example, 40 minutes at an effort rating of 5 gives 200 arbitrary units. This score can organise observations; it is not a measurement of tissue stress, energy expenditure or injury probability.[4]
Keep units explicit. Do not combine kilometres, minutes, kilograms and readiness points as though they were interchangeable. Before importing a log, check duplicate sessions, incorrect dates and activities recorded under the wrong name. Add a short note when a session was deliberately easy, interrupted or performed in unusual conditions.
Readiness, fatigue and recovery
Readiness scores typically compress several inputs into a single label or number. That makes a dashboard convenient, but it can hide the difference between the underlying measurements, the weighting chosen by the developer and the action suggested to the user.
Two apps can interpret the same night differently because their definitions and methods differ. A score of 70 has no universal physiological meaning. It should not be treated as a direct reading of muscle repair, hormone balance or the safety of a planned workout.
Heart-rate variability needs a defined method
Heart-rate variability describes variation in the timing between heartbeats. Optical wearables infer timing from pulse signals, which introduces additional measurement considerations. A small free-living validation study found that agreement with ECG-derived measures depended on recording conditions.[5]
For personal monitoring, keep the device, metric, time and posture as consistent as the protocol requires. Compare like with like. An unfamiliar reading after travel, illness, changed medication or an unusual schedule needs context; it does not identify the explanation by itself.
Research has tested training guided by HRV. A small study in trained cyclists explored this approach, but favourable changes within the HRV-guided group did not establish significant superiority between groups.[6] That is a narrower finding than “HRV-guided training is always better.”
Use several relevant observations
A practical check can include how you feel, whether there are new symptoms, recent training and how a familiar warm-up feels. The dashboard can contribute to this conversation. It should not cancel information that was never entered into the app.
For example, if a device reports excellent readiness while a movement causes sharp pain, the score does not provide clearance to continue that movement. Conversely, one unexpectedly low score with no other concern is a reason to check the inputs and context, not proof that the entire programme has failed.
Persistent unexplained fatigue or declining performance deserves appropriate assessment. The ECSS/ACSM consensus describes overtraining syndrome as a clinical problem requiring consideration of other explanations; no consumer readiness number establishes that diagnosis.[7]
The useful question is not “How can I make today’s score higher?” It is “What information would help me make a sensible decision about today?”
What does an injury or performance prediction mean?
A prediction is only interpretable when its target is clear. “Injury risk” could refer to any pain complaint, a missed training session, a medically diagnosed condition or a specific time-loss injury. These outcomes are related, but they are not interchangeable.
The prediction horizon matters too. Estimating whether an event will occur next week is different from estimating whether it will occur at any time during a season. A credible description identifies the population, outcome, time horizon and information available when the estimate is made.
TRIPOD+AI provides reporting guidance for studies that develop or evaluate prediction models, including those using machine learning.[8] Clear reporting allows readers to inspect what was actually done; it is not a certificate that a model works in every setting.
Ranking and probability are different achievements
Discrimination concerns whether a model tends to give higher scores to cases with the outcome than to those without it. AUROC is one common measure of this ranking ability. An AUROC of 0.80 does not mean that 80% of people will receive a correct diagnosis or that every displayed probability is accurate.
Calibration concerns agreement between predicted probabilities and observed outcomes. Among comparable prediction windows assigned about 10% risk, a well-calibrated model would produce roughly that event frequency across many observations. A model can rank cases reasonably while overestimating or underestimating their probabilities.[9]
Patterns are not personal verdicts
An estimate describes uncertainty within a particular model and dataset. It does not reveal a fixed destiny, and it cannot account for every future event. Contact with another player, a changed surface or an unrecorded symptom may matter without appearing in the inputs.
The same care applies to performance forecasts. A predicted race time depends on the conditions and assumptions behind it. A point estimate without an uncertainty range can suggest more certainty than the evidence supports.
Before acting, ask what decision the prediction is supposed to improve. A model that helps a coach decide whom to speak with first has a different role from a system that automatically removes someone from training.
Why a convincing prediction can still produce false alarms
When a model turns a score into an alert, two questions matter: how often does it identify an event that occurs, and how often does it raise an alert when no event occurs?
Sensitivity is the proportion of actual event cases that receive an alert. Specificity is the proportion of non-event cases that receive no alert. Neither tells you, by itself, how likely an alert is to be followed by the event.
The frequency of the outcome—the base rate—also matters. The following calculation uses invented numbers to make this relationship visible.
| Model output | Event occurs | No event occurs | Total |
|---|---|---|---|
| Alert issued | 16 true positives | 98 false positives | 114 alerts |
| No alert issued | 4 false negatives | 882 true negatives | 886 non-alerts |
| Total | 20 event windows | 980 non-event windows | 1,000 windows |
The model finds 16 of the 20 event windows. But it also raises 98 false alarms among the 980 non-event windows. Only 16 of 114 alerts, about 14%, are followed by an event. Four events occur without an alert.
This is not evidence about any product or a suggested threshold for sport. It is arithmetic showing why a sensitivity figure alone cannot tell a reader what an alert means. The example treats the windows as a simplified evaluation set; real athlete data require attention to repeated observations and changing circumstances.
“Accuracy” can hide the problem
A system that never raises an alert would be correct in 980 of these 1,000 windows: 98% accuracy. It would also miss every event. That makes overall accuracy a poor headline for this hypothetical task.
Moving the alert threshold changes the balance between missed events and false alarms. A lower threshold will generally produce more alerts. Whether that trade-off is useful depends on what happens next.
An extra conversation may be a reasonable response to an uncertain signal. Automatically cancelling many sessions has different consequences. Decision-curve analysis was developed to examine the potential value of prediction-informed decisions while accounting for trade-offs between true and false positives.[10] Such analysis still does not replace evidence about the real intervention.
There is another complication after deployment: an alert may lead to an effective intervention, so the event does not occur. That makes a simple before-and-after tally difficult to interpret. Evaluation needs to consider both the prediction and how people responded to it.
How to judge the evidence behind a model
Developing a model means fitting it to data. Evaluating it means asking how it performs on information that did not determine its final form. Confusing these tasks can make a system appear more capable than it is.
Keep the test genuinely separate
Data leakage occurs when a model benefits from information that would not legitimately be available for the intended prediction. It can happen through future observations, preprocessing performed using the entire dataset, or overlap between training and test records. Research on leakage documents how these mistakes can produce overly optimistic scientific results.[11]
In sport, thousands of records may come from relatively few athletes. Randomly dividing those records can place the same person’s closely related sessions on both sides of an evaluation. That does not necessarily test performance for a new athlete.
The split should match the intended use. Predicting future sessions for known athletes calls for a suitable time-based evaluation. Claims about new athletes require evaluation that holds athletes apart. These are original examples of how the general problem applies to a training database.
Look beyond the development setting
External validation evaluates a model on separate data relevant to its proposed use—for example, from another cohort, club or setting. A later period can also test performance over time. The details matter more than the label “validated.”
PROBAST+AI is a structured tool for assessing prediction-model quality, risk of bias and applicability. Its domains include participants and data sources, predictors, outcomes and analysis.[12] These considerations help distinguish an impressive score from evidence that can support a real decision.
Compare with a baseline
Does the model improve on a simpler method using the same information and evaluation conditions?
Check relevant groups
How does performance vary across the people, sports and circumstances for which the tool is offered?
Measure the consequence
Does using the output improve a meaningful outcome, with acceptable burdens and harms?
A prospective study can collect data forward in time without testing whether acting on the model helps. An impact evaluation examines the consequences of using it. For an injury-prevention claim, the important outcome is whether a defined approach reduces relevant injuries or their burden, not merely whether a risk score separates groups.
Also check the version. A study of one algorithm, sensor or release does not automatically validate a later version. Responsible updating includes checking for changed performance and retaining a clear account of what was evaluated.
AI coaching and programme design
An automated coach can reduce the effort of planning. It might organise sessions around available days, offer exercise substitutions, explain unfamiliar terms or summarise a training block. These are useful functions when the outputs fit the person and can be checked.
Before asking for a plan, describe the goal in practical terms. “Build enough endurance to enjoy a longer weekend walk” gives a different starting point from “prepare for a particular race.” Include your current routine, experience, available time, equipment and preferences.
A plan needs more than exercise names
Review the proposed frequency, duration, intensity and amount of work. For resistance exercise, examine sets, repetitions, resistance guidance and rest. For endurance exercise, look at how easier and harder work are distributed. Include warm-up, transitions and recovery when judging whether a session fits its time limit.
A plan can look varied while repeatedly loading the same tissues or crowding demanding sessions together. A substitution also changes more than a name: replacing cycling with running, for example, changes the activity’s demands and the skills required.
Language fluency needs verification
Generative AI can produce incorrect claims, invented references and internally inconsistent instructions. NIST’s Generative AI Profile describes confabulation and related information risks.[13] Clear prose is therefore a reason to read carefully, not evidence that the content has been checked.
When a recommendation depends on a source, open that source and check whether it exists, concerns the relevant population and supports the claimed conclusion. Ask the system to separate information you supplied from assumptions it added.
A planning prompt you can adapt
“Help me review a general fitness schedule. My goal is [goal]. My current routine is [routine], and I have [time and equipment]. Keep these constraints: [constraints]. First summarise what you know and identify important missing information. Offer a small number of workable options. Explain the purpose, effort and time required for each session, including rest and transitions. List assumptions and show what changed from my current routine. Do not invent performance results or treat a missing entry as a completed workout.”
This is an illustrative prompt, not a validated training protocol. For rehabilitation or a health condition that affects exercise, the plan needs the relevant professional input. Avoid asking a chatbot to diagnose a symptom or independently override an established care plan.
A useful review ends with a schedule you understand. You should be able to explain what each session is for, how to adjust an impractical detail, and what would justify seeking help.
Progress, plateaus and useful adjustments
An adaptive programme changes in response to information. That information might be a completed session, an updated goal, a missed week or a repeated difficulty. Adaptation can be implemented with simple rules or with machine learning; its usefulness depends on the quality of the decision.
The aim is to preserve a coherent training direction while responding to meaningful changes. Constant novelty makes it harder to compare sessions, learn movements and understand which change produced a result.
Distinguish a plateau from ordinary variation
A single slower run or difficult set does not establish a plateau. First check whether the comparison is fair: the same task, similar conditions, consistent measurement and a relevant timescale. Then consider adherence, recent demands, sleep opportunity and whether the programme is still directed toward the goal.
For a strength exercise, more repetitions only mean the intended progress if resistance and technique remain comparable. For endurance, a similar pace at a different effort or on a different route needs interpretation. The dashboard should preserve these details rather than collapse every session into a win or loss.
Change something you can evaluate
Suppose a person keeps missing a long midweek session because work finishes late. An AI assistant could help shorten that session or move it to another day. The problem is practical scheduling; generating a more demanding programme would not address it.
Suppose another person repeatedly finds the planned final repetitions much harder than intended. Checking the load, rest, technique and recent training may reveal a useful adjustment. The appropriate change is individual; there is no universal percentage increase that every app should impose.
These are original examples, not tested protocols. Their purpose is to make the decision visible: identify the problem, choose a proportionate change, and observe the result before changing several more variables.
Let maintenance count
A person may deliberately maintain strength or endurance while caring for family, managing a demanding season or enjoying other activities. A system should allow this goal. Progress does not have to mean increasing every number indefinitely.
Keep a brief change log: what changed, why it changed and what you will review. An assistant can organise that record. This makes a training adjustment easier to understand later, when memory alone may blur the details.
Computer vision and exercise feedback
Computer vision can estimate visible body landmarks, recognise an activity or count repetitions. More elaborate systems combine video with biomechanical models to estimate three-dimensional movement and aspects of the forces involved.
Kinematics describes movement, such as position, joint angles and velocity. Kinetics concerns forces and moments. A camera does not directly measure internal joint force; estimating it requires additional assumptions and modelling.
The 2023 OpenCap study combined videos from two or more smartphones, machine learning and musculoskeletal simulation, then compared selected outputs with laboratory reference methods.[14] It demonstrates a useful research approach to making movement analysis more accessible. It does not validate every app that draws a skeleton over a video.
Match the feedback to what was observed
Camera position, perspective, lighting, clothing and obstructed body parts can affect the information available. A visible angle in a single image is not necessarily the true three-dimensional joint angle. If the relevant movement cannot be seen, the system should acknowledge that limitation.
A repetition counter can be useful without being qualified to diagnose an injury. Likewise, identifying an asymmetry does not establish its cause, whether it is harmful, or whether changing it will improve the person’s outcome.
Choose one understandable cue
For ordinary exercise practice, a useful review might focus on a specific, visible feature relevant to the task. Record under suitable conditions, review the video at a sensible moment, and compare the cue with how the movement feels and the intended technique.
For example, a person checking whether a repeated movement reaches a consistent position may benefit from a clear camera view. A blanket instruction to force every joint into one supposedly perfect alignment is a much broader claim.
Do not divide attention between a demanding movement and a screen that needs continuous interpretation. Where technique, loading or symptoms are uncertain, a qualified professional can help assess what the recording leaves out.
What research has actually demonstrated
“AI works” is too broad a conclusion for a field containing many different technologies. A more useful account describes the system, comparison, population and outcome. The following examples illustrate different kinds of evidence.
Answering common exercise questions
A 2026 study compared ChatGPT 3.5 responses with answers supplied by nine personal trainers to common exercise questions. Blinded evaluators rated the chatbot higher overall on six of nine questions.[15] This supports a potential role in informational assistance under the study conditions.
It was not a trial showing better strength, endurance, adherence or injury outcomes in people following AI coaching. It also evaluated a particular model and question set. Answer quality and the ability to guide someone through months of training remain different outcomes.
Encouraging short bouts of walking
HeartSteps tested contextually tailored activity suggestions in a six-week micro-randomised trial that recruited 44 adults. Walking suggestions increased step counts in the following 30 minutes, with effects that diminished over time; the anti-sedentary suggestions did not show the same benefit.[16]
This is evidence about specific prompts and near-term behaviour. It does not establish lasting fitness improvements from every reminder app. Later HeartSteps work described a reinforcement-learning approach that learns when to offer activity suggestions.[17] That algorithmic work should be distinguished from the earlier trial and from modern conversational coaching.
Predicting running-related injury
A 2026 study followed 142 competitive endurance runners for 12 months and developed models using multidisciplinary information. The best reported AUROC was approximately 0.78. The authors identified the absence of external validation in a separate runner cohort as a limitation and called for trials testing whether model-informed feedback reduces injuries.[18]
This is promising prediction research. It is not proof that a consumer app can prevent an individual runner’s next injury, or that collecting every data type used in the study is worthwhile for everyday training.
Making movement analysis more accessible
OpenCap provides another example: evaluating selected movement estimates against reference methods is an important step toward useful measurement.[14] Further questions concern the intended population, the decision informed by the measurement and the outcome of acting on it.
These examples suggest a practical reading rule: carry the study’s boundaries into the conclusion. A better answer, a short-term behavioural response, a useful measurement and an effective long-term intervention are all valuable—but they establish different things.
Working with coaches, clinicians and everyday life
AI can contribute to a shared training process by reducing administrative work. A coach might use it to organise notes, identify missing records or prepare alternative schedules. A trainee might use it to formulate questions before a session. The person responsible for the plan still needs to review the output.
Before adopting a system, agree who sees the information, who interprets an alert and who can change the programme. A dashboard that sends an alert without a clear response arrangement can create worry without useful support.
Encouragement should preserve choice
Reminders can be helpful when they arrive at a workable time and offer a realistic action. They can also become irrelevant or tiring. Let the person choose their timing, frequency and tone, and provide an easy way to pause them.
A missed notification does not establish lack of motivation. The person might be working, resting, caring for someone, travelling or exercising without a device. These possibilities should remain visible in how a service describes behaviour.
App engagement is also different from participation in life. Opening an app repeatedly is not the same outcome as enjoying a walk, completing a useful strength session or feeling more confident about an activity. Choose measures that reflect the person’s goal.
Keep room for disagreement
WHO’s guidance on large multimodal models in health highlights automation bias: people may overlook errors or delegate difficult judgments because an automated system presents an answer.[19] In fitness, a practical safeguard is to make disagreement ordinary. The user should be able to correct a record, question a recommendation and request an explanation.
A good handover to a coach or clinician might contain the relevant training history, a brief description of the concern and the original observations. A dramatic AI-generated diagnosis is less useful than a clear account of what happened.
For example: “My usual session felt unusually difficult on three occasions, and here are the dates, conditions and changes I noticed.” That information gives a professional something concrete to assess. The system can help organise it without pretending to settle the cause.
Useful technology should leave people better able to understand their own training, communicate and make choices. Dependence on a daily permission signal is a poor substitute for that learning.
Privacy, fairness and control over your information
Training data can be personal in ways that are easy to overlook. A route can reveal a regular location. A video can include other people or the inside of a home. A written note may contain health information that is unnecessary for the planning task.
Start with a clear purpose and share the information needed for it. A general explanation of a training term does not require your medical history, precise location or full account archive.
Understand the service before connecting accounts
Ask who operates it, which organisations receive the data, how long information is retained, and whether content may be used to train models. Check the actual settings and terms; a reassuring answer generated by the chatbot is not a substitute for the service’s policy.
Consider what happens when you disconnect a device or close an account. Exported records, deletion procedures, backup retention and onward sharing are separate matters. Where possible, keep a usable copy of important training information in a format you can understand without the app.
An EU perspective on sensitive data
Under the GDPR, processing personal data requires an appropriate lawful basis; data concerning health receive additional protection under Article 9. Biometric data fall within that special category when processed for uniquely identifying a person—not simply because a fitness device records a bodily signal. Purpose limitation and data minimisation also matter.[20]
In employment, a ticked consent box does not settle whether consent was freely given. EDPB guidance explains how power imbalances can make employee consent inappropriate except in exceptional circumstances.[21] A workplace or team programme should make its purposes and consequences clear and provide meaningful ways to raise concerns.
Ask whose experience the model represents
A model developed in one population may not work equally well for people of other ages, training levels, body characteristics or abilities. Language, equipment and setting can also change the information available. Relevant evaluation should examine both performance and who is excluded from the service.
Fairness is more than a single average score. For example, does the app offer usable alternatives for someone who cannot perform its demonstration movement? Does a missed sensor reading lead to an unjustified label? Can someone who uses assistive equipment supply a meaningful record?
NIST’s framework treats harmful bias, accountability and context as parts of AI risk management.[1] For fitness services, these questions belong alongside prediction accuracy and convenience. The aim is technology people can use, question and leave without being reduced to an unexplained score.
Five steps for putting AI to useful work
You do not need to connect every device or automate an entire programme. Begin with one bounded task and decide what a useful result would look like. This framework is an original practical guide, not a clinical protocol or a validated scoring system.
-
Choose a meaningful job
Write down what you want help with: summarising a log, understanding a term, finding a scheduling conflict or preparing exercise options. Specify the benefit, such as saving planning time or making a programme easier to follow.
-
Check the inputs and boundaries
Use accurate records with dates and units. Include relevant constraints and mark missing information. Decide what the system may suggest, what needs review and what information you are comfortable sharing.
-
Inspect the output before using it
Check calculations, sources, assumptions and the total demands of a proposed session. Ask what changed and why. Correct errors in the original record so they do not keep influencing later suggestions.
-
Try a proportionate, understandable change
Use the output in a way that fits your experience and the task. A clearer log summary or revised calendar is easy to assess. More consequential changes to training or care call for appropriate professional judgment.
-
Review the real result
Ask whether the tool helped the chosen goal, how much attention it required and what went wrong. Keep what is useful, adjust what is awkward and stop features that add burden without a meaningful benefit.
| Ask | A useful response should make clear | Reason to investigate further |
|---|---|---|
| What exactly does this feature do? | The task, inputs, output and intended user. | Broad claims about “optimising everything” without a defined outcome. |
| What evidence supports it? | The studied version, population, comparison and relevant results. | A testimonial or accuracy headline standing in for an evaluation. |
| How are uncertainty and missing data handled? | What cannot be inferred and how users correct records. | Precise recommendations generated from incomplete or contradictory inputs. |
| Can I understand and change a suggestion? | Its assumptions, rationale and a workable review process. | A score that cannot be questioned or traced to relevant observations. |
| What happens to my information? | Retention, sharing, training use, export and deletion arrangements. | Vague policies or unnecessary access to unrelated personal information. |
| What does success mean for me? | An agreed benefit in training, understanding, time or participation. | More notifications or screen time treated as sufficient evidence of benefit. |
What would better future systems offer?
Promising directions include combining useful measurements more carefully, reducing manual entry, improving accessible feedback and processing appropriate tasks closer to the device. These are development possibilities, not guarantees of clinical benefit or complete privacy.
Progress also means better evaluation: clearer uncertainty, independent testing, useful comparisons, documented updates and trials of real outcomes. More data or a larger model cannot replace those steps.
A particularly valuable improvement would be knowing when to stop inferring. A system that asks a relevant question, identifies an unreliable recording or refers a concern to a suitable person may offer more practical intelligence than one that always produces another confident answer.
Common questions
Is every adaptive fitness app using machine learning?
No. Some adapt through predefined rules, others use statistical models or machine learning, and some combine several methods. Ask how the particular feature works and what evidence supports its intended use.
Can AI predict exactly when I will be injured?
A model may estimate a defined outcome over a specified period, with uncertainty. That is different from knowing what will happen to you. False alarms, missed events and differences between the studied population and your situation all matter.
Does high accuracy mean an injury model is useful?
Not necessarily. When events are uncommon, a system can be highly accurate by predicting that nothing will happen. Look at missed events, false alerts, probability calibration and the consequences of acting on the output.
Should I follow a readiness score when it conflicts with how I feel?
Check the underlying record and the context. A score cannot account for symptoms or circumstances it has not captured. New pain or other concerning symptoms require attention regardless of a favourable display; persistent unexplained difficulties merit appropriate assessment.
Can a chatbot write a useful workout plan?
It can help draft and explain options. Review the complete schedule, effort, volume, progression, time requirements and fit with your experience. A readable plan is not automatically a suitable one, especially where rehabilitation or health conditions affect exercise.
Can camera feedback tell whether my technique is safe?
It can describe selected visible or estimated features. It cannot establish safety from appearance alone. Camera conditions, model assumptions, the activity, loading and the person’s experience all influence interpretation.
Do I need a wearable to benefit from AI?
No. A simple, accurate log of sessions, effort and relevant notes can support useful summaries and planning. Collect additional information when it serves a clear purpose, rather than assuming more tracking must produce better decisions.
Will more detailed personal data always improve the advice?
No. Additional data can be irrelevant, unreliable or poorly interpreted. Share information that is useful for the task, understand how it will be handled and recognise that a system can still make mistakes after receiving a detailed history.
Can AI replace a coach or clinician?
It can assist with defined tasks. Its ability to produce information does not establish competence across observation, relationship-building, clinical assessment and accountable decision-making. Decide which work it can support and who reviews decisions that need professional expertise.
What is the best first use?
Choose a task you can check easily: organise a training log, explain a term or compare a proposed schedule with your actual availability. Define the intended benefit and keep using the feature only if it contributes to it.
Sources and further reading
Sources checked 6 September 2026. This educational guide draws on selected original research, methodological papers, professional guidance and official policy sources. It is not an exhaustive review or a ranking of commercial products.
Study findings apply to the methods, populations and outcomes evaluated. Prediction performance, short-term behaviour, answer quality and improved health or training outcomes are assessed separately. The hypothetical prediction table is an original arithmetic illustration; its numbers do not describe an actual system.
The planning prompt, practical framework and everyday examples are original illustrations. Individual symptoms, rehabilitation and medical exercise needs require appropriate professional assessment. Service features, privacy terms, software versions and evidence can change; the sources below make the basis and limits of specific claims available for review.
- National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. 2023.Voluntary cross-sector framework covering trustworthy AI and context of use. It is not evidence that a particular fitness algorithm is effective. 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 ↑
- Foster C, et al. A new approach to monitoring exercise training. Journal of Strength and Conditioning Research (2001), 15(1):109–115.Original validation work in cycling and basketball settings; duration multiplied by perceived effort is a monitoring score, not a direct tissue-load measure or an injury threshold. Return to text ↑
- Lam E, Aratia S, Wang J, Tung J. Measuring Heart Rate Variability in Free-Living Conditions Using Consumer-Grade Photoplethysmography: Validation Study. JMIR Biomedical Engineering (2020), 5(1):e17355.Ten healthy adults using an older optical device and portable ECG; pulse-interval correspondence depended on conditions. Does not establish the accuracy of all current resting-HRV systems. Return to text ↑
- Javaloyes A, Sarabia JM, Lamberts RP, Moya-Ramon M. Training Prescription Guided by Heart-Rate Variability in Cycling. International Journal of Sports Physiology and Performance (2019), 14(1):23–32.Small comparison in 17 trained cyclists after baseline monitoring; favourable within-group changes did not establish significant between-group superiority. Return to text ↑
- Meeusen R, et al. Prevention, diagnosis, and treatment of the overtraining syndrome: joint consensus statement of the European College of Sport Science and the American College of Sports Medicine. Medicine & Science in Sports & Exercise (2013), 45(1):186–205; joint ECSS/ACSM consensus.Professional statement on clinical assessment and exclusion of other explanations for persistent underperformance; not validation of a consumer score. Return to text ↑
- Collins GS, Moons KGM, Dhiman P, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378.Reporting guidance for prediction-model studies. Transparent reporting does not itself establish model effectiveness or clinical utility. 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 ↑
- Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Medical Decision Making. 2006;26(6):565–574.Original methodological paper on evaluating prediction-informed decisions. Decision analysis does not substitute for an intervention trial. 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 ↑
- Moons KGM, Damen JAA, Kaul T, et al. PROBAST+AI: an updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods. BMJ. 2025;388:e082505.Structured assessment of model development and evaluation. Not an endorsement of a specific sports prediction system. 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 ↑
- 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 ↑
- D’hoe B, Kirk D, Boone J, Colosio A. ChatGPT Outperforms Personal Trainers in Answering Common Exercise Training Questions. Journal of Sports Science and Medicine. 2026;25:235–261.Blinded evaluation of answers to nine questions using ChatGPT 3.5. The study did not compare long-term coaching outcomes or validate current models as a group. Return to text ↑
- Klasnja P, Smith S, Seewald NJ, et al. Efficacy of Contextually Tailored Suggestions for Physical Activity: A Micro-randomized Optimization Trial of HeartSteps. Annals of Behavioral Medicine. 2019;53(6):573–582.Six-week study that recruited 44 adults, examining activity following prompts. Suggestion types and effects over time differed; this was not an injury-prevention trial. Return to text ↑
- Liao P, Greenewald K, Klasnja P, Murphy S. Personalized HeartSteps: A Reinforcement Learning Algorithm for Optimizing Physical Activity. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies. 2020;4(1):18.Algorithm-development work on adapting activity suggestions. Distinct from the earlier HeartSteps trial and from generative conversational coaching. 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 ↑
- World Health Organization. WHO releases AI ethics and governance guidance for large multi-modal models. Official guidance announcement. 18 January 2024.Official account of health-sector guidance, including automation bias and oversight. Not direct evidence about every fitness app. Return to text ↑
- European Parliament and Council of the European Union. Regulation (EU) 2016/679: General Data Protection Regulation. Official legal text; Articles 5, 6 and 9; checked 6 September 2026.Primary EU legal text. Application depends on the processing and circumstances; the guide provides a limited educational overview. 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 ↑
Good technology makes the next step clearer.
AI can help us notice patterns, organise information and explore workable choices. Its lasting value is measured in what those choices make possible: more understanding, useful practice, confidence and participation.
Begin with a meaningful goal. Ask what the system knows, how that knowledge was tested and what remains uncertain. Keep learning from the actual result. That is how a powerful tool becomes a dependable part of training and life.