AI fitness coaches are here. The question is how to use them safely.

What readers reacted to

The comment section split in a revealing way. Some readers described AI coaching as inaccurate, flattering, or risky. One commenter called it a "fun, conversational way" to collect inaccurate information. Trainers warned that beginners, older athletes, and people with health conditions need experienced human eyes.

Others reported genuinely useful outcomes: diet plans that respected allergies, marathon and Ironman training plans, strength rebuilding after surgery, rowing workouts, and recovery routines. Several comments shared the same pattern: AI was helpful when people supplied context, kept a feedback loop, and stayed skeptical.

Useful AI is not generic advice

Most fitness advice is already easy to find. Any search engine returns thousands of training plans, periodization frameworks, and nutrition protocols. The value of AI is not another list of workouts. The value is translation: turning your actual logs, constraints, schedule, recovery state, and emotional friction into one next decision that applies to you, this week, given what happened last week.

For everyday athletes, that decision is usually not glamorous. Should I train today? Should I go easy? Should I rest? Should I stop chasing a pace that made sense five years ago? Should I preserve consistency instead of proving toughness? These are the questions that determine whether someone actually makes progress — and they are exactly the questions that generic advice cannot answer, because they depend entirely on individual context.

The lever is specificity. An AI system that knows you ran four days last week, averaged six hours of sleep, reported a soreness rating of seven out of ten on Thursday, and have a goal race in eleven weeks can give you a materially different recommendation than one that only knows you want to run a half-marathon. The difference between those two conversations is the difference between advice and coaching. Most people using general-purpose AI for fitness are operating in the first mode and hoping for the second.

This is why the useful AI fitness products will be built around data collection first, recommendation second. The model's reasoning is only as good as the context it has been given. And getting that context requires friction — logging, importing, answering questions — that most open chat interfaces do not ask for. The products that make that friction low and the output specific are the ones that will actually help people train better.

Where AI fitness coaching can fail

The NYT story also includes important failure modes: weekly totals that do not add up, plans that overweight an old personal best, taper advice that does not fit the athlete, and assumptions about heart-rate zones that may not match age, sex, medication, stress, heat, or individual physiology.

What a safer AI fitness coach should do

A safer AI fitness coach does not start with the goal. It starts with history. Workout logs — even rough ones — are the only honest answer to the question of where someone actually is versus where they think they are. The NYT piece illustrates this exactly: the chatbot's most useful moment was not the encouragement, it was the reality check that the writer's mileage had dropped and the base was not where memory said it was. That check is only possible if the model has seen the data.

Readiness signals come next, and this is where most general-purpose AI completely breaks down. Resting heart rate and HRV are not interchangeable — HRV measures the variation between heartbeats and is a more sensitive indicator of autonomic recovery than a flat resting HR reading. A night of poor sleep drops HRV before it shows up as fatigue you can consciously feel. A coach that sees your HRV trend over a week knows something about your recovery status that you may not have consciously registered yet. Soreness location matters too: generalized muscle fatigue is different from joint tenderness, which is different from the kind of sharp or localized pain that warrants stopping entirely.

Goal context is the third layer. The target race date, current weekly volume, and realistic weekly availability together define the training window. Without them, any plan is essentially a template with your name on it. With them, a model can calculate whether the goal is achievable at the current pace, whether the build rate is safe given the time available, and whether the first week should look like a base phase or a recovery week before any meaningful load begins.

The daily decision — train, easy, or rest — is where all three inputs converge. It is also where the model's conservatism matters most. The correct default when signals are ambiguous is not to train; it is to choose easy or rest. Overtraining injuries accumulate from repeated small miscalculations, not from single sessions. An AI that trends toward caution when the picture is unclear does less harm over a training cycle than one that defaults to action because the user came to it motivated.

Finally, the model needs a genuine escalation rule — not a disclaimer buried in a footer, but a live check that asks whether what the user is describing sounds medical rather than athletic. Persistent pain that does not resolve with rest, pain that changes location or character, pain accompanied by swelling or mechanical symptoms: these belong with a physical therapist or physician, and a well-designed tool should say so directly rather than offering a modified training plan as a workaround.

Where human support still matters

A good coach notices things that do not fit into a log entry. The way you describe a workout — whether you sound relieved it is over or hungry for more, whether you mention the same minor complaint for the third week in a row — carries information that a model working from structured inputs will miss. Experienced coaches also carry pattern recognition across many athletes: they have seen what overreach looks like in the weeks before it becomes an injury, and they can apply that knowledge to your situation without you having to articulate the problem yourself.

There is also a motivational dimension that AI handles poorly. The best coaches calibrate challenge to the specific person — knowing when to push someone who is underestimating themselves, and when to protect someone who trains through everything because stopping feels like failure. That calibration requires a relationship and a judgment call that goes beyond data. No model currently has a reliable way to distinguish productive discomfort from the kind of discomfort that is asking you to stop.

The realistic picture is that most everyday athletes do not have a coach, a physio, and a sports doctor available on a daily basis. They are already making training decisions with whatever is at hand — often ChatGPT, Garmin suggestions, a Reddit post, and memory. The goal is not to replace the humans who could provide better support, but to make the informal guidance people are already using more structured, more conservative, and more honest about what it does not know. That is a narrower claim than "AI coach." It is also a more achievable and more useful one.