Learn
Is an AI personal trainer any good?
Short answer: a good human coach standing next to you is still the best thing we know of, an app that guides each session keeps most of the benefit, and a plan written by a chatbot is usually accurate but often incomplete. That is what the published trials say, and HYLTH is built around them rather than against them. Here is the evidence, where AI coaching falls short, and what we do about each gap. We make no claims about HYLTH's own results, because nobody has studied them yet.
What the trials say about supervision
The clearest recent evidence is a 2025 randomised trial in the Journal of Strength and Conditioning Research. Seventy-nine trained adults followed the same ten-week, three-days-a-week programme in one of three ways: supervised in person, guided by an app, or self-guided from a PDF. Adherence was 88% with supervision, 81% with the app, and 52% with the PDF. Everyone got stronger, and the supervised group gained the most on the squat, but the app group's gains on the bench press were close to the supervised group's. The authors' conclusion was two-sided: in-person coaching produced the best results, and app-based training "shows promise for maintaining adherence, offering a viable alternative when full supervision is not feasible".
A 2023 trial in the Journal of Sports Sciences found the same direction in 36 trained young adults over eight weeks: the supervised group gained more muscle and more squat strength than the unsupervised group on an identical programme, and adhered better.
Read together: the plan on the page matters less than whether you do it. Supervision's biggest effect is on showing up, and an app that stands in for some of that supervision recovers most of the difference.
What the trials say about apps on their own
Apps that simply nudge people to move more have a modest, short-lived effect. A 2019 meta-analysis of nine randomised trials with 1,740 adults found smartphone apps added about 477 steps a day on average, a difference that was not statistically significant, and that apps worked better over periods under three months than over longer ones. A 2025 meta-analysis of 18 trials of standalone digital interventions found a small positive effect on physical activity, rated low-certainty evidence. Both reviews name the same weakness: people stop using the app, and the effect fades with them.
What the research says about chatbot-written plans
Two studies looked at plans written by ChatGPT. A 2024 study in JMIR Medical Education asked it for exercise recommendations for 26 different medical conditions and scored them against the American College of Sports Medicine's guidelines: 91% of the individual recommendations were accurate, but only 41% of the full prescriptions were complete, and most of the errors concerned whether the person should see a doctor before starting. A 2025 study in BMC Sports Science, Medicine and Rehabilitation had seven experienced strength coaches score twelve-week programmes from three versions of ChatGPT: newer versions scored better, the programmes were "aligned with established training principles", and the authors concluded that "the involvement of qualified human expertise remains essential" for plans to be safe and individualised.
So a chatbot plan is a reasonable first draft and a poor long-term coach. The gaps are the same in both studies: it does not know your history, it tends to miss progression and medical screening, and nothing makes it check its work.
What this means for an AI coach
Put the three strands together and the job of an AI coach becomes clear. It is not to write a better week than a chatbot; most weeks are fine. It is to do the things that made supervision work in the trials: know what you actually did, adjust the next week to it, keep you turning up, and refuse to do anything unsafe. And it should know when to hand you to a person.
What HYLTH does about each gap
- Knowing your history. H, HYLTH's AI coach, plans from what you logged last week, your check-ins and the sleep, heart-rate and workout data your watch shares, and rebuilds the week when things change. A chatbot only knows what you paste in.
- Showing up. HYLTH asks you to commit to the week, and a week counts when you do at least three quarters of what you committed to. Twelve such weeks in a row are rewarded. An uncommitted week is never counted as a miss, because the aim is a record of consistency, not guilt.
- Safety that is enforced, not suggested. Calorie targets have a floor and ceiling worked out from your own body, and the database refuses any target outside them. Nobody can override that, including H. H names its limits: it is an AI, it can be wrong, it is not medical advice, and it tells you when a question needs a doctor.
- Supervision when you want it. Every HYLTH client can move from H to a human coach in the app, with their whole history, and back again. The research says in-person coaching is the best option when it is available; HYLTH is built so that choosing it is easy rather than a fresh start.
- Your data. Nothing is sent to an AI provider until you have been told which company receives what and have agreed in the app, and under HYLTH's agreements your data is not used to train the provider's models. The privacy policy names every provider.
What we do not claim
No study has tested HYLTH or H. The trials above tested in-person coaching, generic apps and chatbots, and we have applied their findings to how H is built; that is a design argument, not an outcome. We cannot tell you that H will make you stronger or leaner, and we will not. HYLTH is a fitness and nutrition app, not a medical service; if you have a condition or an injury, speak to your GP before changing how you train or eat.
In short
An AI personal trainer is good at the parts that are about information and adjustment, and only as good as its design at the part that matters most, which is keeping you going. Judge one by whether it knows what you did, changes the plan because of it, refuses to let you do something unsafe, and lets you reach a person. That is the test HYLTH was built to pass.
Sources
- Gavanda S, Held S, Schrey S, et al. Optimizing resistance training outcomes: comparing in-person supervision, online coaching, and self-guided approaches: a randomized controlled trial. Journal of Strength and Conditioning Research. 2025;39(11):1129–1137. doi:10.1519/jsc.0000000000005216
- Coleman M, Burke R, Benavente C, et al. Supervision during resistance training positively influences muscular adaptations in resistance-trained individuals. Journal of Sports Sciences. 2023. PubMed 37789670
- Romeo A, Edney S, Plotnikoff R, et al. Can smartphone apps increase physical activity? Systematic review and meta-analysis. Journal of Medical Internet Research. 2019. PubMed 30888321
- Lee SA, Park JH. Systematic review and meta analysis of standalone digital behavior change interventions on physical activity. npj Digital Medicine. 2025. doi:10.1038/s41746-025-01827-4
- Zaleski AL, Berkowsky R, Craig KJT, Pescatello LS. Comprehensiveness, accuracy, and readability of exercise recommendations provided by an AI-based chatbot. JMIR Medical Education. 2024;10:e51308. mededu.jmir.org/2024/1/e51308
- Genç A, Aydın GR, Kasap M, et al. Comparative evaluation of ChatGPT versions in training program design: scientific approach, accuracy, and practical applicability. BMC Sports Science, Medicine and Rehabilitation. 2025. doi:10.1186/s13102-025-01409-7
Figures are quoted from the papers' published abstracts. Corrections: admin@hylth.co.uk.