Can ChatGPT Write Your Workout Plan? What the Studies Found

Collage of a man in a gym reading his phone while a personal trainer stands beside him with arms crossed, with the headline "AI Workout Plans"

ChatGPT can write a workout plan that looks sensible, and in a 2025 study where 10 experienced coaches graded AI-written programs, overall quality came out moderate, with intensity targets often missing and progression advice generic. Other expert reviews found plans that were realistic but needed modification and lacked individual detail. Treat an AI plan as a first draft, check it against a short list, then adjust it with your own training log.

This is general education, not medical advice. If you have an injury, a medical condition, take medication that affects exercise, are pregnant or postpartum, or are under 18, check with a qualified clinician or coach before following any plan, AI-written or not.

What the studies actually tested

Most of the evidence is on older chatbot versions, hypothetical or single cases and small expert panels. Headlines tend to blur two different questions: can a chatbot answer exercise questions, and can it build a program that fits you. Here is what each paper did.

  • Answering common questions. In a 2026 study in the Journal of Sports Science and Medicine (D'hoe et al.), nine personal trainers' answers to frequently asked questions were compared with ChatGPT-3.5's answers. Twenty-seven graders, nine scientific experts and 18 trainers, initially blinded to ChatGPT's involvement, scored correctness, comprehensibility and actionability. ChatGPT's overall mean was higher on six of the nine questions, and the authors report no trainer answer rated higher on any question or metric. Look at the questions, though: five were about fat loss, and the others covered frequency, motivation, training around pain and time of day. That is question answering, not programming for an individual, and the authors note ChatGPT may lack personalization.
  • Twelve-week strength programs. Washif et al. (Biology of Sport, 2024) asked GPT-3.5 and GPT-4 for 12-week programs for hypothetical 20-year-olds, and the five authors, all with strength and conditioning experience, compared them with NSCA guidance and other literature. They found realistic output that may need modification. GPT-3.5 used the same 90 to 120 second rest across phases with different goals, the male and female programs were objectively similar, and the chatbot offered no real-time adjustment or revision based on feedback or progress.
  • Coach ratings across chatbots. Havers et al. (Biology of Sport, 2025) had 10 coaches rate plans from GPT-3.5 (via ChatGPT and Copilot) and Google Gemini, using 27 criteria on a 1 to 5 scale for a fictional advanced trainee. Overall quality was moderate, and the models differed on only a few criteria. Agreement between the coaches was low (Fleiss' kappa near zero), so ratings of the same plan varied a lot.
  • Newer model versions. Genç et al. (BMC Sports Science, Medicine and Rehabilitation, 2025) built seven prompts with input from seven experts, each with at least 10 years of experience, then compared 12-week resistance training plans from ChatGPT-3.5, 4o and 4.1 for a hypothetical healthy young man. Version 4.1 was more detailed, more internally consistent and better supported by the literature, but the authors still concluded that qualified human expertise is essential for safety and individualization.
  • People with health conditions. Dergaa et al. (Biology of Sport, 2024) had experts review 30-day programs GPT-4 wrote for five health profiles. The programs were generally safety-conscious but lacked specificity for the individual conditions and goals, and often leaned too hard on safety. The authors concluded that AI can serve as a supplemental tool but is not yet recommended as a substitute for professional prescription.
  • One person, three coaches. Li et al. (BMC Public Health, 2025) compared GPT-4 with three coaches on a 16-week plan for a single 24-year-old woman. GPT-4 scored higher on personalization (12.80 versus 11.53) while coaches scored slightly higher on effectiveness, safety and comprehensiveness, and none of the differences were statistically significant. A case study of one person is thin evidence in either direction.

Two limits apply to all of this. The tested models are already a generation or more behind what you can open today, and the authors of the JSSM study expect newer versions to do better. And none of the studies described here had anyone follow an AI plan for months to measure strength or muscle gained; experts scored the documents.

Where AI-written plans tend to fall short

The same weaknesses show up across papers. In the Havers study, the authors listed a lack of detailed intensity parameters, such as proximity to failure and load prescription, as a limitation across all the models. One hypertrophy plan fell back on advice like listen to your body, and its progression advice was generic, along the lines of add weight, reps or sets every one to two weeks. One hypertrophy plan included a strength-style day of three to five sets of three to five reps, which the authors said raises concerns about goal alignment, and one strength plan trained the squat, bench press and deadlift only once a week. The authors also described the plans as conservative, likely prioritizing safety over training effectiveness, which can mean too little stimulus for someone who has trained for years.

Consistency is a separate problem. A 2026 preprint on arXiv (Lee; a preprint, so not peer reviewed when posted) asked Gemini 2.5 Flash for the same exercise prescription 20 times in each of six scenarios, with the temperature setting at 1.0 to expose variability. The wording stayed broadly similar (mean semantic similarity 0.879 to 0.939), but the numbers drifted: variability showed up in quantitative components, particularly exercise intensity, and in resistance-training prescriptions, 10 to 25 percent of outputs had intensity wording that could not be classified. The author concluded that reliability depends substantially on prompt structure. It used a single model, an AI judge did the classifying without validation against expert judgment, and four of the six scenarios were clinical cases, so it speaks only indirectly to a healthy lifter's plan. Treat it as a signal, not a verdict.

The practical reading: the parts of a plan that matter most, how hard and how much and how to progress, are the parts a chatbot is most likely to leave vague. Your prompt has to force them into the open.

Google and Apple are building AI coaches into their health apps

AI-written workout plans are moving from a chatbot trick to a built-in app feature. TechCrunch reported in May 2026 that Google's Gemini-based health coach would leave preview on May 19 inside the rebranded Google Health app (formerly the Fitbit app), at $9.99 a month or $99 a year, with guidance based on goals, routines, equipment access and injuries. It was to be available first to select Fitbit and Pixel Watch users, with other devices to follow. Apple's plans are murkier: 9to5Mac, citing Bloomberg's Mark Gurman, reported in May 2026 that Apple's coach had been scaled back and was not expected until later in the iOS 27 cycle. That is a report, not an Apple announcement, so check Apple's own announcements for the current status.

The studies above tested general-purpose chatbots, not these products, so they say nothing about how well a branded coach performs. Slick packaging does not change the checklist below. For wider context on trends that sell well and trends that hold up, see fitness fads vs. facts.

A prompt that gets a better first draft

Given the papers, the aim is to give the chatbot what it lacks and demand the details it tends to skip. Include these in one message:

  • Goal and timeframe: for example, add muscle over 12 weeks, with a lighter week built in.
  • Training history and current numbers: months of lifting experience, plus recent working sets such as squat 60 kg for 8, 8, 7 and bench 40 kg for 8, 8, 6.
  • Schedule: days per week, minutes per session, and which days you can train.
  • Equipment: what your gym has and what it does not.
  • Injuries and exclusions: movements that hurt, and a note that any pain gets a professional opinion rather than a workaround.
  • Required output: for every exercise, give sets, a rep range, a target for reps left in reserve, rest time and a one-line progression rule. Add total weekly hard sets per muscle group, a rule for what to do after two missed sessions, and a deload rule.
  • Self-critique: finish with, list the three biggest weaknesses of this plan for someone with my history, and the assumptions you made.

Reps left in reserve is covered in RPE and RIR explained. Do not paste identifying details or medical records into a chatbot; a rough description is enough for a training plan.

A safety checklist for reviewing any AI plan

Go through the plan line by line before the first session. Five checks cover most of the failure modes above.

1. Volume

Count the hard sets per muscle per week and compare with what you do now. A 2017 meta-analysis (Schoenfeld, Ogborn and Krieger, Journal of Sports Sciences; 15 studies, 34 groups) found a graded relationship between weekly sets and muscle growth, so within the ranges studied, more weekly sets went with more growth, but it says nothing about how fast you can safely ramp. A plan that doubles your current sets is a red flag; ask for a gradual ramp. Also do the time math: 25 sets at roughly 40 seconds of work and 150 seconds of rest is about 79 minutes, which will not fit a 60-minute slot. For the numbers behind weekly sets, see how many sets you really need.

2. Progression

The plan should say exactly when to add load. A concrete example is double progression: when you hit the top of the rep range on every set, add the smallest jump available. Vague phrases like gradually increase are not a rule. Progressive overload explains the options.

3. Exercise selection

Check that each exercise matches your equipment, your skill and your goal. Every major pattern, squat or hinge, push, and pull, should appear, without five versions of the same movement. A hypertrophy plan built mostly on sets of three to five reps is worth questioning, since that is a strength-style scheme. If a movement is new to you, read the form notes on its exercise page, such as the barbell squat, before loading it.

4. Injuries and pain

A chatbot cannot examine you. Dergaa et al. found the programs for health conditions were safety-conscious but generic, which can look reassuring while missing what applies to you. Whether a movement is off limits with a given injury is a clinician's call, not the model's.

5. Rest and recovery

Check rest between sets against the goal, since the Washif team criticized one model for using the same rest in every phase. Check that the same muscle is not hit hard on consecutive days without a reason, and that a lighter week appears somewhere in the block.

Who should not rely on an AI plan

  • Anyone with heart, lung, kidney or metabolic conditions, or high blood pressure that is not controlled, until a clinician says what is appropriate.
  • Anyone recovering from surgery or an injury, or with pain that is new, sharp or getting worse.
  • Anyone pregnant or postpartum, and anyone under 18.
  • Anyone with a history of disordered eating or compulsive exercise; a chatbot has no way to tell whether more is a good idea for you.

Whatever your situation, stop and get medical attention for chest pain or pressure, fainting, severe or unusual shortness of breath, or sudden severe pain. No training plan is worth ignoring those.

Why your log is what makes any plan adjustable

The Washif authors noted that the chatbot offered no real-time adjustment from feedback or progress. That gap is not really the chatbot's fault: it only knows what you tell it, and many people tell it very little. A plan is a hypothesis about what you can handle, and a training log is the most direct way to test it. Without one, you cannot tell whether week four went well, whether a stall is a recovery problem or a programming problem, or what to ask for next.

A workable loop: log every set with weight and reps, and note effort and any pain in whatever way your log allows. Every two to four weeks, paste a compact summary back into the chat, such as the first and last week of your main lifts, missed reps and anything that hurt, and ask what to change. Then run its answer through the checklist above. If you stall, diagnosing a plateau from your training log shows what to look for first, and why tracking workouts works covers the broader case.

Easy Reps has no AI and does not generate or auto-adjust plans. It is a free workout log for iPhone, iPad, Mac, Apple Watch and Vision Pro with no account. The relevant part here is history: you can open any past session, filter by exercise, see charts and export to CSV at any time, so the data you give a chatbot, a coach or yourself comes from your logged sets, not from memory.

Common mistakes

  • Accepting the first output. The preprint suggests the same prompt can give different numbers on different runs. Regenerate once, and compare the intensity and volume.
  • Prompting with only a goal. No history, equipment or schedule gives you the generic plan the studies describe.
  • Skipping the log. Without data the plan never gets adjusted, so a mediocre plan stays mediocre.
  • Trusting confident wording. Fluent text is not validated text. In the Havers study, coaches still rated the plans moderate overall.
  • Asking a chatbot to override a clinician. If a professional told you to avoid something, a chatbot has no standing to say otherwise.
  • Starting at the plan's volume instead of yours. Begin near your current training and ramp up.

If you would rather start from a conventional structure than a prompt, the beginner strength routine covers the basics.

Bottom line

Studies of AI-written programs consistently conclude that expert oversight is still needed, and in the one that rated overall quality across chatbots, coaches scored it moderate. A chatbot is a decent drafting tool when you give it real context and ask it to state sets, reps, effort, rest and a progression rule. It is a poor substitute for a clinician when health conditions are involved. Check the draft against volume, progression, exercise selection, injuries and rest, start near your current training, and keep a log so that the plan gets adjusted by real results.

Sources