A ChatGPT workout plan is usually fine for a week and wrong by week 3. The structure, the exercises and the first week’s sets are sensible. It cannot see what you lifted, so it progresses you on a schedule written before you touched a bar: weight added after a missed week, no cap on a jump, no memory unless you paste your log.
This is for anyone who has asked a chat model for a program and is deciding how much to trust it.
Is a ChatGPT workout plan any good?
Give it a clear prompt and it will give you back something a coach would recognize. Three days a week, full body, a squat, a hinge, a push, a pull. Four days, an upper and lower split. The exercises are the standard ones, the rep ranges sit in the usual bands, rest periods are stated. Ask it to explain a choice and it will. For a first week, that is a good draft.
The first peer-reviewed look at this asked GPT-3.5 and GPT-4 for 12-week strength programs for a hypothetical intermediate lifter and compared the output with established coaching guidance. Both models produced three four-week phases with intensity rising across them, and the newer model varied rest periods by goal and adjusted its program to training status. The authors concluded the programs were plausible but “likely require additional fine-tuning before application”, and that the tool belongs alongside a practitioner (Washif et al., 2023, Biology of Sport). They also noticed the free model prescribed the same 90 to 120 seconds of rest in every phase, including the heavy ones: the kind of detail a coach fixes in a minute and a beginner never notices.
A model, then, is a competent writer of the first page of a program. The pages it cannot write are the ones that depend on you.
What did the research find about the quality?
A second study, in 2025, took the question to a panel. Ten coaches with exercise-science degrees rated muscle-growth and maximal-strength plans written by three chat models, from detailed prompts describing the same trainee, against 27 criteria. The overall quality came out moderate. The panel’s main complaint was the lack of intensity parameters: how close to failure to take a set, and what load to use. One strength plan trained the main lifts only once a week. Where the plans included methods such as drop sets, it was because the prompt had asked for them. The authors describe the models as tending to “prioritize safety over optimal effectiveness” and conclude that expert supervision remains necessary to refine the plans (Havers et al., 2025, Biology of Sport). Their prompts were already detailed, and they point to earlier work showing that plan quality rises with prompt detail: a one-line request gets a worse plan than a paragraph.
Both studies looked at models from 2023 and 2024, and newer ones write cleaner plans. The finding that survives the upgrades is the missing numbers, and those are the numbers that should come from your training in the first place.
What goes wrong by week 3?
The progression rule a plan should run on is well established. The American College of Sports Medicine’s position stand says to increase the load by 2 to 10% once you can do one to two reps over your target on two consecutive sessions (Ratamess et al., 2009, Medicine & Science in Sports & Exercise). Notice what that rule needs: the reps you got, on two specific sessions. It is a rule that reads a log. The two-session criterion is coaching practice written down; I know of no trial that has tested it against a fixed weekly increment. Most coaches would sign it. That is a different thing from proof.
A plan written in one sitting cannot read a log that does not yet exist. So the model writes a schedule in advance, “add 2.5 kg per week”, which is a guess dressed as a plan, or “progress when ready”, which hands the decision back to you with no number attached. Either way, by week 3 the plan and your reality have separated, in four ways.
It does not know what you lifted. Week 2’s weights were derived from week 1’s weights on paper; the sets you did never entered the calculation. If week 1 was too light, you have wasted two weeks. If it was too heavy and you cut reps, week 2 is built on a set that did not happen.
It adds weight after a week you missed. Tell a chat model you were ill and did one session of three, and it will usually say something kind and continue with next week’s loads as written. A missed week says nothing about whether the load was right, so the correct response is to hold, and a model that has not been told to hold will not.
It has no cap. Mistype 60 as 80, or log a sloppy set as a clean one, and the next number is built on the mistake. Nothing in the conversation says “that is a 33% jump in a week, stop”. A fixed rule with a cap would.
It has no memory. Every week starts from what is in the chat window. Even where a product remembers earlier conversations, it remembers what you typed, not what you lifted. Unless you paste the log each time, you are getting week 3 of a program for somebody who trained perfectly.
A three-week conversation, in numbers
Take a bench press the plan started at 50 kg (110 lb) for three sets of eight.
Week 1. You do 50 kg for 8, 8, 8. The plan adds a plate for week 2. Reasonable, and inside the ACSM band.
Week 2. You get 8, 8, 6. The last set fell short, so by the ACSM rule you repeat the weight. The plan, written three weeks ago, adds another plate for week 3.
Week 3. You catch a cold and train once out of three sessions. On Monday you type “give me week 4”. The model, which saw neither week 2 nor week 3, adds a third plate. You add “I was ill last week and missed two sessions”. It replies that this is completely normal, suggests you listen to your body, and leaves the number in the table. By now the schedule also needs the 1.25 kg plates, and in most commercial gyms those are the plates that are never on the rack.
A rule that reads the log would have moved once, after week 1, and then held twice: once because the target was missed, once because a single session is not a week. Three weeks in, the schedule is two steps ahead of the lifter, and the gap only grows. None of this is a failing in the model’s exercise science. Setting week 4’s weight is a task about the log, and the model is not holding the log. Had you pasted all three weeks and asked it to apply the ACSM criterion, it would likely have got it right. The answer is only as good as the record-keeping and the asking you do, every week, for as long as you train.
A plan written once, against a plan that reads your log
| Written once | Reads your log | |
|---|---|---|
| Knows the weight and reps you did | No | Yes, every set |
| Can hold a weight | No, it continues as written | Yes |
| Can lower a weight | No | Yes |
| Has a ceiling on one week’s jump | No | Yes |
| Memory between weeks | What you paste | The log itself |
| Has a lighter week built in | Only if the prompt asked for one | Yes, at the end of a block |
| Effort from you each week | Paste, re-prompt, check | Log the sets |
The right-hand column belongs to no product. A notebook and a coach who reads it does the same thing. A spreadsheet with a formula does most of it. The one thing it cannot be is a document written in advance, however clever the writer.
How to use ChatGPT for a workout plan well
Use the model for what it is good at and keep the rest yourself.
- Prompt in detail. Your days, equipment, years of experience, the lifts you know, any injuries, the goal. Write a paragraph. It is the cheapest improvement available.
- Take the structure and bin the schedule. Keep the split, the exercises, the sets and the rep ranges. Delete any column that says what the weight will be in week 3. It cannot know.
- Pick week 1’s weights yourself. A weight you can lift for the target reps with two or three in reserve. The model’s number is a guess at your body. Your first session is the measurement.
- Apply one progression rule. When you make the target reps across all sets in two sessions, add a small step, around 2.5% for a trained lifter, a little more for a beginner on a light bar. Our guide on how much weight to add each week gives the numbers; the next-week load calculator runs the rule for one lift.
- Missed most of a week? Hold. Same weights, no catch-up. Between a hold and a guess I would take the hold every time; a week at the same weight costs nothing, and a week too heavy costs the two after it. Write this down somewhere the model cannot override it.
- Keep the log yourself. Sets, reps, weight, every session. If you want the model’s opinion on next week, paste the log and ask it to reason only from that, then check its answer against step 4.
A good test of any plan: on a Monday, ask where the weight it just proposed came from. If the answer is not “from what you did last week”, it is a schedule, not a plan. The rest of our guides assume the second kind.
How BurnX splits the job
BurnX writes the block for your goal, days and equipment, then keeps the log and sets the weights from it. A lift’s weight goes up when you did the work last week, holds when you did not, and the Monday report says why. Nothing, the coach’s suggestions included, can pass a hard ceiling. How BurnX decides next week’s weights walks through a block in numbers.