In 2026, if you pick a random people from the street, here’s one of the first things they’ll probably tell you: “Traditional machine learning models are no good for training prescriptions because they rely on correlation instead of causation, while choosing the correct training prescription is a thoroughly causal question.” No need to beat a dead horse here.

What about large language models (LLMs)? According to the most enthusiastic boosters, these models may have learned some fundamental truths about human physiology by ingesting the entire contents of the public internet. Therefore, LLMs may come up with great training regimes and novel workouts.

I have never used an LLM to plan my training because I have a library of stock training plans from reputable coaches that I modify based on my own needs. However, I did once try to implement an LLM based approach in an analogous domain, namely food. The idea was to develop an LLM chef that would generate high protein plant-based recipes that I’d cook, photograph and put on a website. Over time, I’d have a a stock of plant based recipes suitable for active people.

There was just one problem with this idea: the recipes weren’t any good. No matter the prompt or model I used, the end outcome was invariably a type of food that would be best described as “mushy”. It was mostly edible and probably contained the claimed macronutrients. The taste just wasn’t there.

Sometimes you might decide that mushy food is good enough for your purposes. Maybe you’re just trying to get the nutrition in and it’s already late and you didn’t have time to go to the grocery store. Sometimes you just want to tick an item on your checklist and it’s not that important how it happens.

For example, I recently painted the wall in one of my bathrooms. It wasn’t a high stakes project in any sense of the word. I just wanted to get it done relatively quickly and without causing too much damage. So I looked up some basic instructions from the Home Depot website, bought the suggested items, and painted the wall during one afternoon.

I could’ve asked ChatGPT, Claude or Gemini to provide me with instructions, and I’m sure they’d done a good job at putting together something similar to what I saw on the Home Depot website. I just didn’t see any reason for anyone to spin up any GPUs to produce instructions that I already knew existed on the Home Depot website, presumably written by some person who is relatively competent in painting.

Suppose you want a training plan to be able to finish your first marathon. There are hundreds of books and articles that contain such plans. Most of them are based on the following ideas: ramp up your volume slowly, keep most of your runs easy, and do one or two faster runs or interval sessions a week. You could also ask some LLM to generate a plan like this for you. Given that the principles of endurance training are well-known and formulaic, it’ll most likely do a competent job.

For the average would-be marathon runner, stock training plans and their LLM equivalents are going to be perfectly fine. Like the dozens of LLM-generated recipes that I tested, the training plans may be a little bit mushy, but they will largely get the job done. Based on my unsuccessful AI Chef experiment, however, I highly doubt that LLMs in their current form can come up with training prescriptions that are optimal or even particularly novel.

Here’s why. Consider how you come up with a good recipe for a high protein plant-based dish. First, you might start with the protein itself. That protein turns out to be chickpeas. Then, you might consider the flavours that go together with chickpeas, such as rosemary and olive oil. Next, you might start thinking about the style of the dish, which places you somewhere in Southern Italy. That’ll give rise to the idea of Orecchiette, lemon and garlic. And so on and so forth.

Presumably, in the background, your brain runs a process that shifts through a bunch of counterfactual ingredient combinations, such as chickpeas and raspberries. The vast majority of these combinations are rightfully never brought to your consciousness, thanks in no small to your past experiences of some heinous dishes you’ve “enjoyed”. But even out of those combinations that are in fact brought to your consciousness, most of them will never see the light of day. Maybe you don’t feel like a particular combination works (no capers with rosemary, thanks). Or you just don’t feel like eating a certain type of food today.

When you’re ready to finally cook something, you need to experiment to get it right. I don’t know about you, but for me it takes several attempts before I get a feel for a dish, even if I’m just following someone else’s recipe. And for professionals for whom stuff like this matters much more, I’m sure it takes weeks or months of experimentation to develop a great dish.

Still, it’s interesting that this process of counterfactual reasoning and experimentation tends to result in better food than reading every single publicly available recipe–and every other publicly available document for a good measure–which is what the LLMs are doing. How difficult can it be to “predict” what a good recipe will look like when you’ve seen them all? Apparently, it is very difficult if you cannot experiment, whether it’s experimentation by using your imagination or trying things out real life.

What’s this all to do with training prescriptions? Well, I don’t think coming up with a good training plan is that different from coming up with a good recipe. And even if it is, I’m certain that counterfactual reasoning and experimentation–whether implicit or explicit–is an absolutely key part of it. And this is why LLMs are not only bad chefs but also bad coaches, unless the taste of food or performance in sport is relatively unimportant to you, in which case the mushy stuff will be perfectly fine.