We Didn't Make the AI Smarter. We Gave It the Answer Key

Open ChatGPT or Claude and ask it to build you an optimal training program. It will. It'll sound confident, organised, and authoritative. And depending on what it draws from, it might tell you to rest 60 seconds between sets, chase the 8-to-12 "hypertrophy zone" on every lift, and never eat more than 30 grams of protein in a sitting.

Every one of those is wrong. We've spent this whole blog taking them apart — the rest timer, the rep range, the 30-gram rule. And a general-purpose AI will repeat them anyway, with total confidence.

That isn't a knock on the AI. It's a knock on what it's reading. And it's the exact problem JSON.fit was built to solve.

The problem isn't the AI — it's what it reads

Language models and web search both rank information by prevalence. The more often something appears, the more it surfaces and the more confidently it gets repeated. In most domains that's a reasonable proxy for truth. In fitness, it's closer to the opposite.

The most-repeated training advice on the internet is also the most outdated — decades of bro-science, copied magazine templates, and SEO-farmed listicles all echoing the same myths long after the research moved on. So when an AI reasons from "the internet's consensus," it doesn't hand you the current science. It hands you the loudest version of the old science. Tell it to fact-check itself with a web search and it often just finds more of the same.

A general AI doesn't give you the current science. It gives you the loudest version of the old science.

This isn't really a failing of the model. Large language models are extraordinary at reasoning, synthesis, and applying a framework to your specific situation. They're unreliable at recalling niche, contested, frequently-wrong facts. Fitness is exactly that kind of domain — which means the trick isn't a smarter model. It's better source material.

A worked example: "Should I deload?"

Ask a raw AI whether you should deload and you'll usually get something agreeable and vague: "Yes, deloading is important — take it easier every few weeks to let your body recover." Pleasant, generic, and not actually actionable. It doesn't know your training age, your block length, or what a deload should concretely change.

Now here's what the same model produces when it's reading the file JSON.fit hands it: a deload is the final week of a block, volume cut 40 to 50%, load held, reps-in-reserve raised by two to three — and only if you've earned it by ramping volume first, with short beginner blocks getting none at all. (That's the actual research on deloads, condensed into a rule the AI can apply.) Same model, completely different output. One recites a vibe. The other applies a vetted protocol to your real program.

We did the retrieval so the AI doesn't have to

The fix is simple to describe and tedious to do: don't let the AI guess. When you fill out the questionnaire in JSON.fit, the prompt it builds for your AI doesn't just describe you — it carries curated, citation-backed research files on every lever that matters: rest, volume, proximity to failure, rep ranges, protein, fiber, meal timing, deloads, and more.

So the AI isn't reasoning from the polluted average of the internet. It's reasoning from a vetted evidence base we assembled and keep maintained. Its job shifts from recall — which it's bad at — to application, which it's brilliant at. If you want the technical name: it's retrieval-augmented generation, except we did the retrieval and the curation up front, instead of trusting a live, unvetted search to find the right studies in the moment.

Three things a generic search can't do

Once the AI is reading real files instead of guessing, it can do things a web lookup simply can't.

It branches on your situation. Ask the open internet "what's the optimal eating window?" and you get one answer for everyone. Our files force the AI to check who's asking — someone managing type 2 diabetes, a shift worker, or someone with a history of disordered eating each gets a different, safer answer, because that logic is written into the file.

It refuses to oversell its own topic. Every guidance file ends by ranking its own lever as secondary to the things that actually dominate — total volume, protein, energy balance. Most apps hype whatever feature they're selling. Ours explicitly instruct the AI not to overweight the very thing the file is about.

It tells you how sure to be. Which brings us to the part most apps would never put in writing.

The part most apps would never admit

Every research file is paired with a references file — and those references files grade their own evidence. Each claim is sorted into strong evidence, mixed or limited, or practitioner convention. Where a key study has a conflict of interest, we say so: the deload references note that one of the most relevant trials came from a research group with commercial ties to deload-based programming.

That honesty changes what the AI does with the information. The deload file openly states that the evidence deloads actively improve your results is thin — one trial in trained lifters even found a small strength drop. So the AI is told to treat deloads as low-risk insurance, not a miracle. It ends up calibrated instead of confidently wrong, which is the exact opposite of what a generic model does when it parrots whatever it read most often.

And because the files are public, you don't have to take our word for any of it. The guidance and the full reference lists sit openly at json.fit, for anyone to check. The entire value here is the curation — so we show our work.

This is what "bring your own AI" actually means

You're already paying for ChatGPT or Claude. We don't think you should pay a second subscription for a fitness app that wraps the same AI and bills you monthly — we've made that case before. What JSON.fit adds isn't another model. It's the thing the model is missing: vetted, current, auditable knowledge, fed straight into the prompt so the reasoning you're already paying for gets pointed at the right information.

Workouts are free. Nutrition is a one-time unlock, not a subscription. And the research that makes the whole thing work is free for anyone to read, whether you ever download the app or not.

The honest summary

A general-purpose AI is a brilliant reasoner pointed at a bad library. Ask it about training and it'll confidently hand you the most-repeated version of the field's oldest myths, because prevalence is what it optimises for. JSON.fit doesn't try to make the AI smarter. It gives it the answer key — curated, cited, evidence-graded research files that turn the model from a guesser into an applier — and then publishes those files so you can check them yourself. That's the whole trick. There's no catch buried in the marketing, because here the marketing is the methodology.