Asking AI About Your Training Data: Why It Falls Flat Today, and What MCP Changes
Partager
You have probably tried this. You paste last week's mileage into ChatGPT, add your 10k time and your age, and ask what you should do next.
The answer comes back immediately, it is well-written, it is confidently structured, and it is almost useless. Not because the model is stupid — because you gave it a postcard and asked it to describe a country.
This post is about why that happens, what is changing, and the part most articles skip: what you should ask before you let any AI assistant near your training history.
Why asking a chatbot about your training disappoints
Four separate problems, and they compound.
It has no history
A coach's first question about a hard Tuesday is "compared to what?". Your 10k time means one thing if you ran it off twelve consistent weeks and something completely different if you ran it off three. A pasted summary carries none of that. The model is answering a question about a stranger who happens to share your numbers.
You chose the evidence, and you chose badly
This is the subtle one. You paste what you think is relevant — distance, pace, heart rate. You do not paste the lap splits that show your intervals falling apart in the back half, or the four nights of poor sleep that preceded it, or the fact that the route was 200 metres hillier than usual. The actual explanation is usually in the data you did not think to include.
It forgets
Most of these conversations start from nothing. You re-explain your race, your injury history and your weekly structure every time, and even then you are summarising rather than showing. Across a sixteen-week block, you are having the same first meeting forty times.
It would rather be helpful than honest
Given thin information, a language model does not usually say "I cannot tell from this". It produces a plausible, well-organised answer, because that is what it is good at. Thin input plus confident output is the worst combination available — it feels like insight and is actually a guess, and you have no way to tell which parts are which.
None of this is fixed by a better prompt. It is an access problem.
What MCP is, in plain English
MCP stands for Model Context Protocol. It is an open standard, published by Anthropic and freely available for anyone to implement, for letting an AI assistant read from an outside source of data.
The useful analogy is a plug socket.
Before standard sockets, every appliance needed its own wiring arrangement. Before something like MCP, every combination of app and assistant needed its own custom bridge — and since nobody builds a bridge to every assistant, almost none of them got built. MCP replaces that with one standard connection: a data source exposes a single socket, and any assistant that speaks the standard can plug into it.
Two things it is worth being precise about, because they get muddled:
- MCP is plumbing, not intelligence. It does not make an assistant a better coach. It changes what the assistant can see. The reasoning is still the model's.
- MCP does not grant access by itself. It describes how a connection works. What a given connection is allowed to do — read only, or read and change things — is decided by whoever builds it and by what you approve when you connect. That distinction is the whole of the privacy section below.
Support for the standard has spread beyond its original home, but which assistants support it, on which plans, and in which countries changes fast. Check the current documentation for whatever you use rather than trusting an article — including this one.
What becomes possible when an assistant can see your real history
This is the part worth getting excited about, and it is not "it writes you a plan". It is that the questions you actually have in your head become answerable.
| Your question | Answer from pasted numbers | Answer with your history in view |
|---|---|---|
| "Why was Tuesday so hard?" | A list of nine possible reasons: sleep, fuelling, heat, accumulated fatigue, illness… | Which of those actually applies — your splits fell off from rep four, it was 9°C warmer than your last version of this session, and you had already run 40 minutes more than usual that week. |
| "Am I ready to race in three weeks?" | General taper advice. | Whether your key sessions in this block were actually completed, how your paces at a given heart rate have moved, and where this block sits next to the one before your last race. |
| "What has changed since March?" | Nothing. It has no March. | A real comparison: easy-pace heart rate, how long your long runs have become, whether your hard days have got harder or just more frequent. |
| "Is this niggle showing up in my data?" | Advice to see a physio. Which is correct, but you knew that. | Advice to see a physio, plus whether your cadence or your stride pattern shifted in the fortnight before you first felt it. |
The pattern: with access, you stop getting a list of things it could be and start getting which one it was. That is the difference between an encyclopaedia and a conversation.
The questions to ask before you connect anything
Here is the part almost nobody writes about, and it matters more than the capability.
Your training history is more revealing than it looks. It is a continuous record of where you were, at what time, how often you are at home, how well you sleep, and — if health metrics are included — a reasonable proxy for when you were ill. Treat a connection to it with the seriousness you would treat a connection to your bank.
Before you approve anything, get a clear answer to each of these.
1. What exactly can it read?
Activities only, or health data too? Sleep? Location traces? "Your training data" is a vague phrase covering wildly different levels of exposure. The answer should be a specific list, not a category.
2. Can it write, or only read?
Read-only should be your default, and you should have to be convinced to go beyond it. The reason is simple: a read-only connection that misunderstands you produces a wrong answer, which you can ignore. A connection that can write produces a changed training plan, a deleted session or an edited history — and the mistake is now in your records rather than on your screen.
Write access is not evil and there are good reasons to want it eventually. But it is a decision to make deliberately, after you trust the read-only version, and never as an unexamined default on day one.
3. Can you revoke it, and how fast?
There should be a specific place where you see what is connected and switch it off, and revoking should take effect immediately rather than "within 30 days". If you cannot find that screen before connecting, do not connect.
4. Where does the data actually go, and is it kept?
There is a real difference between an assistant reading your sessions to answer one question and a copy of your history being stored somewhere new. Ask whether anything is retained after the conversation, and for how long.
5. Is it used to train models?
Policies differ by provider and by plan, and they change. This is a question with a findable answer — go and find it rather than assuming.
6. Which account is this, and who else can see it?
If you connect on a work laptop, under a work account, inside a workspace your employer administers, your resting heart rate may now live somewhere you did not intend. Mundane, common, and genuinely worth thirty seconds of thought.
7. Can you see what it did?
Some record of what was read, and of any change made, is the difference between a tool you can audit and a tool you have to trust blindly. Prefer the first.
A good rule of thumb for all seven: connect the smallest thing that answers your question. If you want to understand your running, you do not need to hand over three years of sleep data to find out.
What this will not fix
Access solves the input problem. It does not solve these.
Garbage in, confident garbage out. If your watch is working from an estimated max heart rate, every zone-based number downstream of it is wrong — and an assistant reading those numbers will reason impeccably from false premises. Worth fixing a max heart rate guessed from your age before you ask anything to interpret your training, along with a lactate threshold that has not updated in months. The same applies to readiness scores, which are composites of other numbers and inherit every error in them. If you are not sure which of your metrics are measured and which are inferred, start here.
It still does not know your life. Your data does not contain your job, your childcare, or the fact that you are dreading this race. That gap is the main argument for keeping a human in the loop, and access to more data does not close it.
You still have to know what to ask. An assistant with full view of your history and no question to answer does nothing. The quality of what you get back is still mostly set by the quality of what you asked.
A plausible answer is not a verified one. Even with real data in view, a model can misread it. If an answer surprises you, go and look at the session yourself. "It told me" is not evidence.
Where this is heading
The direction is clear enough. Your watch already measures far more than you can process, and the bottleneck has never been collection — it has been interpretation. A standard way for an assistant to read your actual history, with your permission and under your control, removes the most tedious part of that bottleneck: the copying and pasting.
What it does not remove is the need for something to decide what your data means for next Tuesday. Which brings this back to the line that runs through everything on this site: your watch is very good at measuring and quite bad at telling you what to do next. The numbers are inputs, not instructions.
That is the job Lactate Training does today — it syncs your Garmin activities and health data, builds and adapts a plan from it, and sets your zones from your lactate threshold rather than a guessed maximum, in 24 languages. Letting an outside assistant read that same history is an area we are actively working on; when there is something to announce, it will be announced properly rather than hinted at in a blog post.
Frequently asked questions
Can I connect my Garmin data to ChatGPT?
Not by simply asking it to — an assistant cannot reach your Garmin account on its own. Today most people paste summaries in by hand, or export a file, which is why the answers feel generic. Standards like MCP exist to make a proper connection possible, but what is actually available depends on your assistant, your plan and what the data source has built. Check current documentation rather than an article.
What is MCP?
MCP is the Model Context Protocol, an open standard published by Anthropic for letting an AI assistant read from an external data source. Think of it as a standard plug socket: instead of every app building a custom bridge to every assistant, a data source exposes one connection that any compatible assistant can use.
Is it safe to give an AI assistant access to my training data?
It depends entirely on what the connection is allowed to do. Prefer read-only, check exactly which data is in scope, confirm you can revoke access instantly, and find out whether anything is retained or used for training. Your training history includes location, timing and sleep patterns, so treat it as sensitive rather than as "just running data".
Why does ChatGPT give me generic training advice?
Because it is answering from the handful of numbers you pasted, with no view of your history and no memory of your season. It also tends to produce a confident answer rather than saying the information is insufficient — which makes a guess look like an insight.
Should an AI assistant be allowed to change my training plan?
Not until you trust it with reading. A read-only mistake is a wrong answer you can ignore; a write mistake is a changed plan or a lost session, and it is now in your records. Start read-only and go further only deliberately.
Does connecting an assistant replace a coaching app?
No, they do different jobs. An assistant answers questions you think to ask. A coaching app watches your data continuously and changes what you are going to do next without being asked. Both are useful, and neither is the other.
Or let something watch the data without being asked
An assistant answers the questions you think to ask. Lactate Training does the other job: it syncs your Garmin activities and health data, builds and adapts a plan from them, and sets your zones from your lactate threshold rather than a guessed maximum.