Will AI replace human coaches?

Triathlon Forum
1

TLDR: “AI” will never replace a coach for performance oriented athletes, but will play a role in dumb functionality (which may well be 90% of coaches’ work today) such as templated plans, fitting stuff into calendar, etc. and can be used as a powerful data-processing tool for smart coaches.

First, the term Artificial Intelligence is misleading. We are actually talking about Machine Learning, and more specific Neural Networks. To simplify, its a software that can go through lots and lots of data and look for correlations and patterns in it.

“But, but, but, ChatGPT understands English. Surely it is intelligent,” you might say.
Actually, LLMs like ChatGPT do NOT understand language. LLM is a software that takes textual data and processes the probability of each word coming after any other word, and the probability of these word to come in sequence in a sentence.
Now imagine doing this for the whole freakin’ internet. Everything that was ever written. Including all books. It takes a long time to crunch and costs over 100 million dollars a pop to run this beast (not to mention the pollution and Co2 emissions, but that’s a different topic), but Altman, Zuck and co. can afford it. All they need to do is say “AI”, and the market throws money at them.
Anyway, once you have these probabilities, you can give the LLM a sentence (e.g. “build me a weekly training plan”) and it will calculate what will be the probable next word, and the one after it, and so on and so forth. Add a bit of randomness, and magically, it comes out as English (most of the time). So no, it does not understand anything, hence no intelligence.
With the exception of a chatbot that will give you word of encouragement to motivate you to train (or kill yourself in the case of Charecter.ai), this is a bit of a sidetrack from coaching/training, but the same concept holds for dealing with numeric data.

Machine learning is incredibly powerful at processing tons of data and finding patterns.
Now imagine you give it twenty years of your training data. HR, power, speed, weight, sleep hours, etc., and ask it to figure it out (this is called unsupervised learning). What can it tell you? Remember, it doesn’t really understand the data. Maybe it can find patterns and suggest similar training to what you have done in the past, but that kind of sucks, you can copy/paste yourself. Maybe it can take templated training plans and adjust them to your schedule. Not much better if you really are trying to optimize performance. This is what most “AI” training platform provide, I’m guessing.

But what if you want it to analyze the exact type of training you have done which led to successful results. How would it know which training sessions contributed to which race? Also, one of the biggest problems with machine learning is the quality of the data. How would it know to ignore garbage data when your HR monitor was pace tracking?
For this you can use supervised learning - get an expert that will tell the model what it is trying to find (e.g. race goals), what are the relevant training sessions, thresholds for cleaning up data, etc.

This is oversimplification, but the point is you still need a coach that has intelligence and can apply judgment. “AI” does not understand your data, it just calculates probabilities, so without a human expert you can only get very basic outputs (which, may satisfy 90% of people).

2

I could give the thing 6 years of my data and ask it to give me a Training Plan and I’d expect it to output something pretty similar to what I’m already doing. Which is perfectly fine, and may save me some time and money, but who’s to say what I’ve been doing for the last 6 years is any good? Is it optimal for me? What will it do if I get injured, or if I’m feeling off?

Ok, so that’s just my data, so let’s feed it the data from every pro athlete who has ever logged data. Now am I expected to complete the same/similar workouts to a pro which it will undoubtedly assign? Is it good enough in it’s current state to ‘dumb’ it down to something an amateur athlete who only has 10 hours per week to train can do?

I work in IT (in software development to be precise) so I am exposed and use these tools on a daily basis. Right now, it’s great at basic tasks. Ask it to do something that has been done (correctly) thousands of times and it will do the job and do it well. Boilerplate, base functionality, templating, etc saves me so much time it’s unbelievable. Get more advanced, and more obscure and shit falls apart in a hurry. It’s so confidently incorrect it’s laughable, and it’s a big part of my job to guide my junior developers to utilize these tools but to not rely on them.

What I’m getting at is, could AI be used to template out a basic training plan? Sure. Could/should it be utilized by coaches to do some basic tasks/analysis to save time and money? Absolutely, and I wouldn’t be surprised if it’s not already. Is it good enough to replace coaches? In its current state, nope.

4 Likes
3

But isn’t this the strength of AI-it can review an immense amount of data for thousands of athletes and hundreds of thousands if not millions of sessions to see what works? Or is this still a long way down the road?

4

Works for who though? Clearly Frodo and Dan Lorang worked well as athlete and coach but when Lionel got dished out the same/similar sessions it didn’t work well.

I don’t see AI really every solving that question or any time soon anyway. Humans are unique and there will never be a ‘one size fits all’. The job of AI will be to analyze the subject it’s trying to train and then base a workout off that. So while the LLM may have hundreds of thousands or even millions of workout sessions, it lacks the data of the athlete.

Think about it for a minute. Let’s take a single block of training (4 weeks) and work through it. What’s the state of the athlete… Is it peak season or off season? Is it pre-race or post race? Is their day-to-day nutrition dialed? Are they fueling sessions right? Are they hydrated? Are they at altitude? Are they in a hot or cold climate? Do they feel good or are they fatigued?

20 athletes could be doing the exact same block of training but based on the questions above you could get 20 different outcomes from the same set of training. Even 1 athlete could repeat the same block of training 20 times and have 20 different outcomes based on those questions.

1 Like
5

All those metrics you mentioned could easily be captured by AI and taken into account?

What can a coach interpret than AI can’t? A coach sees you couldn’t complete a session, skipped a session, HR is up, HRV is down etc-what do they do? Why can’t a computer be trained to react the same way, though the result would/could be far more effective with the database of thousands of athletes and the result of that change?

As for Lionel he didn’t work because he doesn’t provide feedback to coaches, perhaps with AI he would even be more inclined to report ‘feeling 5/10 today’ etc

1 Like
6

Ah - thanks for opening up this thread (taking the discussion from the LS thread). Will just copy my latest reply here:

One BIG factor of why having a coach keeps you motivated, is the “commitment” or coach/athlete agreement, that both parties promise to serve in the relationship.
I have worked (in a prior life) trying to automate this to big groups of athletes, such as a e.g. 200 corporate people in cycling, and it is not easy (to automate). The minute the personal aspect leaves the equation and it becomes to machine/programmed, people are falling off. A real person (in this case a coach) looking over your shoulder, checking in on your motivation, well-being, injuries, tiredness, is a HUGE factor for many people.
Can this be simulated in the future?? Yes - it can, and it will be.

7

Yes but the issue is, that data wasn’t captured for the millions of workout sessions that have been logged over many years. Whoop/Garmin/Apple etc are all relatively new and I would gather have been adopted by a very small number of athletes.

I’m not saying you’re wrong, and in fact you’re right. If AI had access to all this data then it could do exactly that.

The thing holding AI back in this discussion is that it doesn’t have the whole range of data required to perform perfectly and I doubt it ever will. Both privacy laws and companies greed will keep the most important personal data under lock and key.

So the entire “garbage in, garbage out” saying with AI still holds true in this scenario.

8

Perhaps, though by that reasoning how many athletes must a coach train before his/her data becomes useful?

Companies like Athletica AI already have access to far more data than a normal coach might obtain in 5 lifetimes.

I might be looking at this far to simplistically but a coach doesn’t exactly have a super complex 4 year degree to become qualified to coach, and I don’t mean that in a insulting way, what I’m meaning is how much knowledge does a good coach need and how easily can this be uploaded into a decision matrix that gets better and better over time. Again I’m speaking from extreme ignorance so please excuse me as I’m sure it’s very very involved and more likely years of passive learning from being involved in triathlon.

For the coaches out there-how long do you spend reviewing an athletes week to build the following week/fortnight? How do you build up and adjust a training plan?

2 Likes
9

An experienced coach, working with a good library and platform (TP or similar) will spend 10-15 mins per week per athlete on planning/programming. If day-to-day communication is part of the service, this obviously is taking up more time, depending on the needs & situation of the athlete.
Plan is optimally (but far from always) weekly micro-adjusted up/down depending on tiredness, injuries, job/family.

1 Like
10

The issue is that sleep tracking ,
Hrv stress score etc is mostly shite data so you get shite results

to add it can be useful to show trends but i would say only a fool would really trust the data yet
and people that use it should study limitations of those devices and the algorithms that are used more and not just the marketing.

1 Like
11

I think where it boils down to is there is prescribers and there is coaches
Prescribers are people like mullen that try to have many people on the books
Coaches are the that have about 15 to 20 people on their books and with ai might few have more , and don’t argue about the number but it can’t be 50 I would say to still be considered coaching and it could be as low as 8 in real high performance coaching

As a coach you are also looking for to educate the athletes help to built a team , looking at stuff is the bike fit , biomechanic issues . And in a simple way remind the athelte to think should they do that hard workout today or not not rest based on algorithms but reality , do they need a rest despite training going well but outside training stress is higher than a app thinks , I would say if on average you dont communicate with an athlete at least 4 times a week, and more like some days twice that’s not coaching that’s prescribing

as a coach I would say you think more in the long term development of an athelte. IE if you take on an athlete who is unfit to do an ironman than that is not coaching for me , so for instance I wonder how long it will take for AI to be actually able to say, for you doing an ironman makes actually no sense.

When it can say you are better of to stay in short curse or long course taking into account if you stay short curse you might get federation support ,if you go long course you might find more sponsors, than it will be really dangerous for coaches

But the first to go are the prescribers , the coaches should have some time and as I said before when coaching goes than there won’t be that many jobs left in any industry. Which I guess we have to decide if that really makes sense to make 70 percent or so of humans irrelevant

So I guess as support coaches ai is great
if it takes over than overall their is likely going to be an issue as how are most people going to make a living if there is not need for them
My 2 cents

And ps IAM not trying to say to be a prescriber or getting prescribed plans is a bad thing as there is people that are very good prescribers and there is athletes that deal very well with prescription . and if one is a good prescriber and also good a bringing people together in a training environment that is pretty good .

The human aspect like club coaches I would think won’t be replaced anytime soon
Would I be wrong here I wonder ?

and since somebody mentioned a dr in the other thread if you just prescribe stuff for symptoms than thats not that great if you look for why the symptoms happen that is more valuable and i guess in coaching thats not so dissimilar i hope that makes a bit of sense .

2 Likes
12

Other than the social/motivational aspect, the coach’s strength is their ability to assess the skills and strength of their athlete and assign them the most efficient training for steady injure free progress

So replace them with AI? Why not but the fundamental law of AI is garbage in = garbage out

The true question for me is, should you use AI to output your training plan:

  • what AI will you use?
  • what dataset will you use to train it (inputs/outputs)
  • how will you test it?

I will make up my mind once the three questions above are answered

13

yes GIGO is a problem, however, if the data set is your historic reactions to training inputs there may be patterns that a human may not see that an AI can find. The test is obvious, do the results of the training show the expected improvements? The one area where AI may not be best is personalize feedback and motivation. But that may change with time.

AI is very good at finding patterns in large datasets, so that is a strength that is hard to match with human bias. If the data is your historic data then it may see unique responses to inputs and be able to modify its prescription based upon the results you get. As the old saying goes to a person that only has a hammer every problem is a nail. Some coaches only know certain things, and AI can be trained on large datasets, then use your particular data to customize just for you. See your responses and then adjust. I see it as being critical for that upper echelon (those after marginal gains) of training in the future.

2 Likes
14

Quote from other thread…

TrainerRoad pushes a response for this each workout. How do you feel? And you pick one of the 5 or 6 options. If you skip it, you have a reminder.

1 Like
15

As far as I understand, current AI solutions can:

  • write customised programs,
  • adjust programs as-hoc based on execution,
  • take into account non-performance wearables output like HRV or resting HR,
  • take into account athlete’s evaluation of how (s)he feels,
  • tailor the season with differently prioritised races (big picture).

Out of curiosity, do you know of any solution that could:

  • take into account blood testing values like hematocrit / hemoglobin, testosterone, cortisol, iron / ferritin, creatinine kinase etc.
  • account for different environmental conditions like altitude and temperature,
  • account for different shoes in which runs were executed?
16

I would think data is just data. It the AI will try to find patterns so if there is added data and it sees that it is indicative of outcomes then as far as I know it is agnostic to what that data is to it is just one more data column to parse. In the area of medical AI there are currently ideas that maybe by using targeted inputs (ie not all of the internet but rather medical texts and vetted data from journals) it will provide even better responses. Already as reported in a recent JAMA article the AI outperformed doctors using AI backup on standardized symptom questions with well defined known correct answers. The Doctors + AI were somewhere around 80% AI alone was around 90% correct diagnosis. AI can be very powerful when used well.

2 Likes
17

could you share a link to the article in your post ,please.

18

here is a discussion with one of the authors:
https://jamanetwork.com/journals/jama/fullarticle/2828679?guestAccessKey=ea9132a5-a774-41e6-9e97-ad8e0f423d29&utm_source=silverchair&utm_medium=email&utm_campaign=article_alert-jama&utm_content=olf&utm_term=122724&adv=000004836329
here is the journal article:
https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2825395

2 Likes
19

That is exactly the point. Stop thinking of “AI” as intelligent. It does not understand the concept of “a session that works” (in fact, it does not understand any concept including the data its processing).
You would need a human expert to clean, classify, identify the data and configure parameters for what “works” means. If you look at companies that successfully use machine learning, that is a huge part of what their data scientists do.

1 Like
20

Since I opened this can of worms :laughing: I’ll reiterate what I mentioned in the Lionel thread…

[quote=“tilburydavis, post:2306, topic:825621, full:true”]
I think that will come down to trust and social context. Some folks am sure will be quite happy to do exactly as prescribed. Many others and my suspicion is this cohort is where 80% pros will sit is that while they may be fully accepting of the ‘message’ coming from A.I. programme prescription wise they will still want the assurance and comfort of a human conveying that message respectful of their social context.

By that I mean if you had to convey message “X” to an American, a Japanese, a German or a Finnish all FOUR of those people would require subtlety different communication styles. Within leadership there is a great text on this by Richard D. Lewis called “When Cultures Collide”.

Interestingly in a recent speech by the Chief Economist of LinkedIn she said the no.1 quality sought in all jobs on their platform… “Communication Skills”.[/quote]

I firmly believe coaching is communication, “buy in”, empathy, nuturing caring or passion for betterment… training prescription is a simply a by product of this.

Some nice further points I’ve read recently going down this rabbit hole…

“If ways of speaking can alter ways of thinking, ways of thinking can alter ways of speaking as well. The dynamic interaction between the two is part of the ongoing story of how we try to make the world intelligible to us, and to make ourselves intelligible to one another” - Manvir Singh / New Yorker 2024

“What we map depends on where we look, what factors we choose to focus on, and what aspects of the terrain we decide to represent. Since these choices will shape the kind of map we produce, there is no perfect map of a terrain. Therefore, making sense is more than an act of analysis; it’s an act of creativity.” - Peter Senge

“If you really want to be great at something, you have to truly care about it. If you want to be great in a particular area, you have to obsess over it. A lot of people say they want to be great, but they’re not willing to make the sacrifices necessary to achieve greatness.” - Kobe Bryant

“Phenomenology, in simple terms, is the study of how we experience the world through our bodies, emphasizing perception, sensation, and the interconnectedness of mind and body. The body is the medium through which athletes encounter the world. Movement isn’t just something they do; it’s a way of being. When we recognize this, coaching becomes less about imposing solutions and more about guiding athletes to discover their own.” - Stu McMillan

A.I. isn’t nuturing that passion and care for one’s sport any time soon, nor are A.I. agents voicing motivation going to do it, humans inately thrive on ‘connection’ / ‘connectedness’ / a sense they are a part of something and elite athletes, in my experience, are constantly striving for a heightened sense of interoception (Interoception is the collection of senses providing information to the organism about the internal state of the body. This can be both conscious and subconscious) so that they can make better autonomous decisions in the moment racing and training.

I don’t see A.I. resolving this either any time soon, but have at my data analytics and prescription (assuming good ‘clean’ data being looked at / filtering erroneous data done) by all means :slight_smile:

1 Like