Why is my AI FTP Detection lower than before?

[velonye]

The new ai version we’ve been forced into this week has dropped my ftp by 26 points overnight. How demoralising is that!

On Wednesday I did an easy 3 hr HR zone 2 ride (outdoors) and now TR says the next 3 days need to be “recovery”. It would generous to say TR is being a bit over cautious here. The “TrainNow” used to be really useful but now it only shows endurance rides which is not what I want*.

If I do all the sessions TR planned for me then according to the new TR ai my ftp will increase by zero.

*I appreciate that ai can be useful to guide and recommend but this current iteration that TR have imposed feels more like a straight-jacket than a helpful coach ..

This all seems very odd

(First draft of this post was written at the height of frustration and annoyance whilst trying to get the training session I wanted. This is the edited/sanitised version of the post - out of respect to fellow TR users)

[mrtopher1980]

Metrics on new AI FTP Detection + TrainerRoad AI

Hey ya’ll! We’re slowly rolling out TrainerRoad AI to people today. There’s been some confusion and misinformation going around, so I wanted to start a new thread to try to clear some of that up. What is the new AI FTP Detection? The updated AI FTP Detection you’re getting puts you around a level 3 threshold level. This will be consistent between detections, and we think this is a good insert point into your overall zone progressions to get faster. The old AI FTP Detection was predicting …

Be best.

[velonye]

What?!? Your automated reply is incomplete and is nonsense. It doesn’t address the issues I have raised.

[Macu88]

I’ve got the opposite situation, my FTP rose by 20 watts to a figure I never actually got close to in the past. I was expecting that maybe the whole trick is that the prescribed workouts will be easier, but when I look at them, they don’t seem easy enough to give me a chance to finish them.

[mrtopher1980]

I mean it literally provides you with a link to a thread started by the ceo of the company that makes the software you are using and addresses the basis for the concerns you have. There are several others plus 2 podcasts with the answers as well.

Sorry you missed the be best joke…

[Yeager139]

Change is hard, trust the process and give it a chance.

[James_Sanderson]

Can you show a screenshot of your calendar. Have you read any stuff that has been posted by the TR staff. And have you actually tried it. Or just upset your FTP isn’t actually what you want it to be?

Also Nate mentioned about have nice normal thread titles. Instead of WTF’s in them.

Can anyone help explain my FTP drop …

[Nate_Pearson]

Your FTP was probably a bit too high before. I know it might feel demoralizing, but we’re recalibrating your training to make you faster.

Here are some recent rides that the AI is tuning in on that make it think your FTP isn’t 205.

Lowest level sweet spot. 3 min repeats at 178w.

Low level Vo2 max. 50 seconds repeats at 236w

The FTP change is from 205 to 182, a 23-watt decrease.

At this new FTP, your threshold intervals would now be 4x8 @182 to start.

So to compare, you were doing 3 min at 178w at a moderate, and now we’re suggesting you do 8 min intervals at 182w and we think that should feel pretty hard/very hard territory.

This is a good example of your FTP going down, but the workouts getting harder. Remember, it’s all about the watts!

You have a heavily modified plan with lots of extra non-cycling work that is impacting the cycling workouts.

Looking back at your training, you’ve keep almost the exact same FTP for one year. You’re 64, so that might be a win. But what the AI is seeing now is that with the same training you’re going to get the same result.

In the first two weeks of the plan, you’re just doing endurance because you’re getting scaled back, and there’s a recovery week. So that’s probably another reason your FTP is expected to remain the same.

I would give more space between rides and more recovery, and see how that goes. I think the AI is behaving as it should for your account, based on what I’m seeing.

Also, we are enforcing the be excellent to each other rule and including ourselves in that. The post you initially did is inflammatory. I’m going to edit it for you. Hope this all helps in your training!

[velonye]

Nate,

I have read the article you directed me to and to associated articles. What is presented has serious logic and data flaws:

1. Logic and Reasoning Issues

1.1. Circular Claims Without Independent Benchmarking

• The article asserts that the “updated AI FTP Detection…puts you around a level 3 threshold level” and that this is “consistent” between detections. However, it does not explain why level 3 threshold is the correct physiological target. It assumes that the chosen threshold level correlates with an optimal FTP magnitude without independent benchmarking against gold-standard physiological measures (e.g., lactate threshold or critical power). This creates a circular logic: the system recalibrates to what it itself defines as correct.

1.2. Conflating Consistency With Accuracy

• The article implies that consistency between detections (i.e., shrinking variance) equates to greater accuracy. Consistency alone doesn’t guarantee that the number reflects true physiological threshold or performance capacity; it can produce a stable but biased estimate. Machine-learning systems can be consistently wrong. Without external validation (e.g., comparison to independent FTP tests or physiological markers), consistency is not proof of correctness.

1.3. Use of Performance on TrainerRoad Workouts as Proxy for Correct FTP

• The article claims that failure rates and expected outcome alignment improve after the new detection. But validating a model solely on its own internal workout structure (i.e., comparatively fewer failures) does not prove improved physiological measurement—just improved fit to the platform’s own prescription logic. In other words, the model may be better integrated with TrainerRoad’s internal expectations, not objectively more accurate FTP.

1.4. Qualitative Feedback Presented as Evidence

• The FAQ includes anecdotal beta tester quotes and aesthetic language about “AI nailing what you’re able to handle.” Anecdotes do not substitute for controlled validation. Objective model evaluation requires statistical measures (e.g., correlation with criterion tests, error distribution), not selective user experience narratives.

2. Data and Methodological Issues

2.1. Lack of Transparency on Data and Validation Protocols

• The article does not provide any quantitative metrics (mean error, variance, confidence intervals, bias measurements) to describe how much closer new detections are to true FTP. It instead shows only relative improvements in “failure rates” of internal model predictions. It is unclear how large the underlying dataset is, what the ground truth is, or how results were statistically validated. This lack of detail undermines claims of data-driven improvement.

2.2. No External Baseline Comparisons

• Unlike the previous TrainerRoad announcement, which touted comparisons with traditional tests (e.g., 20-minute FTP or Ramp Test), this article does not quantify performance relative to those. Without a baseline, readers cannot assess whether the updated AI FTP detection is objectively better or just less internally inconsistent.

2.3. Dataset Bias and Concept Drift Risks

• Machine-learning systems often suffer from concept drift when real-world performance patterns evolve over time, especially when based on training data aggregated across diverse individuals and settings. The article does not address whether the model continuously adapts to changing distributions or if older training data bias persists, potentially reducing real-world validity.

2.4. Performance Metrics Based on Model Output, Not Physiology

• TrainerRoad’s internal metrics (e.g., “failure rates” of workouts) are proxies for training difficulty, not direct measurements of FTP accuracy. In scientific validation, FTP detection accuracy should be measured against independent physiological tests and performance outcomes, which the article does not present. This is a limitation in how the model is evaluated.

2.5. Overfitting to TrainerRoad Environment

• Because the training and validation data are from TrainerRoad’s own user base and workouts, there is a risk that the model overfits to specific training patterns or platform usage behaviors that are not representative of broader cycling performance. Overfitting can produce excellent internal predictive power while performing poorly outside the specific dataset context.

3. Communication and Interpretation Issues

3.1. Lack of Clear Definitions

• The article uses terms like “level 3 threshold level” and “expected outcome improvement” without defining what those mean physiologically. This makes the claims harder to interpret scientifically.

3.2. Marketing Language Vs. Evidence

• The tone blends product marketing with supposed data insights. For example, “your FTP doesn’t mean you’re stronger/weaker—it’s like buying a new bathroom scale that is more precise.” This analogy simplifies a complex physiological construct and obscures methodological nuance.

3.3. Selective Framing of Results

• Figure captions and descriptions emphasise relative improvements while ignoring absolute performance benchmarks or potential negative cases. Balanced reporting should include model limitations, error distributions, and where it fails, especially given community feedback (e.g., unusual FTP shifts, unanticipated decreases in expected performance in real use).

4. Summary of Key Flaws

Logical Flaws

• Equates internal consistency with accuracy.

• Uses internal outcome improvement as evidence of validity without external benchmarks.

• Relies on qualitative feedback instead of quantitative validation.

Data/Methodological Flaws

• Does not present statistical evaluation against independent benchmarks.

• Potentially subject to overfitting and concept drift.

• Evaluates performance through platform conformity, not physiological fidelity.

Communication Flaws

• Marketing framing detracts from scientific clarity.

• Missing definitions and transparency on metrics.

• Lack of balanced reporting about failure cases.

[James_Sanderson]

The irony of criticizing the time and effort trainer road have put into their AI by replying with a load of text generated from an AI chat bot.

[Flashpoint51]

Don’t forget them dismissing the first attempt to help them as ‘an automated reply’

[Pbase]

My impression was that they didn’t realize they are on a public forum with other users trying to help them. They assumed the reply was from TR support. Anyway, it was enough to make me not want to help, so I’m super impressed that Nate put so much time into trying to help someone who almost certainly is just about to flame out.

[kevistraining]

@Pbase

I’m super impressed that Nate put so much time into trying to help someone who almost certainly is just about to flame out.

Same.

From what I can make out their only issue is that their FTP is set lower now? Or am I missing something?

The AI generated text is a low blow - assume they don’t know this is a public forum

[LacticacidMCB]

@velonye

What?!? Your automated reply is incomplete and is nonsense. It doesn’t address the issues I have raised.

Post your training plan and be nice to fellow users. Nobody can help you if you don’t give any information.

[kevistraining]

I just noticed that the OP only got their new FTP yesterday (22th) and hasn’t done any workouts with it.

They must have literally only just received the new lower FTP and then went straight to posting inflammatory posts on the forum. When was the time for support to respond to the support tickets?

The LLM AI generated post which must have had the prompt along the lines of “list all the possible issues with this article” is well out of order when the CEO of the company has reached out to try and help.

I don’t know how Nate does it.

I’m going to have to bow out of this one because I know I’m not being excellent….

[velonye]

Thank you for the link

The first draft of my post was written at the height of frustration and annoyance whilst trying to get the training session I wanted. I have now edited/sanitised the post - out of respect to fellow TR users.

[velonye]

Thank you for the check - The first draft of my post was written at the height of frustration and annoyance whilst trying to get the training session I wanted. I have now edited/sanitised the post - out of respect to fellow TR users.

[txrr9876]

The FTP decrease actually makes sense when looking at the calendar.

1. The plan is to “build endurance” NOT FTP. Building endurance and building FTP are two different things and shouldn’t be conflated.
2. There’s not a deload week in any of these calendar screenshots. Typically a deload week should be a 20-40% reduction in TSS from peak. Adaptations occur when the body can recover. When there’s day after day of stress thrown at the body it has a had time recovering and adapting.
3. Age impacts recovery. At 64 you need to give your body more time to recover.
4. The combination of multisports within the same day is probably doing more damage than you think. Strength + Bike is effective but do you really need bike+strength+row in the same day?