AI Told Me to Rest. My Data Says Work. I'm Testing Both.
On the morning of October 16, my HRV was 74. My resting heart rate was 55.
For context, my one-year average is an HRV of 89 and a resting heart rate of 47. That's a 15-point drop in HRV and an 8-beat jump in resting heart rate. If you track your own numbers, you know that gap isn't noise. Something is off.
I've spent the last few weeks letting AI help me read those numbers and plan my comeback. It's been useful in places. It's also changed its explanation at least five times in ten days, and it rewrote its plan the moment I pushed back.
So I'm doing what I'd tell any client to do when the advice keeps moving: stop debating and run a test. Plan A against Plan B, with rules set in advance, scored on my own baseline, out in the open. I don't know who's right yet. Here's the setup.
How I got here
On Labor Day I hurt my knee and took time off. It's healing well, and I'm working back to full operational status: classes, conditioning, the whole schedule.
During that comeback my numbers slid. From Labor Day through early October my averages held close to normal, with HRV at 87 and resting heart rate at 48. By the middle of October they were drifting the wrong way, and on Day 0, October 16, they landed at 74 and 55.
There's a known piece of physiology that may be part of this. When trained people stop or cut back their training, aerobic fitness starts to slip within a couple of weeks. One review of endurance athletes reports VO2max drops of around 4–7% within the first two weeks, with heart-rate responses to exercise shifting within the first month [1]. I didn't stop completely, and I'm not claiming that's the whole story. But a layoff followed by a ramp back up is exactly the kind of stretch where your numbers can get noisy.
What the AI did
I gave the AI my full Morpheus data, not a single screenshot. It had the weeks, the trends, and the context. Here's how its explanations moved:
Oct 6: sleep debt, then a late meal on Saturday night.
Oct 11: the late meal again. "That's your answer."
Oct 13: "I don't think the meal is the main cause anymore." Maybe early illness. Maybe delayed load from my Sunday double class.
Oct 16: hot yoga plus a heavy day. I pushed back and said rest wasn't working. It reversed to a volume-based plan: "I'm not going to argue it."
Those were the AI's guesses. Not one of them is an established fact about my body.
The clearest example: on Monday, October 12, my numbers were HRV 86 and resting heart rate 48. On October 13 the AI read that as a rebound before delayed fatigue arrived. On October 16 it read the same day as proof that "you respond well to volume." Same number, two opposite stories.
To be fair, it was genuinely good at reading screenshots, pulling out trends, and spotting patterns I'd have missed. But reading data and owning a decision are different jobs.
There's research behind why this happens. A 2024 study of several AI assistants found they often change a correct answer when a user simply asks "are you sure?", and frequently admit mistakes they never made. The researchers traced part of this to training on human preferences, where agreeing with the user tends to get rated higher [2]. That doesn't make AI useless. It means when the advice changes because you pushed, that's agreement, not judgment.
The two plans
Plan A is the AI's plan, as I've actually lived it since Labor Day. Less easy volume and a lean toward rest. Classes continued, and HIIT stayed in. That last part matters: Plan A was not a deload. Calling it one would be inaccurate.
Plan B is mine, and it started Saturday, October 17:
Zone 2 for 45–60 minutes on most days, using the heart-rate zones Morpheus sets each morning from my recovery
Classes as normal
HIIT capped at 1–2 sessions a week, only on good mornings
So the one real variable is added zone 2 volume. HIIT is in both plans.
Why zone 2? Because in my own history, when my HRV dips and my resting heart rate climbs, more easy aerobic work has been what brings them back. That's my experience, not a law of physiology.
And here's the honest part. The research doesn't crown zone 2 as the best tool for HRV. Several kinds of training can improve it, and one 2024 network meta-analysis actually ranked HIIT highest for RMSSD, the HRV measure most wearables report [3]. So Plan B isn't testing a proven protocol. It's testing whether my pattern holds for me, right now.
The rules
A test only counts if you set the rules before you see the results. Here are mine:
Data source: Morpheus.
What gets scored: raw HRV and resting heart rate as 7-day averages.
What doesn't: the app's recovery percentage.
Held constant: dinner 3+ hours before bed, no late-day heat, a consistent sleep window.
Check-ins: Friday, October 23 (day 7) and Friday, October 30 (day 14).
What counts as a win: back to baseline, or within 3 points of it. That means a 7-day HRV of 86 or higher and a resting heart rate of 50 or lower by day 14.
Why 7-day averages? Single-day HRV bounces around. Research on endurance athletes found that weekly averaged HRV tracked changes in fitness and performance far better than any single morning's reading, and that a few valid readings per week is enough to get a stable average [4]. One bad morning tells you little. A bad week tells you something.
Why not the recovery score? On October 15, my HRV was 83 and my resting heart rate was 51, both below baseline, and the app gave me a 94% recovery score. Composite scores like this blend several inputs with formulas manufacturers don't publish. A 2025 review of 14 of these scores from 10 brands found none disclosed how they're calculated, and almost none had peer-reviewed validation. The underlying metrics like HRV hold up much better than the scores built on top of them [5]. So I score the raw numbers.
A fair caveat on the baseline: my one-year average includes the Plan A weeks. It's a reasonable target, not a peak one.
The honest limits
This is one person over a few weeks, with a healing knee, while ramping back to full training. Plenty could move my numbers that has nothing to do with zone 2: the knee itself, the return to harder classes, sleep, stress, travel, or the simple fact that numbers often drift back toward average on their own.
What I'm running is a rough version of what researchers call an n-of-1 trial: testing an approach on one person, with outcomes defined in advance, to answer a question about that individual [6]. A real one would switch between plans more than once and blind what it can. Mine is a lot looser. If Plan B wins, it suggests an answer for me. It won't prove one for you.
The stop rule
This is the part I won't bend on, and it's set from day one: any symptoms, or a resting heart rate of 55 or higher for three mornings in a row, and I stop and get checked.
That's not paranoia. A sustained rise in resting heart rate and a drop in HRV are among the earliest signals wearables pick up when your body is fighting an infection. In a Stanford study of more than 5,000 smartwatch users, 26 of 32 people with COVID-19 showed a sustained spike in resting heart rate during infection [7]. A Mount Sinai study found HRV changes up to a week before diagnosis [8]. Your watch can't tell you why your numbers moved. It can tell you when to stop guessing and see a professional.
If Plan B doesn't move the numbers, my next step isn't Plan C from a chatbot. It's labs and a doctor.
Before you trust an AI with your health data
If you use AI to read your recovery numbers, run it through these questions:
Push back once, on purpose. If the advice flips just because you disagreed, it was agreeing with you, not judging.
Did it ask what it doesn't know? Good advice names its blind spots: illness, stress, sleep, the stuff a device can't see.
Does its explanation fit weeks of data, or just today? One morning is a weak story.
Who's your human check? A coach, a doctor, or someone who knows your body and will tell you no.
The takeaway
AI is a great reader of a screenshot and a poor owner of a decision. The answer isn't to ignore it, and it isn't to obey it. It's to test it against your own baseline, with rules you set before the results come in and a stop rule you respect.
Next update is Friday, October 23. Drop a comment with the metric you'd want me to track, and watch the full breakdown on YouTube.
Stay savage.
This is my personal experience, not medical advice. I'm not telling you to follow Plan A or Plan B. If your numbers change sharply or you have symptoms, talk to a qualified healthcare provider.
Sources
Barbieri A et al. Cardiorespiratory and metabolic consequences of detraining in endurance athletes. Frontiers in Physiology 2024;14:1334766. https://doi.org/10.3389/fphys.2023.1334766
Sharma M et al. Towards understanding sycophancy in language models. ICLR 2024. https://arxiv.org/abs/2310.13548
Yang F et al. Effect of exercise modality on heart rate variability in adults: a systematic review and network meta-analysis. Reviews in Cardiovascular Medicine 2024;25(1):9. https://doi.org/10.31083/j.rcm2501009
Plews D. The Practical Application of Heart Rate Variability: Monitoring Training Adaptation in World Class Athletes (doctoral thesis). Auckland University of Technology, 2014. https://hdl.handle.net/10292/7122
Doherty C et al. Readiness, recovery, and strain: an evaluation of composite health scores in consumer wearables. Translational Exercise Biomedicine 2025. https://www.degruyterbrill.com/document/doi/10.1515/teb-2025-0001/html
Lillie EO et al. The n-of-1 clinical trial: the ultimate strategy for individualizing medicine? Personalized Medicine 2011;8(2):161–173. https://pubmed.ncbi.nlm.nih.gov/21695041/
Mishra T et al. Pre-symptomatic detection of COVID-19 from smartwatch data. Nature Biomedical Engineering 2020. https://www.nature.com/articles/s41551-020-00640-6
Hirten RP et al. Use of physiological data from a wearable device to identify SARS-CoV-2 infection and symptoms and predict COVID-19 diagnosis: observational study. Journal of Medical Internet Research 2021;23(2):e26107. https://www.jmir.org/2021/2/e26107
Comments