When AI Feedback Gets It Wrong — and What To Do

The coach told you to back off, or to add a threshold piece, or that your split last Tuesday was worse than it looked. You read the logbook entry again and the number the advice was built on was not yours. That is a more useful moment than it feels, because the question worth asking is not whether the coach is right but what the advice is standing on.

An automated coach reads a logbook. It can see the sessions you logged, the splits, and whatever you told it in chat. It cannot see the Tuesday you skipped because your shoulder hurt, the sleep you did not mention, or that your last 2k was a bad day in a hot room. Most wrong feedback is not a hallucination. It is a sound rule applied to the wrong person, or to a day that was not there.

You carry the missing half of the picture, and a few checks that cost nothing settle most of these disputes in under five minutes. This is how to run them, and what to do with the answer.

The context

Three different things get called wrong feedback, and they need three different answers.

The first is a premise problem. The advice is about a session, a split, or a ceiling that does not belong to you — a number taken from a population and put into your plan as though it had been measured on you. This is the easiest kind to catch, because the number is written down and you can ask where it came from.

The second is that the advice is technically right and still lands badly, because of what it does to you rather than what it says to your stroke. The Mageau & Vallerand 2003 motivational model in Journal of Sports Sciences argues that coaches who show autonomy support, who provide structure, and who involve the athlete feed that athlete's needs for autonomy, competence and relatedness ([1] Mageau 2003). The same correction, delivered without those three, arrives as dismissal. If your reaction to the feedback was irritation rather than information, that is not a reason to throw the feedback away. It is data.

The third is a signal problem: the advice rests on one session, or on a difference smaller than the instrument can resolve.

The core idea

Wrong feedback and unhelpful feedback are different problems, and only one is the coach's fault. The move that works is to stop asking whether the advice is right and start asking what number it stands on. Every piece of coaching advice has a number under it — a split, a session count, a heart rate, a weekly load. Find that number and check three things: is it yours, was it measured, and is the gap you are being asked to act on bigger than the error in the measurement?

Here are the five ways it goes wrong.

1. The goal was set from somebody else's number. The 220-minus-age rule for maximum heart rate is the clearest case. Tanaka and colleagues pooled 351 studies covering 18,712 subjects and found that age predicts maximum heart rate strongly (r = −0.90); their equation is 208 − 0.7 × age, and they conclude that maximum heart rate is predicted largely by age alone ([2] Tanaka 2001). A ceiling derived from your birth year is a population average wearing your name: reasonable as a starting point for a test, poor as a prescription for a session. The tell is a coach that can state the number but cannot say which measurement produced it.

2. An average was substituted for a distribution. This one is about the scale rather than the athlete. Borg's review points out that ratio-scaling methods such as percentages describe general perceptual variation, while category methods are more useful where you need to describe differences between individuals ([3] Borg 1982). A coach quoting "you are at 85 per cent of your maximum heart rate" is quoting a general rule. A coach asking for your category is asking about you. One session is also an average, and Halson's review notes that very few load markers have strong evidence behind them and there is yet to be a single definitive one ([4] Halson 2014). One hard Tuesday is weather, not a trend.

3. The load recommendation ignored what you have actually been doing. The clearest published demonstration that reading today's signal beats reading last week's plan is a small trial in runners. Twenty-six moderately fit men were randomised either to a fixed schedule of six sessions a week, two easy and four hard, or to a schedule prescribed each morning from that morning's heart rate variability: harder work when the reading was flat or rising, easy work or rest when it fell below its own ten-day reference or trended down for two days. The adaptive group improved maximal running velocity significantly more, though the between-group change in peak oxygen uptake was not significant ([5] Kiviniemi 2007). A small trial, in runners, over four weeks — and still the cleanest published case of a prescription that reads today rather than last week. The way to do this badly is arithmetic. The acute-to-chronic workload ratio, this week's work over the last month or so, survives re-analysis with a meaningless denominator because the ratio is simply a rescaling of the acute load, and neither the ratio nor acute load alone carried any meaningful predictive advantage over a model with no workload terms at all ([6] Impellizzeri 2021). If your load advice is built on that ratio, you are being handed arithmetic dressed up as a finding.

4. The cue does not match your stroke. Stroke rate sets how much time the whole cycle has, and the force takes 0.3 to 0.4 sec to build, at roughly 800 to 900 N on the oar ([7] Secher 1993). If your force collapses in the back half of the drive, a rate cap is something you are surviving rather than holding, and telling the rower to push the rate will not fix it. That is a technique conversation, worth having in those terms.

5. The confidence should have been lower. Halson's review is the reason to discount the tone: the importance of individualising load monitoring cannot be over-emphasised, and a monitoring system should be intuitive and able to report simple, scientifically valid feedback ([4] Halson 2014). If your coach cannot tell you which of your own numbers the advice came from, that is not a reason to trust it more.

And there is a reason to argue rather than simply comply. Feedback you control is worth more than feedback you receive on somebody else's schedule. In a putting task, Yamaguchi and colleagues gave one group knowledge of results only when they asked for it and the other group knowledge of results on a partner's schedule. The self-control group showed enhanced motor learning compared with the yoked group, and the gap was still there at a retention test a day later ([8] Yamaguchi 2026). Twenty-eight adults on a golf green is a long way from an erg, but this is the mechanism an automated coach gets wrong by default: it can fire unsolicited feedback at any moment, which is the yoked schedule.

The how-to

1. Read the body first. Sleep, soreness, illness, travel, heat, a new boat, a different machine. If the prescription assumes a normal day and your day was not normal, it is wrong for reasons that have nothing to do with the model. Say so before you look at any number. A coach will move a session for "I slept four hours" far more readily than for "your algorithm is wrong".

2. Ask what number the advice is standing on. One question: what is this based on? If the answer is a number, you can check it. If the answer is a general rule — because that is all the system can see — you have found the problem.

3. Run a second channel. You have three that cost nothing, and a fourth if the coach's plan is built on effort ratings.

   • The talk test. Comfortable speech is likely possible when intensity is below the ventilatory or lactate threshold, and not likely possible when it exceeds it ([9] Reed 2014). It will not survive hard intervals — the same review notes it may not be practical for high-intensity interval training — but it works for steady work, and it is the only one of the three you can use in a race.
   • Session RPE. After the session, score the whole thing out of ten and multiply by the minutes. Foster's method tracked training across both steady and interval exercise with a highly consistent relationship to heart rate, though the absolute score came out significantly greater than the heart-rate-derived one ([10] Foster 2001). That last detail matters more than it looks: your session-RPE numbers live on their own scale and will not match a stranger's.
   • Morning heart rate. Record it for five minutes each morning and read the trend, not the value. Buchheit's review finds five-minute resting recordings and a short submaximal average to be the most useful of these measures, and that for an individual athlete a change only means something once it is weighed against the error of measurement and the smallest important change the measure can detect ([11] Buchheit 2014).
   • Effort ratings, in scale. Scherr's cohort of 2,560 adults found perceived exertion correlated strongly with heart rate (r = 0.74), and on Borg's 6-to-20 scale an RPE of 11 to 13 is recommended for less trained individuals while 13 to 15 suits more intense but still aerobic work ([12] Scherr 2013). Note the scale. Eleven to thirteen on a 6-to-20 scale is not the same number as an RPE of 11 out of ten, and two coaches using the same word can mean different things.

4. Say what is wrong in the coach's own units. Not "that feels wrong". "My session RPE on Thursday was 7, not the 9 your prescription implied, and I have rowed this session several times at that pace." A coach reasoning from numbers can only answer a number.

5. Let it recompute, then do the recomputed session. Ask for the prescription again with the correction in it, and then row the corrected session rather than the original one. The point is not that the coach has to agree with you. It is that a coach which will not move has told you something you can act on too.

The trade-offs

Almost none of this literature is about indoor rowers. The adaptive-prescription trial was in runners ([5] Kiviniemi 2007), the feedback-control trial in adults putting golf balls ([8] Yamaguchi 2026), the rowing-physiology source in a racing shell rather than an erg ([7] Secher 1993). What transfers is the method — find the number, check the instrument, correct in the coach's units — not the thresholds.

And there are times to ignore the coach completely. Persistent fatigue that does not clear with rest, a sharp pain, illness: those are not calibration problems, and no analysis outranks a rower who feels unwell. Drop the session — a health decision, not a coaching one.

It is also worth being clear about what the coach is better at than you are. It never misses a session in the log, it never has a bad day, and it will notice that your easy pace has crept for eight weeks at a point where you had stopped noticing. That is worth having even when its targets need checking. How the system handles your logbook data is set out in our privacy notice.

What to do tomorrow

Open your last ten sessions and write down the three numbers the coach is most likely to be reasoning from: your recent average split, your weekly session count, your last recorded effort. If you cannot name them, that is the first thing to fix. Then score one session three ways — split, session RPE out of ten, and the talk test — and put them side by side: where they disagree, you have found the disagreement. Send one correction in the coach's own units, one number in one sentence, then ask for the prescription again with the correction in it and row the corrected session. And log your effort from now on: it is the cheapest number you will add to your logbook and the one most likely to contradict the split.