There’s some­thing mild­ly absurd about mod­ern run­ning tech. Every brand talks about AI now. Smart coach­ing, pre­dic­tive train­ing plans, readi­ness scores, recov­ery scores, stress scores, you name it. My watch appar­ent­ly knows my future. It just doesn’t know what hap­pened five min­utes ago.

Take heart rate. I run with a chest strap or the wrist sen­sor, doesn’t mat­ter. Every now and then the data goes com­plete­ly off the rails. Sud­den­ly my pulse jumps to 190 while I’m jog­ging easy, stays there for three min­utes, then drops back like noth­ing hap­pened. No hill, no sprint, no dra­ma. Just noise. Same with GPS. Clean route along the riv­er, then one glitch and the track cuts straight through build­ings like I tele­port­ed. The device shrugs and saves it as truth.

I can live with imper­fect sen­sors. Sweat, move­ment, bad satel­lite recep­tion — physics is messy. What I don’t get is why all that so-called intel­li­gence doesn’t clean up the mess after­wards. Because sta­tis­ti­cal­ly speak­ing, this is the easy part. Out­liers are not some exot­ic phe­nom­e­non. They’re text­book stuff. Every intro course in data analy­sis cov­ers them. If one data point is phys­i­o­log­i­cal­ly implau­si­ble or com­plete­ly detached from the sur­round­ing val­ues, you flag it, smooth it, or drop it. Sig­nal pro­cess­ing has been doing this for decades. Medi­an fil­ters, Kalman fil­ters, sim­ple plau­si­bil­i­ty checks. Noth­ing fan­cy. This isn’t sci-fi AI. It’s basic hygiene.

Garbage in, Glossy Insights out

Instead, most plat­forms skip the present and jump straight into prophe­cy. They’ll hap­pi­ly cal­cu­late my pre­dict­ed marathon time for next Octo­ber but won’t ques­tion a heart rate spike that would put me in the ER. They build elab­o­rate mod­els on top of noisy data and then act sur­prised when the rec­om­men­da­tions feel off. Garbage in, glossy insights out.

And this is where the whole AI nar­ra­tive starts to feel back­wards. Real intel­li­gence would start with doubt. It would ask: does this even make sense? Did a human heart real­is­ti­cal­ly jump 40 beats with­in one sec­ond on an easy run? Did some­one real­ly sprint across a lake at 30 km/h? If not, maybe fix the data first before build­ing train­ing plans and readi­ness scores on top of it. Clean sig­nals beat clever predictions.

There are oth­er exam­ples every­where. VO₂­max esti­mates that swing wild­ly after one bad work­out. »Over­train­ing« warn­ings trig­gered by a sin­gle night of poor sleep. Stress met­rics react­ing more to a loose watch strap than to actu­al life stress. It’s not that mod­el­ing per­for­mance is impos­si­ble — sports sci­ence has sol­id meth­ods for that — but all of them assume rea­son­ably reli­able input. With­out that, the num­bers just look pre­cise while being fun­da­men­tal­ly shaky.

Maybe the next real upgrade for run­ning tech isn’t anoth­er AI coach whis­per­ing split times. Maybe it’s some­thing less sexy: bet­ter fil­ter­ing, more skep­ti­cism, sys­tems that qui­et­ly fix obvi­ous non­sense before pre­sent­ing it as insight. Less crys­tal ball, more com­mon sense. I’d take that any day.