The Instrumented Self: What Wearables Measure, and What They Miss

Wearable devices and biosensors promise something clinical medicine rarely offers anyone outside of illness: a continuous, personal readout of your own body. The promise is only half true. Some of what these devices measure holds up remarkably well against clinical gold standards. Some of it doesn't — and the marketing rarely draws a line between the two. And once a device turns a bodily process into a single daily number — a "readiness score," a "sleep score" — that number has a way of becoming the goal itself, sometimes at the expense of whatever it was built to describe.

Wearing the Data Since 2009

I've been wearing some device that claimed to measure my body since 2009, the year Fitbit shipped its first tracker: that Fitbit clipped to a waistband, then a Jawbone UP on the wrist, a Misfit Shine's small aluminum disc, and later an Oura ring and a Whoop strap. I've never worn a continuous glucose monitor myself, but I've watched the device closely from the outside — family and friends have used CGMs both as a prescribed necessity for managing diagnosed diabetes and, more recently, as an elective wellness purchase with no diagnosis behind it at all. That gap between the two uses — one anchored to a clinician's threshold, the other to a subscription app's dashboard — turns out to matter, and I'll come back to it.

At Google, I also spent time as an internal dogfooder on the health-and-fitness surfaces that shipped on Pixel and on Fitbit's post-acquisition hardware — testing sensor accuracy and scoring logic against the same competitors this essay covers, Whoop and Apple among them, from inside the pipeline rather than only as a consumer strapping on a ring each morning. That's part of why I'm skeptical of both the technology's critics and its marketers: the interesting failures are rarely the obvious ones.

In 2012, I taught at Singularity University alongside Rachel Kalmar, already one of the more rigorous voices in what had just been named the "quantified self" movement. Kalmar, a Stanford-trained neuroscientist, went on to work as a data scientist at Misfit and stayed a visible voice on wearables and quantified-self methodology for years afterward — she was still talking about the field's biohacking fringe on a 2017 a16z podcast — before carrying the same data-driven instincts into health and biotech product roles at Biogen, Tableau, and, most recently, Moderna. Her best-known experiment doubles here as a piece of methodological fieldwork: she wore twenty-one activity and fitness trackers at once, just to see where their readings agreed, where they diverged, and what that divergence implied about what any single one of them was actually measuring. Her conclusion — that the devices disagreed with each other more than their single-number outputs let on, and that each one locked its data in a proprietary silo that made comparison hard even for an expert — is as true of today's wearables as it was of hers.

A Short History of Self-Tracking

Self-tracking predates the smartphone by centuries — Benjamin Franklin's virtue ledger and the nineteenth-century weight-and-pulse diaries kept by physicians' patients are both recognizable ancestors — but the current wave has a specific founding moment. In 2007, Wired editors Gary Wolf and Kevin Kelly coined the term "quantified self" to describe "a collaboration of users and tool makers who share an interest in self-knowledge through self-tracking," and began hosting informal show-and-tell meetups in San Francisco. Wolf's 2010 New York Times Magazine essay, "The Data-Driven Life," brought the framing to a mainstream audience just as the hardware caught up to the ambition: Fitbit (2009), Misfit's Shine (funded on Indiegogo in 2012), Jawbone's UP, and Pebble's smartwatch all shipped within a few years of each other, each wagering that consumers would pay to see numbers about themselves that had previously been visible only in a lab.

Kalmar's twenty-one-tracker experiment belongs to this founding period, and it deserves to be treated as more than a curiosity. It was, in effect, an unfunded validation study of consumer-grade self-tracking before anyone in the industry had run one at scale, and it anticipated — a decade early — the central finding of the peer-reviewed validation literature that eventually followed: devices measuring the "same" thing, steps, sleep, heart rate, routinely disagree with each other by margins their single displayed number never lets on.

The category has consolidated and specialized since. Misfit was acquired by Fossil Group in 2015; Fitbit, after years as the category's dominant name, was acquired by Google in a deal that closed in January 2021 following a fourteen-month antitrust review, and its hardware and software lineage now feeds directly into Pixel Watch. Whoop and Oura, both founded in the 2010s, abandoned the display screen altogether in favor of a subscription-based, algorithmically-scored "recovery" or "readiness" number, betting that consumers wanted an answer more than they wanted raw data. Continuous glucose monitors, originally a prescription medical device for diabetes management, began marketing directly to metabolically healthy consumers in the early 2020s on the premise that glucose variability was a general wellness signal worth watching continuously rather than only when a doctor orders a blood draw.

What These Devices Get Right

The first thing to say in wearables' defense is that some of what they measure, they measure well. A 2022 validation study placed fifty-three adults in a sleep laboratory overnight, wired to polysomnography — the clinical gold standard — while they simultaneously wore an Apple Watch, an Oura ring, a Whoop strap, and three other consumer devices. For the basic binary question of whether a person was asleep or awake at a given moment, agreement with polysomnography was 88 percent for Apple Watch, 89 percent for Oura, and 86 percent for Whoop. For heart-rate variability specifically, Whoop achieved a 0.99 intraclass correlation with the clinical reference, while Oura's was 0.63 — good, but a long way from the near-perfect agreement Whoop showed.

That's not a trivial result. Before continuous consumer sensors, the only access most people had to their own resting heart rate, heart-rate variability, or sleep timing was an occasional clinical measurement, taken at a single moment under atypical conditions — the "white coat" effect is a documented distortion of blood pressure alone. A device that tracks the same quantity every night, in your own bed, builds a personal baseline and a record of its own drift over months that no annual physical could match. The value is less in any single reading than in the trend the readings reveal together — the same reason a string of polls tells you more than any one poll does.

The practical upside follows from that. Millions of people now have continuous, low-cost access to physiological baselines that would have required a research study or a hospital stay a generation ago, and that access has real clinical payoff: early atrial fibrillation detection via consumer smartwatches, still imperfect, has already prompted follow-up care in cases that would otherwise have gone unnoticed between annual physicals.

Where These Devices Fail

The same 2022 sleep-lab study that found 86 to 89 percent agreement for the simple sleep/wake distinction found something much less flattering once the question became harder. Asked to classify sleep into its clinically meaningful stages — light, deep, and REM — agreement with polysomnography fell to 53 percent for Apple Watch, 60 percent for Whoop, and 61 percent for Oura. The devices, in other words, are good at telling whether you are asleep; they are not good, by the standard clinical measure, at telling what kind of sleep you are getting, which is precisely the distinction most wearable marketing emphasizes with the greatest confidence — the deep-sleep graph, the REM percentage, the sleep-stage hypnogram displayed each morning as though it were as solid as the wake/sleep line beneath it.

Continuous glucose monitors show a parallel gap. In diabetic populations, where CGMs were developed and are rigorously validated against blood draws, devices such as the Dexcom G6 and FreeStyle Libre 3 report mean absolute relative differences (MARD) in the range of 8 to 9 percent — good enough for insulin-dosing decisions. But the growing market of CGMs sold to metabolically healthy, non-diabetic consumers rests on much thinner validation: one study of a widely used sensor in thirty-four healthy participants found a MARD of 17.6 percent, roughly double the diabetic-population figure, and most of the studies establishing baseline "healthy" glucose curves have enrolled a few dozen to a few hundred people over one to two weeks — a sample size and duration that would be considered exploratory, not confirmatory, anywhere else in clinical research. That gap between the two use cases tracks something I have observed at close range rather than firsthand, watching family and friends use the same device for different reasons: managed against a diabetes diagnosis, the number is checked against symptoms and a clinician's guidance and treated as an instrument whose accuracy genuinely matters; used electively as a wellness signal with no diagnosis behind it, the same number is noticed rather than verified — which is roughly what the validation data above would predict.

The deeper problem isn't any single quantity's error margin — it's what happens when several imperfectly measured quantities get combined into one proprietary score. A "readiness" or "recovery" number is typically a weighted composite of heart-rate variability, resting heart rate, sleep, and sometimes skin temperature or respiratory rate, combined by an algorithm the manufacturer doesn't fully disclose. Each input carries its own measurement uncertainty; the composite inherits all of it while displaying none of it, presenting a single integer with a false air of clinical precision — the same trick a poll's margin of error pulls when it reports one source of uncertainty and quietly implies it has covered them all.

When the Score Becomes the Goal

Even a perfectly measured quantity creates a problem once it's displayed daily and treated as a target. The economist Charles Goodhart observed in 1975, in a critique of British monetary policy, that any statistical regularity tends to collapse once you put pressure on it for control purposes. The anthropologist Marilyn Strathern gave the idea its now-standard, punchier phrasing two decades later — a line frequently misattributed to Goodhart himself, though he never wrote it, and one any product manager who has watched a team optimize a metric into meaninglessness will recognize instantly:

When a measure becomes a target, it ceases to be a good measure.

— Marilyn Strathern, "'Improving ratings': audit in the British university system," 1997

Wearables supply a clean, almost laboratory-grade illustration of the mechanism. A sleep score is meant to describe sleep. Once someone starts optimizing for a high sleep score specifically — going to bed earlier not because they're tired but because the ring rewards it, treating a low readiness number as a verdict to argue with rather than a noisy estimate to shrug off — the score has become the target. Clinicians now have a name for the resulting anxiety, "orthosomnia": a device meant to reassure you about sleep that instead makes you worse at it. The device hasn't gotten less accurate in Strathern's narrow sense; it's become a worse guide to the actual goal, which was rest, not a number describing rest.

The dynamic gets sharper once a third party attaches money to the score. John Hancock's 2018 decision to sell only "interactive" life insurance policies — premium discounts of up to 15 percent for customers who share Fitbit or Apple Watch activity data — turns a private wellness metric into a monitored, incentivized target for an outside party with its own stake in the outcome. That's the exact setup Goodhart's Law describes, and it's the bridge to the question the third essay in this series takes up directly: what happens once a personal biometric feed crosses an institutional threshold and becomes an input to somebody else's decision.

The Bottom Line

Wearables and consumer biosensors earn their trust unevenly, and the marketing rarely admits it. They validate well against clinical gold standards for some quantities — sleep/wake timing, heart-rate variability — and poorly for others, including the sleep-stage and glucose-variability numbers their interfaces present with the most graphical confidence. That gap between what's well measured and what's confidently displayed is the measurement problem. Layered on top of it is a second problem that shows up anywhere a soft measurement turns into a hard score: once an uncertain number becomes visible, and especially once it's watched by an employer, an insurer, or your own anxious attention, the score starts to shape the behavior it was only supposed to describe. The device that promised self-knowledge can end up manufacturing a slightly different self instead — one calibrated to a number rather than to how it actually feels to be well rested. Rachel Kalmar's twenty-one trackers, quietly disagreeing with each other on a decade-old wrist, made the same point before any of this had a name.