Key Takeaways
  • A 2026 Aging Cell study put 18 epigenetic clocks through two separate tests: run the same sample twice, and sample the same person repeatedly through an ordinary day
  • On the first, most clocks were near-perfect — technical reliability above 0.9 on the standard scale. On the second, most managed only 0.4 to 0.7, and none reached the excellent range
  • The two scores were essentially unrelated (r = 0.017). A clock that gives the same answer twice in the lab tells you nothing about whether it gives the same answer about you on Tuesday and Wednesday
  • A meal or a stressful hour was enough to move several widely sold clocks — DunedinPACE and GrimAge fell into the "poor" range under psychosocial stress
  • The instability propagates: with less reliable clocks, estimates of a diet intervention's effect varied widely and sometimes flipped direction between replicate measurements of the same people
  • The honest verdict: a biological-age number is a real signal with a wide error bar the report doesn't print. Treat one reading as a rough estimate, and a before-and-after with an unreliable clock as close to uninterpretable

The report arrives with two decimal places. Biological age 41.37, chronological age 47 — six years younger, congratulations, here's a chart. It looks like a measurement, the way a cholesterol panel looks like a measurement. And in one narrow sense it is: send the same tube of blood through the same machine twice and you'll get very nearly the same number. A study published this year in Aging Cell asked a different and more useful question. Send the same person through the machine twice — once before lunch, once after — and do you still get the same number? For most of the clocks on the market, the answer is: not especially.

Two Kinds of Reliable

Sehgal and colleagues, publishing in Aging Cell (2026), evaluated 18 DNA-methylation aging biomarkers on a platform called TranslAGE. The list is the field's greatest hits: the original Horvath and Hannum clocks, PhenoAge, GrimAge in both versions, the "pace of aging" clock DunedinPACE, and the newer principal-component and systems models — PCGrimAge, PCPhenoAge, SystemsAge, OMICmAge, CausalAge and others.

What makes the paper useful is that it separated two things people usually blur together.

Technical reliability asks whether the assay reproduces itself. The team ran replicate measurements of the same blood samples across four independent studies — 196 people with two replicates each on the newer 850K array, 128 people with two replicates on the older 450K array, a small set with sixteen replicates per person to isolate the effect of where the sample sat on the slide, and another with three replicates per person to test different DNA extraction protocols.

Biological reliability asks whether the person reproduces themselves. Here the team sampled the same individuals repeatedly over short windows under ordinary, non-medical perturbations: 34 people at four timepoints in a single day around a meal; 34 people at four timepoints in a day around a psychosocial stressor; 21 people at four timepoints over a week at altitude; 16 people at three timepoints over two days of air pollution exposure. The idea is simple. If your biological age is a stable property of you, lunch shouldn't move it.

r = 0.017
Correlation between a clock's technical reliability and its biological reliability across 18 clocks — effectively none. Knowing how well a clock reproduces in the lab tells you nothing about how well it reproduces in a person (Sehgal, Aging Cell 2026)

Both are scored with the intraclass correlation coefficient, a 0-to-1 scale where the paper uses the conventional cut-offs: 0.90 and above is excellent, 0.75 good, 0.50 moderate, below that poor. On technical reliability the majority of clocks scored excellent — above 0.9. The older generation, Hannum and the two PhenoAge variants, came in lower but still acceptable, roughly 0.7 to 0.8. SystemsAge, GrimAge and the PC-based clocks were the most reproducible of all.

On biological reliability the picture changed. Most clocks landed only in the moderate-to-good range, roughly 0.4 to 0.7, and no clock reached excellent. The best performer was PCGrimAge at 0.76. The newer PC-based and later-generation models — PCGrimAge, SystemsAge, OMICmAge — generally held up best. GrimAge in both versions, DNAmEMRAge and DunedinPACE clustered near the low end. And the two scores didn't track each other: the correlation between a clock's technical and biological reliability was 0.017, which is the statistical way of saying they have nothing to do with one another.

What Moves the Number

01

A meal

Around eating, many clocks performed only in the moderate or poor range. This is the perturbation most relevant to how these tests are actually taken: a kit arrives, you prick your finger or spit in a tube at whatever hour suits you, and the report never asks what you had for breakfast.

02

A stressful hour

Under psychosocial stress, several widely used clocks — the authors name DunedinPACE and GrimAge — fell into the "poor" reliability range. These are two of the most commercially prominent clocks, and two of the ones most often used to claim that an intervention "slowed your pace of aging."

03

The air, and the cells in the tube

Air pollution exposure produced sharp drops in stability; altitude, interestingly, less so, with several clocks approaching "good." And the abstract adds one more turn of the screw: most clocks' stability decreased further after adjusting for immune cell composition. Part of what a methylation clock reads is simply the mix of white blood cells in the sample, and that mix shifts with infection, sleep, exercise and stress. Strip it out and the remaining signal is shakier still.

Why This Isn't Academic

A wobbly biomarker would be a curiosity if it stayed in the lab. The paper's second half shows what happens when it's used for the things people actually buy these tests for.

First, prognosis. The team looked at how each clock associated with cognitive performance on the MMSE, and how much that association jumped around between replicate measurements. The variability of the result was tightly tied to the clock's technical reliability — a correlation of −0.91. Low-reliability clocks such as Hannum, PhenoAge and DNAmEMRAge produced effect estimates that, in the authors' words, fluctuated "across the full significance spectrum." Same people, same outcome, different replicate: sometimes a significant finding, sometimes nothing. The most reliable clocks gave narrow, consistent estimates.

Second, and closer to home: intervention response. Using data from a vegan-diet intervention, they asked whether each clock reported a consistent effect across replicates. Again the answer tracked reliability (r = −0.73). The less reliable clocks showed large variability in the size of the effect and often flipped direction — the same diet, in the same people, reading as beneficial in one replicate and harmful in another. The stable clocks kept their estimates consistent.

That is the finding to sit with if you have ever run a self-experiment on one of these kits. If you took a test, spent six months on a protocol, and took it again, the difference you saw is the sum of two things: whatever the protocol did, and the clock's own drift between two ordinary days. For a clock with a biological ICC around 0.5, the second term is not small. You cannot tell from two numbers which one you're looking at, and the two decimal places on the report don't help.

What the Clocks Still Have

It would be easy to read this as "biological age tests are fake." That's not what the paper says, and it isn't true. The original Horvath clock, published in Genome Biology in 2013, was a genuine discovery: DNA methylation at a few hundred sites tracks chronological age across tissues with a precision nobody expected. The later mortality-trained clocks do predict who dies sooner in large cohorts. That is real, population-level signal, and it is why the field exists.

The reliability problem isn't new to the field either. The reason PC-based clocks exist at all is a 2022 paper in Nature Aging (Higgins-Chen 2022) that identified exactly this noise in the original clocks and proposed a computational fix — and in the 2026 comparison those PC clocks are, in fact, the ones that hold up best. The field found a problem, published it, and partly solved it. That is how the process is supposed to work.

The authors also flag their own limits. The perturbations were short-term and acute; longitudinal studies over years are still needed. The populations were specific, not broad. Some of what registers as instability may be genuine short-term biological change rather than error — a point they raise and then argue against, but raise. And newer models may already address some of what they found. Their conclusion is not that clocks are worthless; it is that improving biological reliability is essential before these tools are ready for clinical and interventional use. That is a more modest claim than either the marketing or the backlash.

The Protocol: What to Actually Do

01

Treat one reading as a rough estimate, not a measurement

A single test gives you a number with an error bar the report won't print. If it says you're six years younger, the honest interpretation is "probably somewhere in the younger half, with real uncertainty." Don't restructure your life around the second decimal.

02

If you must retest, control the conditions

Same time of day, same fasted state, same sleep the night before, not the morning after a bad week. This won't remove biological drift but it removes the parts of it you can see coming. And know which clock the company uses: in this comparison, PCGrimAge, SystemsAge and OMICmAge were the ones that held up best.

03

Put the money where the evidence is

The things that reliably move mortality risk in large cohorts — cardiorespiratory fitness, strength, not smoking, the eating pattern, sleep, connection — are measured by cheaper and steadier instruments: a watch, a grip dynamometer, a tape measure, a blood pressure cuff. Our Longevity Calculator puts those together with the study behind each one. If a biological-age test motivates you to do those things, it has earned its price. If it replaces them, it hasn't.

The Honest Scorecard

Is the underlying science real? Yes — methylation tracks age, and mortality-trained clocks predict mortality in populations. Is your individual number a stable property of you? For most current clocks, only moderately — and a meal or a stressful hour can move it. Can one before-and-after tell you whether a protocol worked? With the less reliable clocks, no; the drift can be larger than the effect and can point the wrong way. Which clocks are best? In this comparison, the PC-based and later-generation models — PCGrimAge, SystemsAge, OMICmAge. Evidence grade for the reliability finding itself: Moderate — a well-designed methods study with replicates across four independent technical datasets and four biological perturbations, but short-term, in specific populations, and awaiting longitudinal confirmation.

Common Questions
Are biological age tests accurate?
They're precise more than they're accurate. Run the same sample twice and most clocks agree with themselves almost perfectly (technical ICC above 0.9). Sample the same person through an ordinary day and most clocks score only moderate (roughly 0.4–0.7), with none reaching excellent. The number is stable in the lab and drifts in the person.
Which epigenetic clock is the most reliable?
In the 2026 comparison, PCGrimAge scored best on biological reliability (ICC 0.76), with SystemsAge and OMICmAge close behind. GrimAge, DNAmEMRAge and DunedinPACE sat toward the low end, and DunedinPACE and GrimAge fell into the poor range under psychosocial stress. Technical reproducibility was a separate matter — most clocks were excellent on that, and it did not predict biological reliability.
Can a test tell me if a lifestyle change is working?
Only if the clock is stable enough that the change is bigger than its own noise. In the study's diet-intervention data, less reliable clocks gave effect estimates that varied widely and sometimes flipped direction between replicates of the same people. A single before-and-after with such a clock can't separate a real effect from drift.
So are epigenetic clocks useless?
No, and the authors say so. The conclusion is that biological reliability must improve before clinical and interventional use — a limitation, not a debunking. They also note that some apparent instability could be genuine short-term biology, and that newer models may already address part of the problem.
The Bottom Line

The clock agrees with itself. Whether it agrees with you is a different question.

Eighteen clocks, two kinds of reliability, and almost no relationship between them. That's the finding. A biological-age test can reproduce a sample flawlessly and still describe a person who changed between breakfast and lunch.

So read the number the way you'd read a single blood-pressure cuff reading taken in a doctor's waiting room: real, informative, and not something to act on alone. The levers that move the underlying biology were never in the test kit. They're the same five they've always been, and they're measured by instruments that don't drift when you eat.