The limitations of aging clocks: from scores to outcomes that matter

A biological age result arrives as a single number. It's tempting to read it the way you'd read a lab value against its reference range. But that number is the output of a statistical model, not a direct measurement of a physiological state, and the gap between the two shapes how much weight it can carry.

1. A score is a model output, not an outcome

Every aging clock is a statistical model trained to predict something — calendar age, mortality risk, or the current pace of biological change — from a set of biomarkers. Different clocks are trained to predict different targets, which is one reason two tests can disagree without either being “wrong.”

Before treating any result as meaningful, it's worth asking the question the marketing rarely answers directly: what does this test measure? A score trained to predict calendar age answers a different question than one trained to predict mortality risk, and neither answers the question “how healthy am I, right now.”

2. Reliable for a group is not the same as reliable for you

Most aging clocks are validated by showing that, across a large cohort, the model's predictions correlate with age, disease incidence, or mortality. That is genuine evidence — but it is evidence about the group the model was tested on, not a guarantee about any one person's result.

A model can be well-calibrated on average across thousands of people while still being noisy, or even systematically off, for an individual. Knowing what a biological age score can and can't claim is a useful starting point before treating any single result as personally diagnostic.

Published accuracy figures are group summaries too. The original multi-tissue methylation clock, for example, was built on one set of datasets and tested on others, and its developer called the accuracy on the building data probably overly optimistic. A median error from that kind of testing means half the people tested were further from their calendar age than the figure. It is not a margin around your result, or a threshold for meaningful change between your own tests. Predicting age closely is also a separate question from tracking health. In a statistical analysis illustrated with blood metabolite data, a model that predicted age more accurately was less strongly linked to later mortality than simpler models. And being linked to an outcome is different from matching observed outcome rates; for a survival-based clock, the age result also changes if its reference population changes.

3. Two kinds of repeatability: technical and biological

“Repeatability” gets used loosely, but it splits into two different questions. Technical repeatability asks: if you ran the same specimen through the same assay twice, would you get the same score? It isolates noise from the instrument and the lab process. Biological repeatability asks a broader question: if you took a new sample from the same person on another occasion, would you get the same score? It is a broader question because it captures both assay variation and real short-interval within-person variation across repeated samples.

Technical repeatability can also vary by clock and lab method. In a 2025 cheek-swab study, each sample processed twice in one lab, the original PhenoAge clock's paired results differed by about 4 years on average with one lab method (arrays) and about 6 with another (sequencing), tested in separate groups. A version rebuilt to reduce noise averaged about 1 year with either. Repeatable is not the same as right: the two versions often gave ages many years apart. The authors work for Tally Health, which sells a cheek-swab test; that test was not studied.

A 2026 peer-reviewed comparison of biological versus technical reliability in epigenetic clocks found that, for the datasets and clocks it examined, blood samples taken hours or days apart agreed less well than one sample measured twice. That finding is specific to the clocks and cohorts the study looked at; it shouldn't be read as a universal number for every consumer test, since reliability varies by clock, assay, and provider. The broader point holds regardless: a provider that only reports technical repeatability is answering a narrower question than “would I get this same score again.”

This is the heart of one of the most useful questions to ask of any result: is an apparent change larger than expected variation? A score moving from one test to the next can reflect real change, ordinary biological fluctuation, assay noise, or some mix of the three — and without knowing how much results vary between separate samples from the same person, there's no way to tell which.

4. Why collection conditions can matter

Because biological repeatability includes real within-person fluctuation, how and when a sample is collected can influence the result independent of any underlying change in aging rate. In the 2026 reliability study above, the repeat samples were taken around events such as a meal, acute stress, pollution exposure or a trip to high altitude. Without control groups, it could not separate the effect of those events from the effect of timing. It used research data, not consumer tests, and reliability varied by clock and dataset.

Standardizing collection conditions — same provider, similar time of day, similar circumstances — is a reasonable precaution that may reduce this source of noise and make two results more comparable. None of the sources we reviewed tested whether it does. It does not eliminate normal biological variability, and it doesn't substitute for knowing how much variation is typical for the specific test you used.

5. A changed clock is not yet a health benefit

Even a score change larger than expected variation isn't automatically a health outcome. Biomarker movement — a lower score, a slower pace — does not by itself prove improved function, less disease, longer survival, or that a particular treatment or habit caused it. The geroscience field's own discussion of clinical trial endpoints makes this point directly: candidate aging biomarkers are surrogate measures, and the field is still working out which of them reliably track validated clinical outcomes rather than simply correlating with them at a population level.

A 2026 analysis of epigenetic clocks proposed checking whether reliable clocks respond, whether mortality-trained or pace-of-aging clocks respond, and whether several clocks move in the same direction. The authors present these as confidence checks—not proof of longer lifespan or healthspan. An isolated response can also reflect clock-specific biology rather than noise.

In 2026, several drug developers reported aging-clock results from their own clinical trials, which put checks like these to a real-world test. In the one peer-reviewed example, the company behind an experimental drug for a scarring lung disease reported that six protein-based clocks, including two built to predict risk of death, mostly pointed to a lower biological age estimate in treated patients than in those on placebo (42 patients in all, about ten per group), most clearly after four weeks. The same study shows why that is not yet a health benefit. The protein that contributed most to the change is itself tied to lung scarring, and the researchers say they could not fully separate an effect on aging from a lung disease that was getting better. Two makers of weight-loss drugs reported clock changes too, neither yet peer reviewed: one followed a single treated group without a placebo comparison, and the other analysed five placebo-controlled trials. Semaglutide's established heart and kidney benefits in high-risk patients come from counting clinical events, not from clocks.

6. Beyond the score: feel, function, survive

The outcomes that ultimately matter to most people aren't measured by a clock at all: mobility, cognition, independence, day-to-day symptoms, disability-free years, and survival. The World Health Organization's framing of healthy ageing centers functional ability — the capacity to do the things a person has reason to value — rather than any single biomarker. A biological age score is, at best, an early proxy for some of the processes that eventually show up in those outcomes. It is not a substitute for measuring them. WHO's related concept of intrinsic capacity — the physical and mental capacities a person can draw on — is one input into that functional ability, and it is also not the same thing as a biological-age score.

This is also a point some longevity researchers make outside the peer-reviewed literature. In a public Q&A, Buck Institute president Eric Verdin has described wanting people to worry less about chasing a lower biological age number and more about concrete functional outcomes — his example was being able to get down on the ground and back up to play with grandchildren — treating mobility, not the score, as the goal. A SuperAge profile of Verdin touches on a related theme: prioritizing how longevity efforts affect present-day wellbeing. Those are expert perspective and narrative pieces, not peer-reviewed evidence, but they're a useful corrective to treating the number itself as the outcome.

7. How to use a test without overreading it

None of this makes biological age tests useless — it points to how to use one responsibly. Start with what the test actually measures; the different clock types answer different questions, and knowing which one you're looking at changes what a result can tell you. Stick with the same provider and method if you're tracking change over time, and standardize collection conditions where you can. Before reacting to a shift in the number, check what kind of repeatability figure your provider quotes, if any: a figure from one sample measured twice covers only part of how much your results can vary between tests. See our guidance on testing frequency for how timing interacts with that. And keep the score in context alongside how you actually feel and function, rather than as a verdict on its own.

Sources

Peer-reviewed and official sources support the scientific claims above. The perspective sources are included to illustrate a viewpoint from longevity researchers, not as evidence for any scientific claim.