Biological age is presented as a number that may differ from the years a person has lived. It is an estimate produced by a statistical model, not a reading taken from the body.
The models are trained on chronological age
Most biological age estimators are built by taking a large sample of people, measuring something in each of them, and finding the combination of measurements that best predicts their actual age.
Once trained, the model is applied to a new individual. The output is what the model expects for someone with that measurement profile.
The difference between the prediction and the person's real age is the quantity of interest. A prediction below true age is read as slower ageing.
Epigenetic clocks read chemical marks on DNA
The most studied estimators use methylation, small chemical groups attached to DNA that influence whether a gene is switched on.
Methylation patterns at particular sites change with age in a fairly consistent way across people, which makes them useful as a signal even where their function is not understood.
Different clocks use different site selections and were trained for different purposes, which is why the same sample can return different biological ages depending on which clock is applied.
Other estimators use routine clinical markers
Some models are built from measurements already collected in clinics, including inflammatory markers, kidney and liver values and blood counts.
These have the advantage of being interpretable. If a composite score is raised, the contributing measurements can be examined individually.
They tend to reflect current physiological state, which makes them more responsive to short-term change than clocks built on DNA methylation.
Prediction is not the same as causation
A clock that predicts mortality well has demonstrated that its signal is informative. It has not demonstrated that changing the signal changes the outcome.
Anything that alters the measured markers will alter the score, whether or not the underlying ageing process has moved.
This is the central limitation of consumer testing built on these models, and the reason researchers treat the numbers as research outputs rather than clinical results.
What the numbers can reasonably support
At population level, these estimates are useful for comparing groups and for testing whether an intervention shifts a measurable trajectory.
At individual level, the measurement error across repeat samples can be large enough to swamp the differences people are trying to detect.
Treating a single result as a verdict on personal ageing overstates what the method was designed to do, and any health decision arising from one belongs with a clinician.