Benford's Law: Why So Many Numbers Start With One

Take a large collection of real-world numbers — the populations of towns, the lengths of rivers, the amounts on a company's invoices — and look only at the first digit of each.

Intuition says each digit from 1 to 9 should appear roughly equally often, about eleven per cent each.

Intuition is wrong. In many such datasets, roughly thirty per cent of numbers begin with 1, about eighteen per cent with 2, and only about five per cent with 9.

This is Benford's law. It was noticed twice, decades apart, by people examining logarithm tables and observing that the early pages were more worn than the later ones — people were looking up numbers beginning with small digits far more often.

It sounds like a curiosity. It is used in practice to detect fabricated data.

Why small leading digits dominate

The clearest explanation comes from thinking about growth.

Suppose something grows steadily — an investment, a population, a company's revenue. Start at 1,000. To move past a leading digit of 1, it must reach 2,000: a doubling. That takes a long time.

Once it reaches 9,000, moving past a leading digit of 9 requires only reaching 10,000 — an increase of about eleven per cent. That happens quickly.

Then it is back to a leading digit of 1, and must double again to escape.

So anything growing by percentages spends much more time with small leading digits than large ones. Not because small numbers are special, but because the proportional distance between 1 and 2 is far greater than between 9 and 10.

The same logic applies to collections of numbers spanning many orders of magnitude. If a dataset includes values from tens to millions, the way those values distribute across scales produces exactly this pattern.

The deeper version of the explanation is that the law describes what happens when data is distributed evenly across logarithmic space rather than linear space — which is how many naturally growing quantities are distributed.

Using it to detect fabrication

The practical application follows from a simple observation: people inventing numbers do not produce this pattern.

Asked to make up plausible figures, humans distribute leading digits far too evenly, and often avoid 1 because it feels unconvincing. Fabricated data tends to look more uniform than real data.

So auditors and investigators compare the leading-digit distribution of a dataset against the expected pattern. A significant departure is a flag — not proof of anything, but a reason to examine more closely. The technique has been applied to accounting records, expense claims, scientific data, and reported statistics.

Two important limits on this use, and both matter because the technique is easy to misapply.

A deviation is a signal to investigate, not a finding. Plenty of legitimate datasets deviate for innocent reasons — a price point that recurs, a threshold that clusters values, a business that mostly transacts in similar amounts.

Someone who knows about the law can fabricate data that conforms to it. It catches naive fabrication, which is most of it, and not sophisticated fabrication.

Where it does not apply

Blind application produces false accusations, so the conditions matter.

Data must span several orders of magnitude. Adult human heights cluster tightly and never come close. Neither do exam scores or shoe sizes.

Numbers must arise naturally rather than being assigned. Phone numbers, invoice numbers issued sequentially, postcodes, and identification numbers follow whatever rule assigned them.

There must be no strong floor or ceiling. Values constrained to a narrow band — percentages, ratings out of ten, anything capped — do not follow the pattern.

The dataset must be reasonably large. Small collections vary too much for the comparison to mean anything, which is the point made in the law of large numbers.

The general lesson is worth more than the technique. Real data carries structural fingerprints that people do not intuitively reproduce, because those fingerprints come from the processes that generated the data rather than from anyone's sense of what looks plausible.

That is a broadly useful idea: authenticity often shows up in properties nobody thought to imitate, precisely because nobody knew they were there.