Benford's Law: Why Fabricated Numbers Always Look Wrong to a Trained Eye

Pick any big pile of numbers that occurred naturally: river lengths, electricity bills, stock prices, the populations of every town in your country. Count how often each number starts with a 1, a 2, a 3, and so on up through 9. You'd guess roughly equal odds for each, about 11% apiece. You'd be wrong, and the gap is not small.

Photo by Zacharias Korsalka on Pexels.

About 30% of those numbers will start with 1. Only around 4.6% will start with 9. The distribution falls off in a smooth logarithmic curve: P(d) = log10(1 + 1/d) for digit d. This is Benford's Law, named after physicist Frank Benford, who noticed in 1938 that the early pages of logarithm tables (used to look up small leading digits) were far more worn and smudged than the later pages. Someone had already clocked this pattern decades earlier, mathematician Simon Newcomb in 1881, but Benford did the legwork across twenty different datasets and got his name on it.

Why would leading digits behave this way? Because many real quantities grow multiplicatively, not additively. A number has to travel through the '1-something' range (1 to 1.999...) before it can double into the '2-something' range, and that doubling takes proportionally longer at each stage as the number grows. Spend time thinking in terms of ratios rather than differences and the skew stops looking weird. It starts looking obvious.

Here's where it gets genuinely useful. Benford's Law doesn't care what the numbers measure. Populations, river lengths, physical constants, accounting entries, tax filings, invoice totals: as long as the data spans several orders of magnitude and isn't artificially capped or assigned (phone numbers don't count, zip codes don't count), the first-digit distribution snaps to the Benford curve with eerie consistency.

Fabricated numbers don't. Humans making up figures lean on digits that feel 'random' to them, which usually means an overrepresentation of middle digits (5, 6, 7) and an underrepresentation of 1. We're bad intuitive random number generators, and we're especially bad at mimicking a logarithmic skew we've never consciously noticed. This is why forensic accountants and tax auditors run Benford tests on expense reports, election vote counts, and corporate filings. A dataset that should follow Benford's Law but instead shows a flat or inverted digit distribution is a flag, not a proof, but a very good reason to look closer.

The method has real limits worth naming before you go full detective on your coworker's spreadsheet. It works best on data spanning multiple orders of magnitude (think company revenues from $10,000 to $10,000,000, not a dataset capped between 1 and 100). It falls apart on assigned or bounded numbers, like ages or standardized test scores. And a Benford violation on its own proves nothing; seasonal invoicing patterns, rounding conventions, or a company that only ever bills in increments of $500 can all produce non-Benford distributions without a single fraudulent entry.

graph TD
    A[Collect dataset spanning orders of magnitude] --> B[Extract leading digit of each value]
    B --> C[Compare observed frequency to Benford curve]
    C --> D{Big deviation?}
    D -->|No| E[Looks organically generated]
    D -->|Yes| F[Flag for manual review, not automatic guilt]

What makes Benford's Law worth knowing beyond fraud detection is the reminder it gives about what 'random' actually looks like. Most people's mental model of randomness is uniform: every outcome equally likely, nothing favored. Reality disagrees constantly. Earthquake magnitudes, city populations, the sizes of files on your hard drive: skewed distributions are the norm, not the exception, once growth is multiplicative rather than additive.

Next time someone hands you a spreadsheet of numbers they swear is untouched, try a leading-digit count before you trust the summary stats. If the 1s aren't winning by a wide margin, ask why.