The Divide Inside One Country
Is Wider Than Between Two

We argue about AI as if we already know what people do with it. The arguments are about work: whose job is next, which desk it clears, how many people one prompt replaces. It is a loud debate, and most of it is conducted without looking at what anyone is actually typing.

As it happens, some of that is public. Every month Anthropic publishes the Anthropic Economic Index, an anonymised, aggregated record of what people do with Claude, broken down by country, by task, and by occupation. The latest release is 1.6 million rows in a single file, free to download, no key and no licence. I ran it through a pipeline on Bonacci Studio. Here is what was in there.

Less than half of it is work

The first number undoes the premise of the whole jobs argument. Sort every conversation by whether it is for work, for someone's personal life, or for schoolwork, and work is a minority of it.

Two in five conversations are somebody's own life, not their job. Meals, holidays, arguments, money and health. Another one in six is a student doing homework. The single biggest reason people show up is writing, which beats software development roughly two to one. And the single most common thing Claude hands back is not code. It is an explanation.

The same task, timed twice

Every conversation carries an estimate of how long the task would take a person alone, and how long it took with help. Across everything, the file says four and three-quarter hours becomes about forty minutes. Call it seven times faster.

There is a trap in getting there, and it is the kind that ships to production looking correct. The two columns sit next to each other in the same file, but one is measured in hours and the other in minutes, with nothing to warn you. Divide them at face value and you get 0.12, as if AI made every task eight times slower. The pipeline has to reconcile the units before it computes a single ratio. It is one line of code and the entire finding depends on it.

Does it do the work, or help you do it?

Every task is scored on a spectrum from Claude helped me do it to Claude just did it. Across the whole file it is almost an even split: 49% automation, 51% working alongside a person. The interesting part is not the average, it is where each kind of work lands.

• Fixing a broken system is 90% automated. Troubleshooting, log-keeping, keeping a dashboard alive: Claude mostly gets on with it.
• Mentoring someone in maths is 9% automated. Teaching, editing to someone's taste, tutoring: the human stays in the loop.

The pattern is legible without a statistics degree. The closer the work gets to another person, the more it stays a conversation. The closer it gets to a machine, the more it becomes automation.

The gap everyone quotes, and the one they miss

Now geography, where the headline hides a bigger story underneath it. One in five conversations anywhere on earth comes from the United States. But America is also large, so the fairer measure is usage per person, indexed so that 1.0 means a country uses Claude exactly in line with its share of the world.

America is 3.87. India is 0.30. A gap of about 13 times, and that is the comparison every write-up reaches for. One country against another, the familiar shape of the digital divide.

Then you stay inside America and look state by state, and the same divide is already there.

Washington DC runs at 3.32. West Virginia at 0.25. The gap between two American states, on the same day, under the same laws, is wider than the gap between America and India. Whatever this divide is, it is not really about borders.

What the data does not say

Three honesties, because the numbers are easy to over-read. First, this is Claude, not AI. Every figure here is a share of one company's traffic, not the whole market. Second, the time savings are the model's own estimate of a task, not a stopwatch held to a real worker.

Third, and this one is a data-engineering trap as much as a caveat. 114 countries were counted in April, 121 in May. The seven that appeared did not start using AI that month. They were always there, they just finally crossed the sample floor and became reportable. A cell that is below threshold is published as absent, not as zero, so filling the blanks with zero would quietly libel exactly the places the story is about. The quietest countries in this data are the ones it can say the least about.

Watch it get built

Below is the whole thing, the file going in, the numbers coming out, and every chart above being drawn from the same 1.6 million rows. Under two minutes.

The loudest arguments about AI are about work. The data, it turns out, is mostly about people, and about which of them get to use it at all.