Build With AI — From Wherever You Are

In Brazil, 14.76% of sampled conversations were matched to software development tasks. In Korea that figure is 10.42%. What the data says about Brazil The number that would surprise anyone assuming a simple rich-country/poor-country ordering is 57.35% — the share of AI conversations from Brazil that are work-related. The global figure is 43.36%, and Brazil’s Anthropic Usage Index of 0.96 sits far below Korea’s 3.78. A naive reading would predict lighter, more casual use. The data shows the reverse: conversations from Brazil skew heavily toward getting work done, more so than almost anywhere else in this dataset. Coursework-like conversations, meanwhile, run at 11.52% against a global 16.45%, which reinforces the same picture from the other side. One plausible interpretation — and it is a guess from a single snapshot — is that usage in Brazil is concentrated among people who have already integrated these tools into their working day, rather than spread across casual experimentation. ...

In Ghana, 23.4% of sampled conversations were matched to education & learning tasks. In Korea that figure is 11.21%. What the data says about Ghana The number that should rearrange your assumptions is 29.57%. That is the share of conversations from Ghana matched to coursework or learning tasks, against a global figure of 16.45% — nearly double. Someone expecting a simple rich-country/poor-country ordering might assume that lower overall usage (Ghana’s Anthropic Usage Index is 0.38, ranked 95 of 121) means shallower use. Instead, the single largest structural feature of Ghana’s snapshot is that conversations there leaned far more toward learning than the world average, while work conversations came in at 37.78%, below the global 43.36%. ...

In India, 19.31% of sampled conversations were matched to education & learning tasks. In Korea that figure is 11.21%. What the data says about India The number that upends the usual assumptions is 0.3. India’s Anthropic Usage Index sits at 102 of 121 countries — by raw usage, barely on the chart. Korea, for comparison, posts 3.78. If you stopped there, you’d expect India’s conversations to be shallow or tentative. They are not. The mix of what actually gets asked is one of the most technically concentrated in the dataset: 16.22% of sampled conversations are matched to software development tasks, against a global average of 11.51%, and 19.31% match education and learning, against 13.23% globally. Low volume, high density. Whatever is limiting overall usage — access, cost, device mix, language coverage — it is not a lack of serious intent. ...

In Indonesia, 21.92% of sampled conversations were matched to education & learning tasks. In Korea that figure is 11.21%. What the data says about Indonesia The number that stands out in Indonesia’s data is not the usage index itself — 0.46, at rank 86 of 121 — but what fills the conversations. 45.88% of conversations matched to coursework or learning tasks, nearly three times the global share of 16.45%. Someone assuming a simple ordering where richer countries lead on everything would not predict this: on Education & Learning as a request topic, Indonesia runs at 21.92%, against a global average of 13.23% and Korea’s 11.21%. That is a 95.5% higher share than Korea, a country whose index of 3.78 is more than eight times Indonesia’s. The volume per capita may be smaller, but the shape of the demand is unmistakably oriented toward learning. ...

In Kenya, 15.83% of sampled conversations were matched to software development tasks. In Korea that figure is 10.42%. What the data says about Kenya The number that should stop you is this: conversations matched to software development tasks make up 15.83% of Kenya’s sample, against 10.42% for Korea — a share 51.9% higher than the country whose usage index is more than ten times larger. Kenya’s Anthropic Usage Index sits at 0.34, ranked 100 of 121, while Korea’s is 3.78. If you expected a simple ordering where higher adoption means more technical use, the data disagrees. Where usage is lower, it is also more concentrated: fewer casual conversations, more ones that look like building things. ...

In Nepal, 25.72% of sampled conversations were matched to education & learning tasks. In Korea that figure is 11.21%. What the data says about Nepal Nepal’s data refuses the obvious ordering. Overall usage is light — an Anthropic Usage Index of 0.41, ranking 92nd of 121 countries, against Korea’s 3.78. Yet conversations matched to software development tasks make up 16.29% of the sample in Nepal, versus 10.42% in Korea. On that share, Nepal leads Korea by 56.3%. Education and learning runs the same direction at higher volume: 25.72% of conversations in Nepal against 11.21% in Korea, a 129.4% gap. The country with far lower overall intensity is devoting a larger slice of its conversations to code and to learning than the country near the top of the index. ...

In Nigeria, 38.63% of sampled conversations were matched to content creation & copywriting tasks. In Korea that figure is 27.3%. What the data says about Nigeria The most interesting number in Nigeria’s data is 38.63%. That is the share of sampled conversations about content creation and copywriting — against a global average of 22.72%, and against 27.3% in Korea, a country whose overall usage index (3.78) is more than fifteen times Nigeria’s 0.24. A simple rich-country/poor-country ordering would predict Nigeria trailing on everything. Instead, on the single largest category of use, Nigeria’s conversations lead Korea’s by 41.5%, and sit 70% above the world average. Low aggregate volume, highly concentrated purpose. ...

In Philippines, 16.63% of sampled conversations were matched to software development tasks. In Korea that figure is 10.42%. What the data says about Philippines Here is the number that should flip any easy assumption about who uses AI “seriously”: 16.63% of sampled conversations from the Philippines were matched to software development tasks, compared with 10.42% from Korea. That is nearly 60% higher than a country whose overall usage index is more than eight times larger (3.78 against the Philippines’ 0.44). Fewer conversations in total, but a much heavier concentration of code inside them. If you expected intensity of use and technical depth of use to travel together, this country’s data says otherwise. ...

In Rwanda, 25.68% of sampled conversations were matched to education & learning tasks. In Korea that figure is 11.21%. What the data says about Rwanda The number that breaks the usual ordering is this: conversations matched to Education & Learning tasks make up 25.68% of sampled usage in Rwanda — nearly double the global share of 13.23%, and 129.1% higher than Korea’s 11.21%. A reader assuming that richer countries lead on every dimension of AI use would expect the comparison to run the other way. It does not. And the pattern repeats at the level of overall conversation type: 34.76% of conversations in Rwanda look like coursework or learning, against a global 16.45%. Careful reading matters here — this describes what the conversations resembled, not who was having them. But taken at face value, the data suggests these tools are being used heavily as study instruments: explaining material, working through exercises, drafting and checking written work. ...

In Vietnam, 18.01% of sampled conversations were matched to software development tasks. In Korea that figure is 10.42%. What the data says about Vietnam Vietnam’s overall Anthropic Usage Index sits at 0.53, roughly a seventh of Korea’s 3.78. If you stopped there, you would assume conversations about software are proportionally rarer too. They are not. 18.01% of sampled conversations from Vietnam matched software development tasks — against 11.51% globally and 10.42% in Korea. On the thing I care most about, conversations from Vietnam are 72.8% more likely to look like software work than conversations from Korea. Lower overall usage, heavier engineering flavour in the usage that exists. ...

It arrived. Your message is in my inbox and nobody else can see it. I read everything. I answer as fast as I can, which is not always fast — if it’s been more than a week and you haven’t heard back, send it again or reach me on GitHub. Persistence is not rude; it’s the thing this whole page is about. If you left a repository link, I’ll look at the code before I reply, so the answer may take a little longer and should be worth more. ...