Anthropic publishes the raw data behind its Economic Index, under an open licence, in files that run to hundreds of megabytes. It has never written about the Middle East. We pulled the May 2026 country rows and looked.
In Jordan, the largest single category of request people bring to Claude is Education & Learning, at 21.76% of all requests. It is the number one topic in the country. In Egypt it is also number one, at 21.50%. In the United States, the same category is fourth, at 9.28%.
That is not a survey of what people say they want AI for. It is a classification of what they actually typed.
What the numbers say
A short note on the method first, because it decides how much weight the finding can carry. Anthropic samples conversations, runs classifiers over them without humans reading transcripts, and maps each one to a task in O*NET, the US Department of Labor's occupational database. It publishes country-level cuts for 121 countries. Every figure below is our own computation from the raw file for May 2026, not a number Anthropic has published.
| Metric | Jordan | Egypt | United States |
|---|---|---|---|
| Education & Learning, share of requests | 21.76% | 21.50% | 9.28% |
| Rank of that category in-country | 1 | 1 | 4 |
| Coursework share of all conversations | 30.02% | 28.15% | 8.94% |
| Tasks mapping to Educational Instruction | 14.52% | 13.99% | 11.92% |
| Tasks mapping to Training and Teaching Others | 5.48% | 6.11% | 2.22% |
| Work share of conversations | 37.10% | 40.78% | 41.32% |
| AI Usage Index | 0.64 | 0.36 | 3.87 |
Four readings fall out of that table.
Coursework is three times the American share. Just under a third of everything a Jordanian brings to Claude is classified as coursework, against under a tenth in the United States. Jordan is not even the extreme case: Tunisia is the highest of all 121 countries at 51.52%, with Algeria second at 47.12% and Morocco at 34.89%.
Teaching is over-represented, not just studying. The share of tasks mapping to the "Training and Teaching Others" work activity is 5.48% in Jordan and 6.11% in Egypt, against 2.22% in the United States. Somebody on the other end of a meaningful share of these conversations is preparing to explain something to someone else.
Work use is only slightly lower. Jordan sits at 37.10% and Egypt at 40.78% against the US at 41.32%. Education is not displacing work use. It is stacked on top of a comparable base and pushing the personal category down instead: Americans use Claude for personal purposes in half of all conversations, Jordanians in a third.
And all of this is happening at very low volume. The AI Usage Index compares a country's share of global usage against its share of the world's working-age population. A value of 1.0 means exactly proportional. Jordan reads 0.64 and ranks 75th of 121. Egypt reads 0.36 and ranks 98th. The United States reads 3.87. Jordan is 0.11% of global usage; Egypt 0.62%; the United States 20.16%.
What it means, and what it does not
The first thing to say is what this is not. It is one company's users, in one product, and Claude is not the assistant most people in Cairo or Amman reach for. The population being measured is small, self-selected, and by the usage index below population parity in both countries. Cells below Anthropic's sample floors are suppressed, so a country missing from a cut has not been observed to be zero — it has not been published. None of these figures should be read as "Jordanians use AI for education", only as "among Jordanians using this product, this is what they use it for."
With that stated, the pattern is consistent enough across countries to be worth taking seriously, and it points somewhere specific.
Where AI is scarce, it arrives as a study tool before a work tool. Where AI is abundant, the opposite: Anthropic's own global series shows coursework falling from 19% of conversations to 12% while personal use rose from 35% to 42%, and educational use as a whole growing from 9% to 15% of conversations over 2025. The global trend and the regional level are pointing in different directions, which is exactly what you would expect if adoption in a market begins with the people who have the least to lose and the most immediate need — students — and only later reaches the enterprise seat.
Every EdTech roadmap written in San Francisco assumes the reverse sequence. Consumer, then prosumer, then work, then enterprise. In Jordan and Egypt the entry point is already coursework, and the enterprise seat is a long way behind it.
Set this against the labour market and the reason becomes uncomfortable rather than encouraging. The ILO's August 2026 youth report puts youth unemployment at 26.2% in the Arab States and 22.6% in Northern Africa — the highest rates in the world — with roughly a third of young people in neither employment, education nor training. The region's youth employment-to-population ratio was 18.5% in 2023, the lowest of any world region and half the global rate.
A population with the world's weakest attachment to formal work is using AI predominantly to learn. That is a demand signal that does not appear in any funding table, any market-size forecast, or any ministry's procurement plan.
The part that should worry a founder
Anthropic's student report, drawn from about a million higher-education conversations, classified what the model was actually producing against Bloom's taxonomy — the standard ladder of cognitive demand, from remembering facts at the bottom to creating new work at the top. The distribution came out inverted: Creating 39.8%, Analyzing 30.2%, Applying 10.9%, Understanding 10.0%, Remembering 1.8%. Roughly 47% of conversations were direct requests for an answer with minimal engagement.
Students are delegating the top of the ladder, not the bottom. That finding was reported as a US campus problem. If coursework is half of all AI use in Tunisia and a third in Jordan, it is not a US campus problem.
There is also a distribution problem underneath the demand. EdTech Hub's regional scan finds AI use in MENA concentrated among urban elites, private schools and well-resourced universities, while public-sector schools lag on infrastructure and class size. The students showing up in this dataset are the ones who already had the most.
What a founder should do about it
Build for the study session, not the enterprise seat. The market is telling you what it wants in its own typing. A product designed for a procurement cycle is being built for a customer who has not arrived yet in these two countries.
Price for a market at 0.36 and 0.64. Below-parity usage is not a reason to discount the opportunity; it is a constraint on what a user can pay per month and how much inference you can afford per session. Model the unit economics at the regional price point before designing the feature set, not after.
Design against the inverted pyramid. If the dominant use is answer-seeking at the top of Bloom's taxonomy, a product that simply answers faster is competing with a free general assistant on its strongest ground. The defensible position is making the delegation productive rather than blocking it, and that is an instructional design problem before it is a model problem.
Publish your own cut of this data. It is open, it is CC-BY, and it takes an afternoon. Nobody has run these numbers for the Arab world before, and the first company that publishes its own market's usage composition will be quoted by everyone who comes after.
The number one thing people in Jordan and Egypt ask AI to do is help them learn. Nobody in the region is currently building the thing they are asking for.