Search for a study measuring what generative AI has done to employment, wages or hours in an Arab country and the count is zero.
Not a small literature. Not a contested literature. No employment study, no wage study, no productivity trial, in Jordan, Egypt, Saudi Arabia, the UAE, Morocco, Tunisia or anywhere else in the region. Every figure in regional circulation is one of two things: a labour-market statistic collected before the question existed, or an exposure projection calculated from American occupational data.
The usual caution applies and we will state it once. An absence found by a bounded search is not proof of absence, and if such a study exists we would like to be shown it. But we looked hard, and the organisations whose job it is to compile this literature have not found one either. The most recent regional scan of AI in education, commissioned by the UK's development agency and published in 2026, cites no effect sizes and no controlled trials from anywhere in MENA.
What "measured" means, and what other countries have
The distinction that does the work here is between a projection and a measurement. A projection says how much of a job could in principle be affected. A measurement says what actually happened to the people doing it.
The measurements that exist, in full:
Denmark. Humlum and Vestergaard linked two adoption surveys to national administrative payroll records — 11 exposed occupations, 25,000 workers, 7,000 workplaces — and found nothing. A precise null on earnings and on hours, tight enough to rule out effects larger than 2% two years after adoption. Users saved an average of 3% of their time. The null held for intensive users, for early adopters, for workplaces that invested heavily, and for early-career jobs.
The United States, payroll records. Stanford's Digital Economy Lab, working from ADP data through June 2026, finds employment of 22-to-25-year-olds in AI-exposed occupations 19% below where it would be had it tracked their less-exposed peers. In levels: that age group in the two most exposed quintiles fell about 11% between November 2022 and June 2026, while the three least-exposed quintiles grew about 10%. The adjustment runs through hiring, not firing. The authors are careful to call these descriptive indicators, not causal estimates, and the caveat should travel with the number every time.
The United States, freelance platform. Hui, Reshef and Zhou found a 2% fall in jobs and a 5.2% fall in monthly earnings for freelancers in the most affected occupations after ChatGPT, with the top-rated freelancers hit hardest.
The United States, firm-level. Hosseini Maasoum and Lichtinger report that firms adopting generative AI cut junior headcount by about 9% within six quarters while senior headcount held. We have not read the paper and are relying on secondary coverage of it; this literature revises its headline figures often, so treat that one as indicative.
Alongside those sit the productivity experiments, all of which found large positive effects on narrow tasks: 14% average and 34% for novices across 5,179 customer support agents; 40% less time and 18% better quality across 453 professionals writing; 55.8% faster for 95 developers on a single synthetic coding task; and, in the study everyone quotes half of, more than 40% higher quality for 758 consultants working inside AI's capability frontier and 19 percentage points less likely to be correct when working outside it.
Denmark, the United States, one online platform, one consultancy. That is the map.
The two answers disagree, and the reason is not the technology
Look at the sharpest contradiction in that list. Danish payroll records say the effect is indistinguishable from zero. A freelance platform says earnings fell 5.2%. Both studies are credible, both are recent, and the occupations overlap.
The reconciliation is institutional. Danish employment carries employment protection, firm-specific capital, and slow demand reallocation. A freelance platform carries none of those. The same task becoming cheaper produces a null in one setting and an earnings loss in the other, and the variable that decides which is the labour law, not the model.
That is precisely why neither answer transfers to an Arab labour market, and why the gap in the middle cannot be interpolated.
Consider what the region's institutions look like. Around 40% of Jordanian employment is public sector, and 72% to 87% of employed nationals in Saudi Arabia, Kuwait and Qatar — figures from a 2019 study on 2000s and 2010s data, and the most recent comparable series available. Meanwhile, two-thirds of MENA youth employment is paid work and the vast majority of those young paid workers are informal. The region contains both institutional extremes at once: a state sector with stronger protection than Denmark's, and an informal sector with less protection than a freelance platform. A single national estimate would average across the two and mean nothing.
What the region has instead
It has projections, and good ones. They should not be mistaken for the other thing.
The ILO applied its exposure index to detailed occupational distributions in August 2026 and found 5.8% of youth jobs in the Arab States and 4.0% in Northern Africa sitting in the most exposed categories, against 14.3% in high-income countries. It also published a scenario — explicitly a scenario — in which 10% of exposed jobs disappear entirely, giving 94,900 youth jobs in transition in the Arab States and 77,200 in Northern Africa. The World Bank, covering 25 low- and middle-income countries, put high exposure at 12% of workers in low-income and 15% in lower-middle-income countries, and stated in the paper that exposure "does not equate to job loss."
Those are careful documents whose authors say what they are. The problem is downstream, where a scenario becomes a forecast and a forecast becomes a slide.
It is not that the region cannot run the study
The most instructive precedent has nothing to do with AI.
In Jordan, Groh, Krishnan, McKenzie and Vishwanath randomised female community college graduates across four arms: a wage-subsidy voucher, 45 hours of employability skills training, both, and a control. The training produced no significant employment impact across three follow-up survey rounds. The voucher moved employment by 40 percentage points in the short run, but mostly into informal work, and the effect was no longer statistically significant four months after the subsidy ended. That result is consistent with the broader meta-analytic finding, across more than 200 programme evaluations, that training shows near-zero short-term employment effects and only turns positive at two to three years.
The point is not the null. The point is that a properly designed randomised labour-market experiment was run in Jordan, with administrative follow-up, and published in a peer-reviewed journal. The capability exists. It has simply never been pointed at this question.
What it would take is not exotic: an AI-adoption module attached to an existing national labour force survey, linked to social security or payroll records, with two waves. Jordan runs a quarterly labour force survey. The Gulf states run social insurance registries that cover their nationals completely. The instrument is a design problem, not a data problem.
What a founder should do about it
Stop citing foreign effect sizes as though they transfer. If your deck contains a productivity number, name the country, the sample and the task. "Fifty-five percent faster" describes 95 contractors building an HTTP server. It does not describe your customer.
Say the absence out loud. "No study has measured this in any Arab labour market" is a finding. It is more credible than an estimate, and it is the sentence that separates a founder who read the literature from one who read a summary of it.
Treat the missing study as an asset, not an obstacle. The first credible measurement of AI's labour-market effect in an Arab country will be cited for a decade by ministries, funders and every competitor in the category. It is available to whoever runs it, and nobody has claimed it.
Do not wait for the number to build the product. The absence is an argument against confident forecasting, not against building. Sell the outcome you can measure in your own product, in your own market, and publish it.
Fikr is investing in a region where the single most consequential labour-market question of the decade has no local evidence attached to it. That is not a reason to be quiet about the number. It is the reason to go and produce it.
[FIKR TO CONFIRM: whether Fikr Foundation will commission or co-fund a first measurement study — which country, which partner institution, and whether the design attaches to an existing national labour force survey or stands alone.]