The best number
in the world,
and the widest gap

The conversation about women and artificial intelligence in Asia is usually held as a pipeline story: not enough girls in STEM, not enough women in engineering, fix the supply and the rest follows. The data no longer supports that framing. In two Asian economies the supply problem is closer to solved than anywhere in the West. What has not moved is everything downstream of it.

Start with the numbers that are genuinely world-leading

On the Stanford AI Index measure of relative AI skill penetration, which compares how often AI skills appear on a country's professional profiles against the global average, Indian women score 1.9, the highest of any country's women in the index. That is up from 1.61 the year before. American women sit at 1.71, Canadian women at 0.97, British women at 0.90.12 India leads the overall index too, with AI skills appearing on Indian profiles at roughly three times the global average, and ranks second in the world for total AI authors and inventors.2

In the Gulf, Saudi Arabia records the highest female share of AI talent of any country in the index, 32.3%, ahead of Australia at 30.1% and Canada at 29.6%.3 In a single year, 666,000 Saudi women completed training in data and artificial intelligence.4

Set that against the global baseline. Women are roughly 22% of AI professionals worldwide and about 12% of AI researchers.5 On LinkedIn's own series, women's share of people listing AI engineering skills rose from 23.5% in 2018 to 29.4% in 2025, and the gender gap in AI talent narrowed in 74 of the 75 economies tracked.6

Figure 1

Women's relative AI skill penetration, selected economies

India1.90

United States1.71

Global average1.00

Canada0.97

United Kingdom0.90

These are not consolation figures. On the metric the industry itself uses to locate talent, the women of one Asian economy are ahead of every other country's women, and another Asian economy has the most gender balanced AI workforce in the world. Any argument that begins "women in Asia lack the skills" is now arguing with the evidence.

Now the number that runs the other way

The same region contains the widest digital gender gap on earth. In low- and middle-income countries women are 14% less likely than men to use mobile internet. In South Asia the gap is 32%, the widest of any region.7 Of the 885 million women in low- and middle-income countries who do not use mobile internet at all, around 60% live in South Asia or sub-Saharan Africa.7 Progress that ran from 2017 to 2020 has since stalled.7

Both facts describe Asia. They do not describe the same women. A woman writing model evaluation code in Bengaluru and a woman in rural Sindh without a handset in her own name are not at different points on one curve. They are on two curves, moving at different speeds, in a region that reports a single headline.

Aggregating them produces a regional average that describes nobody, and a policy response aimed at the average.

The exposure inversion

Here is the part that gets least attention and matters most.

The International Labour Organization's occupational exposure index finds that clerical and administrative work carries the highest generative AI exposure of any occupational group. Because women are concentrated in exactly that work, women are more than twice as likely as men to hold a job in the highest exposure category. Globally the top exposure band covers 4.7% of female employment against 2.4% of male employment.8

Applied to ASEAN, the same index finds roughly 80 million workers, or 22.9% of regional employment, in occupations with some generative AI exposure, and 11.7 million in the highest exposure band.9 Exposure tracks how office-based an economy is, which is why the wealthiest economy in the group carries the highest figure.

Figure 2

Share of total employment in occupations exposed to generative AI

Singapore42.2%

Philippines28.1%

ASEAN average22.9%

Indonesia21.7%

Vietnam20.8%

Thailand20.6%

The uncomfortable implication is structural. Clerical and administrative work has been, across much of Asia, the principal on-ramp for women into formal, contracted, socially acceptable paid employment. It is respectable, it is office-based, it is often the one category a family will agree to. It is also the single most exposed occupational group in the index.

So the same technology is running two effects on women in the same region at the same time: a widening opportunity at the top of the skill distribution, and a narrowing one at the entry point that most women actually use. A national figure that reports only the first is not wrong. It is half an answer.

Training is not the binding constraint at the top

If skills were the constraint, the numbers would improve with seniority once the pipeline filled. They do the opposite.

In India's technology sector, women's share of the workforce has roughly doubled over a decade to about 36%. Within that total, representation runs at 43% at entry level and between 4% and 8% at executive level.10 Globally, women hold 12.2% of STEM C-suite roles.6 The research figure is similar: about 12% of AI researchers, and roughly 16% of AI academic faculty.5

A drop from 43% to somewhere near 6% is not a skills gap. Nobody unlearns engineering between 28 and 42. It is a retention and progression failure, and it is invisible to every metric currently being celebrated, because those metrics count entries rather than exits.

Why the good numbers are the easy numbers

Note what the world-leading figures have in common. Skill penetration counts a term appearing on a profile. Training completions count attendance. Enrolment counts registration. Every one of them is an input, and inputs are cheap to collect, instantly available and flattering.

The number nobody publishes is the one that decides whether any of it worked: of the women trained, how many were in a data or AI role twelve months later, and at what earnings? We could not find that figure for any of the large Asian AI training programmes, including the ones we consider genuinely impressive. 666,000 women trained is a substantial achievement and a real commitment of public money. It is also, on its own, a receipt rather than a result.

This is not a criticism specific to AI programmes. It is the general failure mode we have written about elsewhere: a baseline nobody collected cannot be reconstructed afterwards, so the follow-up question has no answer in the data, permanently. If the twelve-month employment status of a 2026 cohort is not recorded in 2026, no analyst in 2028 can recover it.

It is also worth noticing that the United Nations' own indicator for women and technology, under SDG target 5.b, is the proportion of individuals who own a mobile telephone, disaggregated by sex. That is an access proxy. It was a reasonable proxy when it was written. In a period where the question is whether women hold, keep and advance in AI-exposed work, it no longer reaches the thing being asked.

What we would measure instead

Five indicators. None is expensive. All of them have to be designed before delivery starts, which is the only reason they are rare.

• Twelve-month role retention. Of participants who completed training, the proportion in a role using those skills one year later, with the follow-up sample and the non-contact rate published alongside it. Attrition reported, not quietly dropped.
• Earnings retained, not earnings gained. Income at a fixed interval after the intervention rather than at the point of placement. A placement figure measures the programme's last day. A retention figure measures whether it mattered.
• The seniority transition rate. The proportion of women at one level who reach the next within a defined window, compared with men in the same cohort. This is the indicator that would have exposed the 43% to 4-8% collapse a decade before it appeared in a report.
• Exposure-adjusted displacement. For any programme placing women in clerical or administrative work, the occupational exposure band of the destination role, recorded at placement. It costs nothing to record and it converts a success today into a risk that can be tracked.
• A connectivity baseline taken before, not after. Handset ownership in her own name, data affordability, and whether use is supervised by a household member. In a region with a 32% access gap, a digital skills programme that did not measure access first cannot attribute its own result.

The pattern across all five is the same. Each replaces a count of activity with a state observed at a later date, and each requires somebody to go back and ask. That step is the entire difference between a number that survives scrutiny and one that does not.

What this analysis cannot tell you

We would rather state the limits than have a reader find them.

• Skill penetration is measured on professional networking profiles. That population is urban, formally employed, English-language and self-selecting. It describes a real and important group of women, and it is not a sample of women in Asia. India's 1.9 is a genuine achievement by the women it counts, and it says nothing about the women it never sees.
• Author and inventor gender is usually inferred from names. Name-based inference performs unevenly across Asian naming conventions, so country-level gender shares in research and patent data carry more uncertainty than a single decimal place suggests.
• Exposure is not displacement. The ILO index measures overlap between task content and model capability. It is a risk map, not an outcome, and the ILO's own position is that large-scale disruption has not yet been observed.
• Pakistan is largely absent from these datasets. There is no comparable AI talent series for it. What exists is context: last of 148 economies on the Global Gender Gap Index, with women about 22.8% of the labour force.11 We work there, and we will not fill that gap with a plausible estimate.

The close

Asia is not behind on women in AI. On the measures the sector has chosen for itself, parts of Asia are ahead of everyone. That is the finding, and it should change how the region is discussed.

What has not been established is whether the women counted at the top of that distribution stay, advance and earn, and what happens to the far larger number of women whose route into paid work runs through the most automatable job category there is. Neither question is unanswerable. Both require somebody to decide, before a programme starts, that they intend to find out.

Until then the region will keep producing world-leading numbers about its own inputs, and no numbers at all about what those inputs bought.