Two claims crossed my feed in the same week, both about women and AI, both citing real research, both apparently true — and pointing in opposite directions.

The first: 86% of the workers most at risk from AI are women — and they're in admin, finance support, retail, and care. Low-paid, low-margin, hard-to-retrain roles.

The second: the workers most exposed to AI are more likely to be female, more educated, and higher-paid — coaches, consultants, executives, strategists.

So which is it? Are the women in the firing line the lowest-paid or the highest-paid? It can't be both.

Except it is. The two claims aren't contradictory. They're measuring two different things, and the confusion between them is the single most important thing to get right about AI and work. Once you separate the axes, the whole picture snaps into focus — and the policy response for each group turns out to be almost opposite.

Axis one: exposure

Exposure answers a narrow technical question: can current AI do a meaningful share of the tasks in this job? It says nothing about whether anyone gets fired. It's a measure of overlap between what the model can do and what the role involves.

Anthropic's Labor Market Impacts of AI report (March 2026) measured this directly and found that the more-exposed group is "16 percentage points more likely to be female" and "earn 47% more, on average, and have higher levels of education." The IMF reached the same directional conclusion back in January 2024: in advanced economies, women and college-educated workers are more exposed — because the tasks AI is good at (drafting, summarising, analysis, structured knowledge work) are the tasks those jobs are made of.

This is the second claim. And here's the part the viral version leaves out: the same Anthropic report found no significant rise in unemployment for highly exposed workers since ChatGPT launched. Exposure is not displacement. A senior consultant whose job is 40% automatable doesn't get fired — she gets a tool that does the 40%, and she's expected to do more of the other 60%. High exposure plus high adaptive capacity equals adaptation, not foreclosure.

So when an Instagram carousel takes the "most exposed are educated, higher-paid women" stat and concludes those women are about to lose their jobs, it's borrowing a real number to tell a story the number doesn't support.

Axis two: welfare risk

Welfare risk answers the question that actually matters to a human being: if this job is disrupted, how badly will this person struggle to recover? That depends on savings, age, how transferable the skills are, and whether there are other jobs nearby.

This is what the Brookings report (January 2026) measured, and it deliberately built a model to separate it from raw exposure. Its finding:

"Some 6.1 million workers (4.2% of the workforce in the sample) will likely contend with both high AI exposure and low adaptive capacity… about 86% are women."

That's the first claim — and it's solid, with one correction the viral version got wrong: this is a US study, not a global one. The 6.1 million are drawn from a US workforce sample. And critically, Brookings excludes the high-paid, high-exposure workers — the lawyers, the software developers, the consultants — from this group, because they have the savings and the transferable skills to adapt. The 6.1 million are clerical and administrative workers, overwhelmingly women, for whom disruption isn't an upgrade to their toolkit. It's a cliff.

The ILO's data (March 2026) confirms the shape globally: 29% of female-dominated occupations are exposed to generative AI versus 16% of male-dominated ones, and in the highest-risk band the split widens to 16% versus 3%. Their refined 2025 index put female employment in the highest-exposure category at 4.7% globally versus 2.4% for men — and 9.6% versus 3.5% in high-income countries. Women are more exposed than men in 88% of countries analysed.

The map

Plot the two axes against each other and every worker — every woman — lands in one of four quadrants.

Rendering diagram...

The two viral claims sit in opposite corners of the same chart, and both corners are disproportionately female:

• Top-right — "Adapt and thrive." High exposure, high adaptive capacity. The educated, higher-paid women in the Anthropic data. AI eats part of their job; they have the savings, mobility, and skills to absorb it and move up. This is the group the second claim is about. They are exposed. They are mostly fine.
• Bottom-right — "Highest welfare risk." High exposure, low adaptive capacity. The 6.1 million in the Brookings data. Clerical and admin women for whom the same disruption means a cliff, not an upgrade. This is the group the first claim is about. They are the ones who actually need help.

Same gender skew. Same "high exposure." Opposite outcomes — because the y-axis is doing all the work, and neither viral claim mentions it.

Why this matters beyond getting the stat right

The reason to separate the axes isn't pedantry. It's that the two groups need opposite interventions, and conflating them sends help to the wrong place.

The top-right group needs access and encouragement — and here the gap is real and measurable. Lean In's March 2026 survey (n=1,015) found men are 27% more likely to be praised for using AI at work and 23% more likely to be encouraged by a manager to use it. Women's adoption is climbing fast — Deloitte found US women's GenAI adoption tripled in a year, outpacing men's 2.2× — but they're recognised at roughly half the rate. For this group, the fix is cultural: stop treating a man with Copilot as innovative and a woman with Copilot as cutting corners.

The bottom-right group needs something completely different — transition support: retraining funded before the disruption, not after; income bridges; geographic mobility. Encouragement to adopt AI does almost nothing for a 55-year-old accounts clerk in a town with one employer. She doesn't have a recognition problem. She has an adaptive-capacity problem, and no amount of "lean in and learn prompting" closes it.

When a viral post collapses both groups into one "women are most at risk" headline, it pulls attention and resources toward the louder, more online, more adaptable group — and away from the group Brookings actually flagged as most at risk. The conflation isn't just inaccurate. It's regressive.

The takeaway

This is the third piece I've written about who AI actually comes for. Protect the Juniors was about who by career stage. The Measurer Trap was about who by function. This one is about who by demographic — and it carries a method the other two didn't need: always ask which axis a displacement statistic is measuring.

Exposure tells you whose tasks overlap with the model. Welfare risk tells you who can't recover when they do. A number that doesn't tell you which one it's measuring isn't telling you anything you can act on.

The women in the top-right corner will mostly be fine, and they're getting most of the airtime. The women in the bottom-right corner are the 86%, and they're getting a masterclass funnel. The most useful thing you can do with either statistic is to notice, every time, which corner it's actually about.