The headline numbers are familiar. Women represent roughly a quarter of AI and data science professionals globally. At senior levels, the proportion drops to the low teens. At the C-suite level of AI-focused companies, it drops further. These numbers are cited so frequently that they have become background noise — acknowledged, lamented, and unchanged.

The reason they are unchanged is that most interventions target the wrong problem. The dominant framing is a “pipeline problem” — not enough women studying computer science, not enough women entering the field, not enough women in the candidate pool. This framing leads to pipeline interventions: scholarships, coding bootcamps, mentorship programs, recruitment initiatives. These interventions are well-intentioned and largely ineffective at changing the senior-level numbers, because the pipeline is not where the gap is created.

Where the gap actually forms

The gender gap in AI is primarily a retention problem, not a recruitment problem. Women enter AI and data science at higher rates than the headline numbers suggest. They leave at higher rates than their male colleagues. The gap that matters is not the gap at the entry point. It is the gap at the three-year mark, the five-year mark, and the ten-year mark.

A 2024 study by the AI Now Institute found that women in AI roles were 1.7 times more likely to leave the field within five years than men in equivalent roles. The reasons cited were not technical difficulty or lack of interest. They were organizational: exclusion from informal decision-making networks, attribution of their contributions to male colleagues, disproportionate assignment to support work rather than core development, and a culture that rewarded aggressive self-promotion over collaborative competence.

These are not pipeline problems. They are workplace culture problems that cause women who are already in the pipeline to exit it.

The attribution gap

The most pernicious dynamic I have observed is the attribution gap. In meetings, in code reviews, and in project retrospectives, women’s contributions are more likely to be attributed to the team or to a male colleague. A woman who identifies a critical data quality issue is described as having “raised a good point.” A man who makes an equivalent contribution is described as having “solved the problem.” The language difference is subtle and persistent, and it accumulates over time into a difference in perceived competence.

I watched this dynamic in a model review meeting. A female data scientist presented an analysis that identified a feature interaction the team had not considered. The team lead acknowledged the finding and then spent twenty minutes discussing the implications with a male colleague who had not contributed to the analysis. After the meeting, the male colleague received credit for the insight in the project update email. The data scientist did not correct the attribution because doing so would have required publicly contradicting her team lead.

This is not an unusual story. It is a pattern that research on gender dynamics in technical fields has documented extensively. And it has direct career consequences: perceived competence, not actual competence, drives promotions, project assignments, and compensation decisions.

The support work trap

Women in AI teams are disproportionately assigned to support work — data cleaning, documentation, stakeholder communication, project coordination — rather than core model development. This assignment is often framed positively: “She is so good with stakeholders.” “She is great at keeping the project organized.” “We need her on the data quality work because she catches things others miss.”

The framing obscures the career consequence. Core model development is where technical reputation is built. Support work is essential but invisible in performance reviews and promotion discussions. When women are consistently assigned to support work and men are consistently assigned to core development, the two groups accumulate different career capital. After three years, the men have a portfolio of models they built. The women have a portfolio of projects they organized. The promotion committee evaluates model portfolios.

The fix is not to eliminate support work. The fix is to distribute it equally, to make it visible in performance evaluations, and to ensure that every team member has a rotation through both core development and support work. The fix is structural, not attitudinal, which is why mentoring programs and awareness campaigns have not produced the desired results.

What actually moves the numbers

Three interventions have shown measurable impact on the gender gap at the senior level, and none of them is a pipeline program.

Structured promotion criteria. When promotion decisions are based on documented, pre-defined criteria rather than subjective assessment, the gender gap in promotions narrows. Subjective assessment is where bias operates. Removing the subjectivity does not eliminate bias, but it constrains its impact.

Equitable work assignment. When project assignments are managed explicitly — with tracking of who gets core development work versus support work — the assignment gap narrows. This requires a manager who is willing to track the data and act on it, which requires organizational support for the tracking.

Pay transparency. When compensation bands are published and individual compensation is audited for gender parity, the pay gap narrows. Pay transparency is uncomfortable for organizations that have been compensating inequitably, because it exposes the gap. But the exposure is a prerequisite for closing it.

The organizational cost

Organizations that lose women from their AI teams at higher rates than men are paying a cost that they may not be measuring. The cost is not just the loss of the individuals, though that cost is significant given the expense of recruiting and onboarding technical talent. The cost is the homogeneity of the perspectives that shape the AI systems being built.

AI systems that are built by homogeneous teams encode the assumptions and blind spots of that homogeneity. Recommendation systems that assume a male user’s content preferences. Hiring models that encode the patterns of historically male-dominated candidate pools. Health models that underrepresent conditions that present differently in women. The technical quality of these systems may be high. The representational quality is not, because the teams building them lack the perspectives that would surface the blind spots.

This is not an argument for diversity as a moral imperative, though the moral case is strong. It is an argument for diversity as a quality imperative. Teams that are more representative of their user populations build better systems, because they catch assumptions that homogeneous teams do not.

The uncomfortable truth

The gender gap in AI persists because the interventions that would actually close it are structurally uncomfortable. Structured promotion criteria reduce managerial discretion. Equitable work assignment requires managers to give up the convenience of assigning support work to the people who are good at it and willing to do it. Pay transparency exposes compensation inequities that the organization would prefer to keep internal.

These interventions work. They are also resisted, because they require people in positions of power to accept constraints on how they exercise that power. Until organizations are willing to make structural changes rather than pipeline investments, the headline numbers will remain familiar, the hand-wringing will continue, and the women who leave AI will continue to be described as a pipeline problem rather than an organizational failure.