AI Leadership Is Still an Overwhelmingly Male Room

Count the people who run the labs, lead the research teams and set the deployment policies at the largest artificial intelligence companies. The list is short, and it is overwhelmingly male.

The observation is not new, and neither is the pattern; the technology industry has had this problem for decades. What has changed is the stakes.

Hiring, lending, healthcare and policing

• These systems are deployed in hiring, lending, healthcare and policing, where bias has direct consequences.
• Training data reflects historical conditions, and the decision to measure that is a choice someone makes.
• Products are built for users whose circumstances the designers may not share.
• The field’s voice in public debate is concentrated among a narrow group.

Intake improved and seniority did not

The usual explanation is a pipeline problem: too few women in computer science, too few in machine learning. That is part of it, and the numbers at graduate level have improved substantially while leadership has not. The gap between improved intake and unchanged seniority is a retention problem, and retention problems are cultural.

The departures that get noticed are the senior ones, and several have been among the most consequential events in the industry’s recent history. Each departure is a signal about what the environment is like for the people who remain.

Pay bands, promotion criteria, leave that is not a penalty

The interventions that have evidence behind them are unglamorous: transparent pay bands, documented promotion criteria, parental leave that does not become a career penalty, and accountability for attrition numbers at the executive level. None of these are specific to AI. All of them are the reason the field’s demographics are, slowly, changing.

Demographics by seniority are rarely published

Most large technology companies publish workforce demographics in aggregate. Very few publish them by seniority in research organisations, which is where the decisions that shape deployed systems are made.

The absence of that data makes the problem easier to dismiss and harder to fix. An organisation that does not know its attrition rate by gender at each level cannot tell whether it has an intake problem or a retention problem.

Accents, darker skin, gendered defaults

The consequences are concrete rather than abstract. Speech recognition that performs worse for some accents, facial analysis that fails on darker skin, translation that defaults to gendered forms — these are documented failures that trace to training data and evaluation choices made by teams.

None of those failures required malice. They required the absence of anyone in the room who noticed, which is what a homogenous team produces.