5 Ways AI Models Make Women Look Less Suited For Leadership

Cambium AI , a company building synthetic populations for research, has been testing what happens when large language models behave like particular kinds of people. Why does this matter in the workplace?

Sometimes inequality appears in the data because inequality exists in the world. For example, if women carrying caregiving responsibilities have historically earned less , a model may learn the relationship. If employers have historically passed over Black women for leadership, the data may preserve that too. AI can stop caricaturing women and still reproduce the conditions behind their disadvantage.

Imagine surveying 100 women and 100 men and measuring how entrepreneurial they are. Plot the results on two bell curves and both groups would cluster in the middle. In real life, the curves would overlap heavily. Ask an AI model to simulate the same groups, and it shoves the two hills more than three times farther apart. Cambium AI found LLMs can indeed take a tiny difference and blow it up into a cartoon version.

Flawed examples of AI models

Leadership potential.A woman's 360 reviews describe her as supportive, collaborative and thoughtful. The AI summarizes her as a people-oriented supporter rather than a decisive leader. A man with comparable reviews "builds alignment." She "supports the team." Same behavior, softer verdict.

The stretch assignment she never gets.Her reviews call her careful, thorough and consensus-driven. An AI talent system converts those words into risk aversion, so she never makes the shortlist for the assignment she needs before the next promotion. It didn't deny her a promotion, but it denied her an opportunity capable of producing one.

The idea that reads as "not entrepreneurial enough."A woman pitches a new product and walks through customer risk, implementation constraints and what needs testing first. The AI reads caution and scores her lower on entrepreneurial drive. A pitch full of "disrupt," "capture the market" and "move aggressively" scores higher. Her risk analysis becomes evidence she isn't a builder.

The conflict she resolved becomes a conflict she'd avoid.A woman acknowledges another person's concerns, explains her position and proposes a compromise. The AI emphasizes empathy over her ability to hold a boundary. Her skill at resolving conflict becomes an assumed inability to fight one.

The ambition that doesn't sound like a pitch.A woman says she wants senior leadership, but also the right role, team and business problem. A colleague says, "I want to run a business unit within three years." The succession tool gets a concrete signal from him and may score her ambition lower. Do this enough times, and a company may conclude its women are opting out of advancement.

The importance of intersectionality

Michael Birdsall, cofounder and CEO of Cambium AI, also told me that current LLMs struggle to express emotions, like anger, as intensely as humans feel them. Coincidentally, in a recent study, Hogan Assessments found nearly all professionals carry at least one moderate or high-risk "derailer." A derailer is a personality trait, like confidence, charisma, or attention to detail, that helps someone succeed under normal conditions but tips into a liability (arrogance, attention-seeking, perfectionism) when they're under pressure, stressed, or fatigued.

Picture a company testing a controversial decision against thousands of synthetic women. The system reconstructs their ages, incomes, families and occupations, yet still underestimates their anger. A synthetic woman says, "I'm concerned." A real woman says, "Absolutely not." Tone, facial expression and body language carry information AI may miss. Derailers surface because of real human pressure, and if AI can't replicate the intensity of that pressure or emotion, it can't reliably model what emerges from it either.

Cambium AI's fix? Intersectionality. Instead of telling a model less about a person, tell it more.

If all you know is "she's a woman," one label does a lot of work. Learn she's also Asian, a mom and a city resident, and "woman" stops carrying so much explanatory weight. There is no single female experience: income, race, education, marital status, health, childhood, religion, career and politics also shape who she is. Women may work at the same company, in the same job title, without living the same reality.

In 2018, Amazon's AI recruiting tool , trained on a decade of male-heavy resume data, taught itself to downgrade any resume containing the word "women's" and penalize graduates of women's colleges. So we've seen this before. Now the risk isn't just resume screening, it's personality scoring: models trained to flag derailers without knowing whether they're reading real risk or a pattern baked in by biased history.

When a manager makes a real decision about a woman's career, compensation or opportunity because a machine made certain traits look more predictive than they actually are, that's a problem. Managers using AI need to become more demanding about evidence. Ask what the model knew about a candidate beyond her gender. Ask which behaviors it actually observed versus which characteristics it inferred. Don't accept "the AI recommended it" as evidence on its own.

This article was originally published on Forbes.com