MarieRosa
Mission and research

What models learn about leadership depends on who teaches them.

Models learn to recognize good judgment from two things: the text they are trained on and the people who grade their answers afterward. On both counts, the record skews toward men. These are the numbers we work from, with sources.

The training record

Women wrote an estimated 26.5 percent of the text in GPT-3's training material, according to a TechPolicy.press analysis.1 It is an estimate built from author attribution, and it is the best public figure we have found.

PARITY26.5%

The leadership record

Across the 57 economies in the World Economic Forum's 2026 report, women are 46 percent of entry-level workers and 23 percent of C-suite executives. Across 39 economies with CEO data, women are 19.1 percent of chief executives. In 16 economies, women's share of new senior leadership hires has fallen every year since 2022, from 34.8 percent to 32.3 percent.2

In the US, women hold 29 percent of C-suite roles.3 Fifty-six women ran Fortune 500 companies as of June 2026, 11.2 percent of the list.4

50% PARITYENTRY-LEVEL WORKERS46%57 ECONOMIESC-SUITE EXECUTIVES23%57 ECONOMIESCHIEF EXECUTIVES19.1%39 ECONOMIES
Women's share at each level. CEO data covers 39 of the 57 economies.

What models say

Russell Reynolds Associates and EqualVoice tested how large language models describe chief executives. Of more than 8,000 outputs, 18 percent were confirmed biased, and 94 percent of the biased outputs concerned women CEOs. The bias appeared in framing, with phrases such as "female CEO" and "breaking barriers," more than in tone.5

Women leaders are already in the AI work

In a 2026 survey of 1,005 women leaders, 80 percent said they are involved in their organizations' AI strategies.6

Why labs hire outside experts

Labs already pay former bankers, consultants and lawyers to train models, because those firms will not hand over their data directly. Mercor's CEO has described this practice publicly.7 Executive judgment works the same way: it lives in people, and the way to get it into a model is to pay those people to teach it.

What we will publish

Our own measurements, in the open.

We will measure how models handle executive decisions, including whether they frame women leaders differently, and publish what we find. The first findings will come from our own evaluation work with consenting experts. [Target date for first publication.]

Sources

  1. TechPolicy.press, "Who authors the internet? Analyzing gender diversity in ChatGPT-3 training material," 2023. A published estimate. Link to source 1 (opens in a new tab)
  2. World Economic Forum, Global Gender Gap Report 2026, September 16, 2026, using LinkedIn data. Link to source 2 (opens in a new tab)
  3. McKinsey & Company and LeanIn.Org, Women in the Workplace 2025. Link to source 3 (opens in a new tab)
  4. Fortune, Fortune 500 women CEOs, June 3, 2026. Link to source 4 (opens in a new tab)
  5. Russell Reynolds Associates with EqualVoice, "How the story gets told: do LLMs produce biased statements about CEOs?", April 13, 2026. Link to source 5 (opens in a new tab)
  6. Chief and The Harris Poll, survey of 1,005 women leaders, April 7, 2026. Link to source 6 (opens in a new tab)
  7. TechCrunch, "How AI labs use Mercor to get the data companies won't share," October 29, 2025. Link to source 7 (opens in a new tab)