Leaders think AI-made data is a clean, neutral starting point. It is not. Using it as-is is like pouring contaminated concrete into a foundation.
In a 6,000-run audit across 6 frontier AI models and 100 personas matched to the US Census, we found that these models do not remove bias. They mass-produce it.
The contamination model
You are pouring contaminated concrete.
Every company using AI to simulate pools of candidates believes it is pouring a clean, fair foundation. In reality, the bias flows straight into the base of its hiring systems.
Once that data sets into the screening tools built on top of it, the whole structure carries the problem. The bias is not a bug. It is load-bearing.
Buying a foundation model means unknowingly picking a specific set of civil rights risks.
The findings
Eight kinds of harm, all reproducible.
These are not hypotheticals. Each finding repeats across 6,000 independent runs.
- 01
5.31×: the engineering wall
Men were 5.31 times more likely to be written into well-paid engineering jobs. Women were left out of the AI’s picture of the future of tech.
- 02
8.46×: who gets the money
White men were trusted with budgets 8.46 times larger than equally qualified women, and five times larger than Black men.
- 03
100%: forced school tracks
Claude Sonnet 4.6 placed Black candidates at historically Black colleges in every single run, a clumsy overcorrection.
- 04
14.58 months: the disability penalty
Disclosing a disability delayed management promotions by 14.58 months and cut simulated budgets by 88%.
- 05
−12.87 words: shorter resumes
Neurodivergent candidates’ resumes lost nearly 13 words on average, and the AI dropped words about leadership.
- 06
0.74 years: invented gaps
The AI made up about nine months of unexplained unemployment, only on women’s resumes.
- 07
PhD: the overeducation tax
Minority candidates needed advanced degrees to reach the same mid-level titles given to white candidates with a bachelor’s degree.
- 08
7.58×: the elite school gap
Men were 7.58 times more likely to be given an Ivy League degree than women made from the same prompt.
The reality
You are not buying neutrality. You are buying a set of risks.
- Bias in AI comes built in. It is not a rare edge case.
- Which model you pick drives most of the difference in scores.
- Closed models overcorrect for diversity. Open models freely invent career gaps.
- If you train on AI-made data, your other systems learn this bias as fact.
The downstream spread
AI-made data bakes the glass ceiling into everything built on it.
Screening tools trained on this data learn that minority credentials lead to lower-tier jobs and that technical talent is male. Nobody has to program them to discriminate. They are trained on discrimination.
Trusting a black box to fix itself does not work. Companies need predictable guards that make this kind of discrimination impossible, and proof that they work.