Our bias audit results, published in full.
The Eightfold Matching Model was independently audited for disparate impact across gender and race/ethnicity. We passed. Here is every number, every group, and every finding — no summaries, no spin.
Most AI organizations say they care about fairness. Few publish the data to prove it.
This page presents the full results of the 2026 bias audit of the Eightfold Matching Model, conducted by BABL AI Inc. under the requirements of New York City Local Law 144. The audit evaluated disparate impact across gender and race/ethnicity, internal governance, and risk assessment. The Eightfold Matching Model passed all three sections.
Prepared by BABL AI Inc. | March 3, 2026 | Signed: March 26, 2026
Overall results
Three categories audited. Three passing opinions.
The audit was conducted by BABL AI Inc., an independent auditing firm whose lead auditors are ForHumanity Certified under the NYC AEDT Bias Audit standard. BABL AI independence conforms to the ForHumanity and Sarbanes-Oxley definitions. Fees are fixed and unrelated to the opinion rendered.
The audit covered data from January 2024 through December 2025, analyzed across more than 29 million candidate assessments where demographic data was self-declared.
The system
What the Eightfold Matching Model does.
The Eightfold Matching Model evaluates a candidate’s skills and experience relative to a specific role and its requirements. It produces a match score — rated from 0 to five stars in increments of 0.5 — used primarily in the initial phase of external applications and occasionally to support internal mobility decisions such as promotions.
The model does not receive demographic inputs as part of its scoring logic. Scoring rates were measured after the fact, using self-declared demographic data, to assess whether model outputs produced disparate outcomes across groups.
Methodology
How scoring rates and impact ratios work.
The audit used the scoring rate method — the proportion of candidates within a demographic group who scored at or above the overall median score of the full population.
Impact ratios are calculated by dividing each group’s scoring rate by the scoring rate of the highest-scoring group. Under the federal Four-Fifths Rule (UGESP, 1978), an impact ratio below 0.80 is generally regarded as evidence of adverse impact.
In plain terms: if female candidates score above the median 62.8% of the time and male candidates do so 60.5% of the time, the male impact ratio is 0.962 — well above the 0.80 threshold. All groups in this audit remained above 0.80.
Disparate impact results · Gender
Gender scoring rates across 23.8 million assessments.
| Group | Candidates assessed | Scoring rate | Impact ratio |
|---|---|---|---|
| Female | 10,435,471 | 62.8% | 1.000 — reference |
| Male | 13,381,312 | 60.5% | 0.962 ✓ |
Female candidates scored at or above the median at a slightly higher rate than male candidates. The male impact ratio of 0.962 is comfortably above the 0.80 Four-Fifths threshold. No gender group showed adverse impact.
Note: An additional 74,997,062 candidate records were excluded from this calculation due to an unknown gender category — a reflection of real-world data collection limitations in self-declaration.
Disparate impact results · Race/Ethnicity
Race and ethnicity scoring rates across seven groups.
| Group | Candidates assessed | Scoring rate | Impact ratio |
|---|---|---|---|
| Hispanic or Latino | 2,120,351 | 67.2% | 1.000 — reference |
| American Indian or Alaskan Native | 122,484 | 66.3% | 0.986 ✓ |
| Native Hawaiian or Pacific Islander | 46,422 | 66.1% | 0.984 ✓ |
| Two or More Races | 746,927 | 65.7% | 0.978 ✓ |
| Black or African American | 2,246,235 | 64.8% | 0.965 ✓ |
| White | 5,291,543 | 64.8% | 0.964 ✓ |
| Asian | 4,880,107 | 63.0% | 0.938 ✓ |
Hispanic or Latino candidates had the highest scoring rate in this dataset, making them the reference group for impact ratio calculations. All seven race/ethnicity groups showed impact ratios above 0.80. No group triggered the adverse impact threshold.
The range across all groups was narrow: from 63.0% (Asian) to 67.2% (Hispanic or Latino), a spread of 4.2 percentage points across 15.4 million candidates with known race/ethnicity data.
Note: An additional 85,587,944 candidate records were excluded due to an unknown race/ethnicity category.
Disparate impact results · Intersectional analysis
Gender and race/ethnicity combined.
NYC Local Law 144 requires intersectional analysis, examining every combination of gender and race/ethnicity — not just each dimension separately. This is a more rigorous standard than most audits require. The reference group (highest scoring rate): Hispanic or Latina female candidates — 70.5%.
Female candidates
| Group | Candidates assessed | Scoring rate | Impact ratio |
|---|---|---|---|
| Hispanic or Latina Female | 1,247,002 | 70.5% | 1.000 — reference |
| American Indian or Alaskan Native Female | 57,387 | 70.0% | 0.994 ✓ |
| Native Hawaiian or Pacific Islander Female | 23,811 | 69.1% | 0.980 ✓ |
| Two or More Races Female | 358,847 | 68.1% | 0.966 ✓ |
| White Female | 2,421,935 | 67.8% | 0.962 ✓ |
| Black or African American Female | 1,216,990 | 66.4% | 0.941 ✓ |
| Asian Female | 1,643,219 | 62.1% | 0.880 ✓ |
Male candidates
| Group | Candidates assessed | Scoring rate | Impact ratio |
|---|---|---|---|
| Non-Hispanic Asian Male | 2,732,449 | 64.1% | 0.910 ✓ |
| Two or More Races Male | 321,033 | 63.7% | 0.903 ✓ |
| American Indian or Alaskan Native Male | 55,747 | 63.0% | 0.894 ✓ |
| Hispanic or Latino Male | 782,366 | 62.9% | 0.893 ✓ |
| Native Hawaiian or Pacific Islander Male | 18,789 | 62.9% | 0.892 ✓ |
| Non-Hispanic Black or African American Male | 809,717 | 62.7% | 0.890 ✓ |
| Non-Hispanic White Male | 2,282,433 | 62.2% | 0.882 ✓ |
All intersectional groups — including those with the lowest observed ratios — remain above the 0.80 Four-Fifths threshold. The lowest value in the dataset is 0.880 (Non-Hispanic White Male and Asian Female), both of which exceed the threshold. A consistent pattern appears across the data: female candidates scored at higher rates than their male counterparts across all race/ethnicity groups.
Governance · Audit finding: Pass
Who owns fairness at Eightfold AI.
Governance of bias and fairness risk at Eightfold AI is managed by a cross-functional Responsible AI working group. This team includes the Chief AI Compliance Officer alongside representatives from product, engineering, legal, and security, ensuring fairness considerations have direct influence over product decisions.
The governance structure passed all three audit criteria: the accountable party is identified, duties are clearly defined, and those duties were demonstrably carried out prior to the audit date.
Contact for bias audit inquiries: legal@eightfold.ai
Risk Assessment · Audit finding: Pass
How Eightfold AI identifies and monitors bias risk.
The audit reviewed the Eightfold AI internal risk assessment process, including the risk register, risk prioritization methodology, and evidence of ongoing monitoring. BABL AI reviewed screenshots from risk register dashboards, meeting minutes, and received verbal testimony from risk register maintainers.
The risk assessment covered risk identification, stakeholder impact, severity scoring, likelihood scoring, risk sources, and controls — all required dimensions under BABL AI’s Criterion Audit Framework, modeled after PCAOB Auditing Standard 1105.
Independent auditor
BABL AI Inc.
BABL AI Inc. is an independent AI auditing firm based in Iowa City, Iowa. Lead auditors are ForHumanity Certified Auditors under the NYC AEDT Bias Audit standard.
The BABL AI audit framework — the Criterion Audit Framework — is modeled after financial auditing practice and was published in the Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’24).
BABL AI independence is codified by the Sarbanes-Oxley Act of 2002 and the ForHumanity Code of Ethics. Fees paid for the audit are fixed and unrelated to the opinion rendered. The opinion is grounded solely in the audit criteria.
Signed by Dinah Rabe, Lead Auditor, BABL AI Inc. · March 26, 2026
Scope and limitations
What this audit covers — and what it does not.
This audit was designed to satisfy the requirements of New York City Local Law No. 144 of 2021. It does not certify that the Eightfold Matching Model is “bias-free” — no audit can make that claim — and it is not intended to demonstrate compliance with any legislation other than the NYC AEDT law.
Publishing this section is a deliberate choice. Transparency means being clear about what an audit covers, not only what it found.
— Assessed in this audit
- Gender (Male, Female)
- Race/ethnicity (seven groups per NYC LL 144)
- Intersectional groups (all gender × race/ethnicity permutations)
- Internal governance structure
- Risk assessment process
Our commitment
This is not a one-time event.
NYC Local Law 144 requires annual bias audits. We conduct them because the law requires it — and because we believe responsible AI is an ongoing practice, not a moment in time.
Questions about our methodology?
Contact our legal team for questions about our audit methodology, results, or responsible AI practices.
Curious about Responsible AI at Eightfold?
See how fairness and transparency shape every product decision we make — from model design to ongoing governance.
Learn more
Fairness is built into the model, not bolted on after.
See how the Eightfold Matching Model works — and how responsible AI thinking shapes every product decision we make.