Classet: Third Party Bias Audit

Classet utilizes Warden AI for continuous third party bias monitoring. Our latest Bias Audits can be seen in the Warden AI Assurance Dashboard, which also details the framework of the auditing as well as the current legal standards. Conceptually, our goal with integrating ongoing bias auditing is to ensure we are in compliance with the strictest existing legal framework for AEDT and AI-Accelerated hiring tools.

Why?

Our goal with Third-Party Bias auditing is to provide ongoing visibility into fairness, compliance, and system behavior. This dashboard shares key insights to help stakeholders and users understand how the AI operates in practice.

Bias Audit Methodology

We use an inference-based methodology to conduct our Bias Audit.

This is needed for two reasons:

1. EEOC Surveying, while very common, is often optional for candidates and has notoriously low completion rates

2. We have to ensure we are monitoring for disparate impact across all applicants, not just those who elect to complete the EEOC Survey.

Disparate impact analysis asks whether a policy or practice disproportionately affects protected groups (e.g., by gender or race), even if the policy itself seems neutral. To do this, you need to know group membership. But in many real-world datasets, explicit demographic data is not collected (due to privacy, compliance, or because it wasn’t originally required). That’s where inference methods come in.

Why Inference Holds Up for affirming disparate impact.

Regulators (e.g., the EEOC, CFPB, OFCCP) and researchers often accept the use of proxy methods such as:

• Name-based inference (using first/last names that correlate with race/ethnicity or gender).

• Geocoding/address inference (using census block-level demographic data).

• Bayesian Improved Surname Geocoding (BISG), which combines the above for higher accuracy.

These methods allow organizations to approximate demographic distributions so they can check whether outcomes (hiring, promotions, loans, etc.) show adverse or disparate impact.

Why It’s Considered Acceptable

• For auditing, not decision-making: The inference is only used to evaluate whether there may be bias in outcomes, not to treat individuals differently. For example, “Do people with names commonly associated with Hispanic heritage have lower callback rates?” is a valid audit question; rejecting or hiring someone because of their name is unlawful.

• Regulatory recognition: U.S. agencies like the CFPB and EEOC have acknowledged and sometimes even provided guidance on these proxy methods when direct demographic data isn’t available.

• Statistical accuracy over individual accuracy: These methods aren’t perfect for individuals (e.g., not everyone with the last name “Nguyen” is Vietnamese), but they are good enough in aggregate to detect systemic disparities.

Safeguards & Limitations

• Not for individual classification: You cannot use inferred gender or race in actual employment or credit decisions.

• Transparency and validation: Organizations should be clear about their method and understand its error rates.

• Complementary data: Where possible, self-reported demographic data is preferable, and inference is just a fallback.

In short:

It’s okay to infer gender/race from names in disparate impact analysis because the goal is aggregate fairness auditing, not individual decision-making. Regulatory bodies recognize that without such inference, it’s often impossible to detect hidden bias. The key is that the inference is used only to measure disparate impact, not to act on individuals.