Navigating NYC Local Law 144: A Practical Guide to AEDT Bias Audits

The Slack Thread that Started It All

A few weeks back, we found ourselves in a heated Slack discussion about the implications of NYC Local Law 144. As a team of engineers at Fitlane AI, we knew our hiring algorithms needed to comply with this new regulation regarding automated employment decision tools (AEDTs). The stakes were high: if we didn't get this right, we risked not just financial penalties but also damage to our reputation as a fair hiring platform.

Understanding Local Law 144

NYC Local Law 144 mandates that any company using AI in hiring must conduct regular AEDT bias audits. This law aims to ensure that AI hiring tools do not discriminate against candidates based on race, gender, or other protected classes. The law applies to both the software we develop and the companies that use our services, making it crucial for us to ensure compliance.

The Challenge of Compliance

As we dove deeper into the requirements, we realized the complexity involved. One particular scenario highlighted this: a client using our platform noted discrepancies in candidate evaluations where certain demographic groups consistently received lower scores. This raised a red flag about potential biases in our machine learning models. We needed a systematic approach to audit our algorithms effectively.

Initial Approaches and Dead Ends

Initially, we considered a simple statistical analysis of the candidate scores across various demographic groups. However, this approach felt inadequate. Relying solely on statistical metrics could mask deeper issues related to the model's behavior. After much debate, we decided to abandon this method and explore more rigorous auditing frameworks.

Implementing the Technical Solution

Finally, we settled on a structured framework that included both quantitative and qualitative assessments of our algorithms. This involved:

1. Developing a standardized checklist for bias audits.
2. Implementing third-party evaluations to ensure independence.
3. Creating a feedback loop with clients to refine the models based on audit findings.

Here’s a snippet of the code we used to facilitate audit logging:

import logging

logging.basicConfig(level=logging.INFO, filename='audit_log.txt')

def log_audit(candidate_id, score, demographic_info):
    logging.info(f"Audit log for {candidate_id}: Score {score}, Demographics {demographic_info}")

This solution allowed us to track the performance of our models more transparently and ensured compliance with the audit requirements outlined in Local Law 144.

Observable Changes in Our Product

The most significant change post-implementation was the introduction of the audit logs. Clients using our hiring tools could now access detailed reports on how candidates were evaluated, helping them understand not just the results but the reasoning behind them. This transparency has positively affected our client relations, especially in the context of /for-companies, where we emphasize our commitment to fair hiring practices.

Key Lessons Learned

Through this process, we gleaned several insights that we hadn’t anticipated:

• Proactivity is critical: Waiting for compliance issues to arise can lead to more significant problems down the line.
• Independent audits add value: Third-party evaluations often reveal insights that internal teams might overlook.
• Documentation matters: Maintaining detailed records is not just a compliance necessity; it enhances the overall quality of our hiring algorithms.
• Stakeholder feedback is invaluable: Regular discussions with clients about their experiences can lead to continuous improvements in our models.

Implications for Candidates

For candidates, this audit process means that your evaluations will be more transparent and fair. You can trust that the algorithms assessing your applications are regularly scrutinized for bias, ensuring a more equitable hiring process.

Implications for Recruiters

Recruiters will benefit from clearer insights into how candidates are being evaluated. With access to audit logs and reports, they can better understand the decision-making processes behind candidate selections, leading to more informed and fair hiring decisions.

Looking Ahead

While we’ve made significant strides in compliance with Local Law 144, there’s still much to monitor. We’re particularly focused on refining our auditing processes and staying updated on any legal changes. If we could revisit this project, we would prioritize integrating real-time bias detection mechanisms, allowing for immediate adjustments rather than waiting for periodic audits. Our commitment to transparency and fairness in hiring remains strong, and we look forward to evolving alongside these regulatory requirements.