AI systems make decisions affecting billions of people — from loan approvals to medical diagnoses to criminal sentencing. AI auditing ensures these systems work as intended and don't cause harm.

What Is an AI Audit?

An AI audit is a systematic evaluation of an AI system assessing:

• Fairness — does the system treat all groups equitably?
• Accuracy — does the system perform as claimed?
• Transparency — can decisions be explained and understood?
• Safety — does the system avoid harmful outputs?
• Privacy — does the system protect personal data?
• Compliance — does the system meet regulatory requirements?

Why Audit AI?

High-profile AI failures demonstrate the need:

1. Amazon Hiring AI — automated resume screening that penalized women's applications
2. COMPAS Recidivism — criminal justice risk scores with racial disparities
3. Healthcare Algorithms — disease prediction that underserved Black patients
4. Facial Recognition — error rates 10-100x higher for darker-skinned individuals
5. Content Moderation — automated systems disproportionately censoring certain communities

Types of AI Audits

Different audits serve different purposes:

• Pre-Deployment Audit — evaluate before launching an AI system
• Continuous Monitoring — ongoing assessment of live systems
• Incident Investigation — deep dive after a problem is identified
• Compliance Audit — verify adherence to specific regulations
• Third-Party Audit — independent evaluation by external experts

The Growing Audit Market

AI auditing is becoming a professional discipline:

• Companies like ORCAA, Holistic AI, and Credo AI provide audit services
• Regulatory requirements (EU AI Act, NYC Local Law 144) mandate audits
• Professional certifications are emerging for AI auditors
• Insurance companies are requiring AI audits as part of underwriting

NYC Local Law 144: A Real Compliance Deadline

New York City's Local Law 144, effective January 2023, requires employers that use automated employment decision tools (AEDTs) in hiring or promotion decisions to conduct annual bias audits by an independent auditor and publish the results. Vendors selling AEDT products must also provide bias audit summaries. This was the first binding municipal regulation specifically requiring third-party AI audits — and it has been followed by similar legislation in Illinois, Maryland, and at the EU level. If your organization uses AI for hiring, promotion, or compensation decisions in covered jurisdictions, audit compliance is not optional.

The Amazon Hiring AI Case Study

Amazon built a resume-screening AI trained on historical hiring decisions. Because the company had historically hired more men in technical roles, the AI learned to penalize resumes that included signals of female applicants — including phrases like "women's chess club" or attendance at all-female colleges. The system penalized resumes from two all-women's colleges explicitly. Amazon disbanded the team in 2018 when they could not eliminate the bias. The lesson for auditors: when training data reflects historical human decisions, the AI inherits those decisions' biases. Auditing the model alone is insufficient — you must audit the training data labels as well.

Common Pitfall: Auditing Only at Launch

Many organizations conduct a thorough pre-deployment audit and then treat the AI as safe indefinitely. In practice, AI systems can develop new biases or accuracy degradations after deployment through distribution shift — the real-world data the model encounters diverges from what it was trained on. A model that was fair at launch can become unfair six months later if the user population changes. Build continuous monitoring into every AI deployment, with quarterly metric snapshots and automatic alerts if fairness metrics shift beyond acceptable thresholds.