AI Safety for Hiring: HR, Bias, and NYC Local Law 144

TL;DR

If HR uses a model to rank, screen, or chat with candidates, you are a deployer of a high-impact system. New York City Local Law 144 is one example of bias-audit and notice rules for automated employment decision tools. Safety training for HR is literacy plus a policy, not a CISO cert.

A model score is not a hire. A named human must be able to override it.

Key Takeaways

• Safety = protect people from the model (accidental harm: bias, hallucinations, leaking PII, over-reliance) — guardrails. Security = protect the model from attackers (adversarial: prompt injection, breaches, exfiltration). Intent is the key distinction.
• Over-reliance on a fluent ranking is a safety failure even when no one attacked the model.
• Name a human who can override the tool before it moves someone’s job.

Why hiring is a safety lane

A biased ranking does not need an attacker. It needs a model trained on yesterday’s workforce and a recruiter who treated the score as a fact. That is accidental harm: bias and over-reliance. Security still matters if someone poisons the corpus. HR’s first file is still the safety policy.

NYC Local Law 144 in plain English

New York City requires independent bias audits and candidate notice when automated employment decision tools are used in hiring and promotion, with exceptions and evolving guidance. It is not the EU AI Act. It is a reminder that a deployer in HR cannot say “the vendor’s model did it.”

What HR should write this week

Allowed tools. Forbidden inputs (health, protected class guesses, secret notes). Who reviews a reject. How candidates are told a tool was used. Where logs live. Module 0’s one-pager is the start. Door C on the homepage is this exact pain: staff pasting data they should not, plus a policy HR will sign.

Enter the free classroom

Watch Module 0 — Why AI Safety Is Not Security — with no login. Write and quiz stay in your classroom. Unlock M1–M5 for $199. If you are a deployer, also read the EU AI Act deployer page.