Ban the Box in the Age of AI Hiring

For two decades, the fair-chance movement worked to take one small box off the job application: the one that asks, before anyone has read your resume, whether you have a criminal record. Dozens of states and more than 150 cities and counties now restrict when and how an employer can ask. The logic is simple and humane: judge people on what they can do, not on a checkbox that ends the conversation before it starts.

Artificial intelligence is quietly threatening to put that box back, not on the form, but inside the algorithm.

What fair-chance hiring actually requires

"Ban the box" is shorthand for a patchwork of state and local fair-chance laws. The details differ, but the core rules rhyme: an employer cannot ask about criminal history until later in the process, often not until a conditional offer. On top of that sit two federal pillars. The Equal Employment Opportunity Commission has long warned that blanket use of criminal records can create unlawful disparate impact, and expects an individualized assessment of the offense, the time passed, and its relevance to the job. The Fair Credit Reporting Act governs background checks themselves, including the timing of disclosures and the adverse-action notices a candidate is owed before being turned away.

The throughline of all of it is timing and transparency. Certain information must stay out of the early decision, and when it does come in, the candidate has a right to know and to respond.

Where AI breaks the model

Automated hiring tools are built to do the opposite of what fair-chance law demands. They evaluate everyone, early, at scale, on signals no human reviewer explicitly chose.

• They surface what the law tried to delay. Sourcing and screening tools scrape and rank candidates from public data long before a conditional offer. A model does not respect the conditional-offer line.
• They infer what they were never told. Even a tool that is never fed a criminal record can learn proxies for one, gaps in employment, ZIP codes, name-based correlations, and weigh them invisibly. The record is "considered" without ever appearing.
• They make compliance impossible to prove. Fair-chance and EEOC compliance depend on showing your work: the individualized assessment, the timing, the adverse-action steps. An opaque model produces a score, not a defensible record. "We did not know the system factored it in" is not a defense; it is an admission.

This is exactly the territory that new AI-hiring rules are starting to police. New York City's Local Law 144 now requires bias audits and candidate notice for automated employment decision tools, and more jurisdictions are following. The compliance question is no longer just "did we ask too early," but "can we explain what the machine decided, and prove it was lawful."

What GRC and compliance teams should do now

1. Inventory every AI tool in the hiring path, including sourcing and ranking tools that HR may not think of as decision tools.
2. Demand transparency from vendors on what signals each model uses, and treat "proprietary" as a red flag, not a reassurance.
3. Map tools to the fair-chance timeline. Anything that can surface or infer criminal history before a conditional offer is a liability, regardless of intent.
4. Keep a human in the adverse-action loop, with a documented, individualized assessment that a regulator could read.
5. Audit for disparate impact on the model's outputs, not just its inputs.

The principle underneath

Fair-chance laws exist for the same reason the best privacy practices do: a person deserves to be judged on their qualifications, by a process they can see and answer, not by an invisible automated verdict rendered before they ever get a fair look. The box came off the form. The work now is making sure it does not quietly come back inside a model no one can audit.

That is not just compliance. It is dignity, and it is the standard every employer, and every hiring tool, should be held to.