Automated tools can make recruiting faster, but speed is not the same as quality. A résumé ranker, chatbot, recorded-interview score, assessment, sourcing recommendation, or “fit” score can shape who receives human attention. Employers remain accountable for the resulting employment process even when a vendor owns the model.
Start with the decision, not the product label
California’s employment regulations regarding automated decision systems address technology used to make or facilitate employment decisions, including recruitment, screening, hiring, promotion, and decisions about pay, benefits, or leave. Create an inventory based on function.
- What data enters the tool, and where does it come from?
- Does the tool score, rank, filter, recommend, predict, or generate interview content?
- Can a recruiter override it, and is that override visible and reviewed?
- Which applicants never reach a person because of the output?
- Does the vendor retrain or change the system without approval?
Review four risks before launch
1. Discrimination and adverse impact
A neutral-looking factor can produce an unequal result. Test the actual workflow and applicant population, not only the vendor’s general benchmark. Define who reviews results and what action follows when a disparity appears.
2. Disability access and accommodation
The EEOC warns that algorithmic tools may screen out qualified applicants with disabilities or create prohibited disability-related inquiries. Tell candidates how to request an accommodation and give recruiters a documented alternative process that does not penalize the request.
3. Explainability and human oversight
“A recruiter makes the final decision” is not meaningful oversight if the recruiter sees only candidates the tool already filtered. Identify the stage where a person can see, question, and reverse the system’s influence.
4. Records and defensibility
Retain job criteria, configuration, output, human decisions, accommodation handling, testing, and relevant vendor documentation for the required period. A screenshot of a vendor dashboard rarely explains a past decision after the product changes.
Questions to ask the hiring-tech vendor
- Which inputs most influence the result, and which inputs are prohibited?
- What job-related validation supports the intended use?
- What demographic or accessibility testing has been performed, on what population, and when?
- How are model, feature, and threshold changes communicated?
- Can we export decision records in a usable format?
- What subcontractors or third-party models process applicant data?
- What happens to customer and applicant data after termination?
A workable governance rhythm
- Before purchase: HR, legal, recruiting, IT, and privacy agree on the intended use and unacceptable uses.
- Before launch: validate the workflow, train users, document accommodation steps, and define monitoring measures.
- During use: review outcomes, overrides, complaints, false positives, and process changes at a set cadence.
- At renewal: confirm whether the product, model, data, contract, or regulations changed.
The smallest effective control is an owner, a review date, and evidence. A policy that says “use AI responsibly” without those elements is difficult to operate.
Common questions
Can a California employer use AI in hiring?
Yes, but buying a tool does not transfer the employer’s responsibility. Evaluate the system under applicable employment, disability, privacy, and recordkeeping requirements with qualified counsel.
Does a résumé filter count?
Potentially. Focus on whether it uses computation to make or facilitate an employment decision—not whether marketing calls it AI.
Is human review enough?
Only when it is informed and real. A person should understand the tool’s role, have access to relevant context, be able to disagree, and have that disagreement handled consistently.