California's automated-decision system (ADS) regulations took effect October 1, 2025. Here's what they cover, what they require, and how to build a hiring process that can actually defend itself, with an honest note on what technology can and can't do for you.
Artificial intelligence (AI) has quietly become part of how most companies hire. Resume screeners, chatbots, automated knockout questions, scored assessments, and AI interviewers now sit between applicants and jobs at scale. California has decided that these tools don't get a pass on decades of anti-discrimination law, and as of October 1, 2025, employers using them need to understand exactly how the rules apply.
The California Civil Rights Council finalized regulations under the Fair Employment and Housing Act (FEHA) that make one thing unmistakable: an automated system used to make or help make an employment decision is subject to the same anti-discrimination obligations as a human doing the same job. This article breaks down what the rules mean, what they cover, and what they require.
What the rules actually say
At their core, the regulations clarify a principle rather than invent a new one. FEHA has long prohibited employment discrimination on the basis of protected characteristics. The new rules confirm that this prohibition applies fully when the decision is made or assisted by an automated-decision system (ADS) - and, critically, that an employer can be liable even when it never intended to discriminate, if the tool produces a discriminatory outcome.
That's the shift most employers underestimate. This is a disparate-impact regime. A facially neutral tool that disproportionately screens out a protected group is a problem regardless of anyone's intent, unless the employer can show the practice is job-related and consistent with business necessity.
What counts as an "ADS"
The definition is deliberately broad, and this is where most scoping mistakes happen. An ADS is any computational process that makes a decision or facilitates human decision-making regarding an employment benefit, and it may be derived from or use artificial intelligence, machine learning, algorithms, statistics, or other data-processing techniques.
Two features of that definition matter enormously:
- It's about function, not sophistication. The tool doesn't have to be "AI." A plain, deterministic knockout question that automatically rejects applicants - the kind that has lived in every applicant-tracking system for fifteen years - is an ADS in California, even with zero machine learning in it. If it's automated and it gates candidates, it counts.
- "Or facilitates" is the wide part. A tool doesn't have to make the final call. If it produces a score, ranking, or recommendation that a human then relies on, it "facilitates" the decision and is covered. "We only use it to assist; a person decides" is not, by itself, an exit from the rules.
The regulations offer a helpful illustrative list. Covered tools include those that screen resumes, administer tests or assessments, measure skills or "cultural fit," direct job advertisements to targeted groups, analyze facial expressions or speech patterns in interviews, and evaluate applicant data from third parties. The only carve-outs are trivial utilities - word processors, spreadsheets, calculators, spellcheck - and only when they aren't making an employment decision.
Who and what it covers
- Who: Employers with five or more employees doing business in California. The trigger follows California employment - where the work is located and performed - not the applicant's home address. A California-based role is covered; a candidate who merely lives in California and applies to a role located entirely elsewhere generally is not what these employment rules reach. (Remote work blurs this line, and is a fact-specific question worth legal input.)
- What decisions: Not just hiring. The rules reach across the employment lifecycle - recruitment, screening, hiring, promotion, selection for training, and other terms and conditions of employment.
- Whose conduct: The rules extend potential liability to an employer's "agents" - third parties who exercise a function traditionally performed by the employer. Notably, the final regulations narrowed this from an earlier draft: a company that merely builds or sells an ADS tool is not automatically on the hook. The exposure concentrates where a vendor, staffing firm, or recruiter actively performs a hiring function on the employer's behalf. The distinction is between providing a tool and doing the selecting.
What the rules require and prohibit
- No discriminatory outcomes. The central obligation: don't use an ADS (or any selection criterion) that discriminates against applicants or employees based on a protected characteristic. Disparate impact counts, intent or not.
- Watch the proxies. The rules newly define a "proxy" as a characteristic closely correlated with a protected class, and prohibit using one to do indirectly what you couldn't do directly. ZIP code standing in for race is the textbook example; graduation year (age) and continuous-availability requirements (disability, religion, caregiving) are common traps.
- Specific high-risk capabilities. The regulations single out tools that analyze tone of voice, facial expressions, or physical characteristics in interviews - these can adversely affect candidates by national origin, disability, or race, and may require a reasonable accommodation. Tools measuring reaction time, dexterity, or similar abilities carry parallel disability risk.
- Existing FEHA limits now explicitly reach ADS. An automated system may not inquire into criminal history before a conditional offer, and may not conduct medical or psychological inquiries prohibited under FEHA. Automating the question doesn't launder it.
- Anti-bias testing - not mandated, but decisive. Here is the part employers most often misread. California does not require you to test your ADS for bias, and does not require an independent auditor. Instead, testing is evidence. Proactive anti-bias testing can support an affirmative defense to a discrimination claim - and the absence of testing can be used against you. In a dispute, courts and the agency will weigh the quality, efficacy, recency, and scope of your testing, its results, and - importantly - how you responded to what it found. Testing that surfaces a disparity you then ignored is worse than no testing at all.
- Four-year recordkeeping. The rules extend the record-retention period from two years to four, and the scope is wide: "ADS data" includes the data used in and resulting from the tool, plus data used to develop or customize it. In practice, employers need to retain enough - inputs, selection criteria, outputs, and the human decision - to reconstruct and defend an individual decision years later. Most organizations don't keep this today, and it's the most concrete, immediately effective obligation in the entire regime.
- Reasonable accommodation. Where a tool interacts with a protected characteristic - a speech-analysis interview and a candidate with a disability, for instance - employers may need to offer an accommodation or an alternative process.
How California compares to New York City
Employers familiar with NYC's Local Law 144 should note the contrasts, because they cut both ways. California is easier on process: there's no independent-auditor mandate, no prescribed public audit summary to publish, and no waiting period before evaluating candidates. But California is broader and heavier on substance: its ADS definition sweeps in deterministic tools NYC's law doesn't reach, its recordkeeping burden is greater, and its testing standard - though procedurally lighter - is substantively demanding, because the quality of your testing is what actually defends you. NYC wants an audit to exist and be posted; California wants your program to hold up under scrutiny.
What employers should do now
The regulations are already in effect, so the work isn't hypothetical. A sensible starting sequence:
- Inventory your tools. Identify every automated system in your hiring process, including the "simple" ones, and map which touch California-located roles.
- Fix recordkeeping first. It's the live obligation with a running clock. Make sure you can reconstruct any ADS-assisted decision for four years.
- Stand up bias testing. Build a documented, recurring testing program that follows a recognized methodology and - the load-bearing part - records what you did when it found something.
- Guard the proxies. Audit your screening criteria for ZIP code, graduation year, and other correlates of protected class.
- Review your vendors. Understand where third parties act as your agents, and make sure contracts allocate testing, recordkeeping, and liability sensibly.
- Give notice and an alternative. Disclose AI use to candidates and provide a genuine accommodation or human-review path.
What "built to defend" looks like in practice
Full disclosure: this is the problem my company works on, so weigh the next few paragraphs accordingly. The design principles hold whatever platform you use, and the honest limits matter more than the features.
The California rules reward employers who can do four things: show their tools were tested, reconstruct any individual decision, keep a human genuinely in the loop, and prove the employer, not the vendor, set the criteria and made the call. A purpose-built platform can make each of those far easier. It's the thinking behind how we built aiAvenu and our AI interviewer, Eva:
- Reconstruct-any-decision recordkeeping. The platform retains each candidate's inputs, the exact scoring configuration, the output, and the human decision, versioned, for the four-year window. Thus, "show us how this person was evaluated in 2026" still has an answer in 2029.
- Testing you can actually produce. Rather than leaving every employer to invent bias testing, the platform runs it centrally and generates a dated, methodology-documented artifact you can hand to counsel, the exact evidence the affirmative defense turns on.
- A design that keeps humans deciding. Eva defaults to an Evidence View, structured findings against the role's stated requirements, with a required human decision, rather than a single automated score. And because it evaluates interview transcripts rather than faces or voices, it sidesteps the facial- and voice-analysis capabilities the California rules specifically flag.
- Guardrails against the obvious mistakes. The human-decision step can't be skipped, and every threshold and rule is recorded as the client's decision with a timestamp, which supports the employer's "we decided" position and keeps the vendor on the right side of the agent line.
Now the part that matters more than any feature above:
None of it is bulletproof, and any vendor who tells you otherwise is selling you a liability. A well-built platform gives you instruments and evidence. It does not give you immunity. If an employer configures a knockout that screens out a protected group, ignores a test result showing disparity, rubber-stamps the AI instead of genuinely reviewing, or points the tool at a use it was never validated for, the technology won't save them. And the same recordkeeping that defends good practice will faithfully document the bad. A test result showing bias that you did nothing about is not a defense; it's an exhibit.
That isn't a flaw in the design; it's the point. These rules are outcome-based on purpose. Good tools can make you defensible, but only you can make you compliant. The right way to read a platform like ours is as scaffolding for a defensible process that is tested, documented, and human-supervised, not as a certificate that makes the obligation someone else's problem.
The bigger picture
California's ADS rules are not the whole story. Separate tracks, including proposed legislation on automated decision-making in the workplace and the California Privacy Protection Agency's rules on automated decision-making technology, add their own notice and opt-out obligations, and some of those turn on California residency rather than job location. Employers building a compliance program should treat the FEHA ADS regulations as the foundation, not the ceiling.
The throughline is simple, even if the details aren't: California has decided that "the algorithm did it" is not a defense. Employers who treat their hiring tools with the same rigor they'd apply to any other selection procedure - tested, documented, and defensible - will be well-positioned. Those who assume automation is a shortcut around anti-discrimination law are in for an expensive correction.
Source: California Civil Rights Council employment regulations on automated-decision systems under the Fair Employment and Housing Act (2 CCR §§ 11008 et seq.), effective October 1, 2025. Disclosure: The author leads aiAvenu, which builds hiring technology discussed in this article. This article is general information, not legal advice. Several areas, particularly remote-work coverage and the boundary between providing and operating a hiring tool, are genuinely unsettled, and application depends on the specific facts. Consult qualified California employment counsel before making compliance decisions.
Kaylynn Kim, Esq. is Founder & CEO of aiAvenu. This article was first published on LinkedIn.
Related reading
- Can Your Hiring Software Explain Its Own Decisions? Starting January 1, California Says It Has To (also on LinkedIn)
- The Software Rejected Them, Not Us
#AIinHiring #CaliforniaADS #FEHA #AutomatedDecisionSystems #HRCompliance #BiasAudit #EmploymentLaw #ResponsibleAI #TalentAcquisition #HRTech




