California AI hiring law news has centered on the workplace AI bills Governor Newsom signed on September 30, 2026. The headline bill governs how AI is used to discipline and fire workers. Hiring and termination are two ends of the same decision chain: a tool that helps decide who gets in can also help decide who gets disciplined or let go, and California now expects evidence at both ends.
For HR technology buyers, that raises a concrete question. How does a bias audit detect discrimination, which outcomes does it compare, and what can that comparison establish?
Talk with Warden AI about independent evidence for employment AI
California's new measures do not create one universal audit form. Senate Bill 947, the No Robo Bosses Act, addresses automated decision systems used in discipline and termination. Senate Bill 951 adds technology information to some Cal-WARN notices. Assembly Bill 1883 limits AI workplace surveillance that reads emotions or collects neural data, and Assembly Bill 1331 restricts surveillance in workplace bathrooms. None replaces a careful test of selection outcomes for hiring, promotion, or other employment decisions.
What California AI Hiring Law News Means for Bias Audits
A bias audit is an evidence review of how an automated tool behaves for different groups. The reviewer starts with a defined decision, such as who advances from a resume screen or who receives a recommendation for an interview. The reviewer then compares the tool's outputs, and the human decisions that follow, across relevant demographic groups.
The basic unit is a rate, not a label. If 100 applicants from Group A enter a process and 60 advance, the selection rate is 60 percent. If 100 applicants from Group B enter the same process and 30 advance, the rate is 30 percent. The impact ratio for Group B is 30 divided by 60, or 50 percent. That difference is a signal for investigation. It is not, by itself, a final legal conclusion.
The Uniform Guidelines use a practical indicator called the four-fifths rule. A selection rate for a race, sex, or ethnic group that falls below four-fifths of the highest group's rate generally calls for a closer look.
29 C.F.R. section 1607.4 states that rule. It also says smaller differences can still show adverse impact when they are significant in both statistical and practical terms. A larger gap is not automatically adverse impact. Sample size and the facts of the process still matter.
The EEOC's questions and answers on the Uniform Guidelines frame the same point. A selection procedure with an adverse impact is discriminatory unless it is validated or otherwise justified under federal law.
How Do Bias Audits Detect Discrimination?
A useful audit does not begin with a claim that a model is biased or unbiased. It begins by defining the decision, the comparison groups, the time period, and the data needed to reproduce the result. The detection method runs in seven steps. The adverse-impact testing guide covers each one in full.
- Define the decision and population. Name the job, location, process stage, model version, and dates. A vague population makes a precise-looking rate misleading.
- Separate outcomes by group. Count who was screened out, advanced, interviewed, hired, promoted, or retained. A resume ranker and a scheduling tool are different decisions.
- Compare selection rates. Compare each group with the highest rate or another justified reference group, at the model stage and at the final human decision.
- Test uncertainty and sample size. Record counts and a test of statistical and practical significance, not the four-fifths threshold alone. Report small samples as inconclusive when needed.
- Check intersections and controls. Examine intersectional groups where data and privacy safeguards allow, and test whether job, location, recruiter, or workflow explains the gap.
- Trace the result to the product and process. Identify which features, thresholds, prompts, or human overrides shaped the outcome.
- Record the conclusion and limits. State the population, method, findings, limitations, remediation, and retest date. Separate evidence of a disparity from proof of intent.
The process detects a pattern in observed outcomes. It does not read intent from code, and it does not prove that a tool caused every unequal result. It gives the employer and vendor a disciplined way to find a disparity, test plausible explanations, and decide whether the tool or process needs remediation.
What Evidence Should an HR Tech Buyer Request?
Buyers should ask for evidence that can be understood and reproduced. A polished statement that a system was tested is weaker than a specific report. The useful report names the decision stage, groups, counts, rates, impact ratios, statistical treatment, and limits.
California's Civil Rights Department regulations sit beside this evidence review. The Civil Rights Council's June 30, 2025 announcement says the Office of Administrative Law approved the rules on June 27, 2025. They took effect October 1, 2025.
The rules explain how existing antidiscrimination law applies when AI is used in hiring, promotion, and other employment decisions. Employers must keep automated-decision data and related records for at least four years. The rules do not turn every audit into a named annual filing. The finalized California FEHA regulations guide covers those duties.
A bias audit can show disparate outcomes. It cannot, by itself, prove that a tool meets every FEHA obligation or that a selection procedure is job-related. It also cannot clear the employer's whole process. Federal law still looks at adverse impact in context, including whether a procedure is validated. Buyers should treat an audit as evidence for legal review, not as a compliance certificate.
How California's AI Employment Laws Change the Evidence File
SB 947 changes the evidence question for automated decision systems used in discipline and termination. The authenticated chaptered text, Chapter 859, was approved by the Governor and filed with the Secretary of State on September 30, 2026. It adds Labor Code Part 5.5.5, beginning with Section 1520, and the operative date is July 1, 2027.
The statute defines an automated decision system as a computational process that issues a simplified output. The process may draw on machine learning, statistical modeling, data analytics, or artificial intelligence. The output includes a score, classification, or recommendation that assists or replaces human discretion and materially affects a person. Spam filters, firewalls, antivirus software, and identity management tools are excluded.
Section 1522 sets use limits, not an audit filing. An employer may not use an ADS to violate labor, safety, employment, or civil rights law. An employer may not use an ADS to infer protected status under Government Code Section 12940. An employer also may not use an ADS to predict and punish the exercise of legal rights, and may not rely solely on an ADS for a disciplinary or termination decision.
When an employer primarily relies on an ADS output for that kind of decision, the statute requires a human to corroborate it. The listed examples of corroborating information include supervisory or managerial evaluations, personnel files, employee work product, peer reviews, and witness interviews. If the output cannot be corroborated, or the reviewer finds it inaccurate, incomplete, or misleading, the employer may not use it for the decision. The employee may request a meaningful, objective description of the employee's own data used by the ADS. Other people's personal information must be anonymized in that description.
Section 1524 then requires a written post-use notice when the employer primarily relied on an ADS for the decision. The notice is a separate, plain-language communication. It must say that the employer primarily relied on an ADS and that a human reviewed and corroborated the output. It must also name a human contact and state that retaliation for using these rights is prohibited.
Violations carry a civil penalty of $500 and may be enforced by the Labor Commissioner or through a civil action, according to Covington's analysis. A narrow carve-out covers aircraft development and certain national security, military, space, or defense work. It applies only when ADS use is reasonably necessary to meet a federal statute, regulation, or binding federal contract.
Those duties are not the same as a hiring bias audit. A vendor may need to preserve what output was used, what data supported it, who reviewed it, and what the reviewer decided. A selection-rate analysis can still help for hiring or promotion. It does not prove that an employer corroborated a termination or sent the required notice. Which laws name a bias audit is covered in the guide to bias audit legal requirements.
SB 951, AB 1883, and AB 1331 were signed in the same September 30, 2026 package, and they create different records. The October 2026 Morgan Lewis analysis summarizes SB 951 and AB 1883. SB 951 adds Cal-WARN disclosure when AI or automation substantially causes a covered layoff, relocation, or termination. AB 1883 takes effect January 1, 2027. It restricts AI workplace-surveillance tools used to infer emotional state or to collect neural data, with limited exceptions, and it gives employees a private right of action. AB 1331 separately restricts surveillance tools in workplace bathrooms.
The Governor's September 30, 2026 signing announcement confirms the package. A sentiment or tone feature needs its own legal and product review. A bias audit should name the function that was tested rather than treating every AI tool as one use.
Where AB 1405 Fits: A Registry for AI Auditors
A separate bill, AB 1405, signed on September 9, 2026, addresses the auditors rather than the employers. It directs the state to build a registry of AI auditors with standards for independence, transparency, and integrity. It also sets what an audit report must contain, including the results, the documentation behind them, and the limits of the audit. Its author described it as a bill that "does not require auditing of AI systems and models." The registry is due by January 1, 2029.
As described in the Senate Judiciary Committee analysis, the registry rules reach audits conducted under a state statute that requires one. No California statute today requires an independent bias audit of hiring tools. For employment AI, AB 1405 is best read as a benchmark: buyers can ask whether their auditor already meets independence and disclosure standards of this kind.
What Bias Audits Cannot Establish Alone
A disparity is a warning signal, not a complete explanation. The groups may have entered different jobs, locations, or applicant pools. A small group may produce unstable estimates. A human may have changed the model's recommendation. A legally relevant characteristic may be missing, incomplete, or protected by privacy rules. These limits do not make testing optional. They determine how carefully the result must be interpreted.
An audit also cannot establish discriminatory intent from an impact ratio alone. A low ratio can prompt a review of job-relatedness, validation, data quality, workflow design, and possible remediation. It can support a buyer's due diligence and help a vendor find a product defect. It should not be converted into a promise that the system is compliant in every jurisdiction or decision context.
Passing one test does not guarantee future performance. Models, applicants, jobs, thresholds, and recruiters change. A responsible evidence program sets a monitoring interval and a trigger for retesting after a material model, data, or workflow change. It keeps versioned records so a later reviewer can tell which system produced which result. The trade-offs between point-in-time and continuous testing are covered in AI bias audit frequency.
What HR Tech Vendors Should Prepare Now
Vendors selling into California should prepare a decision map before they prepare a marketing claim. The map should show where the product can influence hiring, promotion, discipline, termination, layoffs, scheduling, monitoring, or access to an employment opportunity.
- Identify the output. Record whether the product produces a score, rank, recommendation, classification, alert, or generated explanation. Define what a customer can do with that output.
- Preserve the evidence path. Keep the model version, relevant inputs, output, timestamp, threshold, and any human override available under an appropriate privacy and retention policy.
- Support subgroup analysis. Make it possible for an authorized reviewer to compare outcomes by relevant group without exposing unnecessary personal information.
- Explain limits. State what the test covered, what it did not cover, how small samples were handled, and whether the review examined intersections or downstream human decisions.
- Separate use cases. Do not use a hiring selection-rate report as proof about an emotion-recognition feature, a layoff notice, or a disciplinary decision. Each use requires its own evidence and legal analysis.
- Plan remediation. Identify who can change a threshold, retrain a model, pause a feature, or require human review. Schedule a retest and retain the prior result for comparison.
Independent assurance can make this evidence easier for an enterprise buyer to evaluate. Warden AI's work with HR tech vendors is third-party testing and documentation of employment AI. It is not a software seat that makes a product compliant with SB 947.
The AI audit trail resource covers how to keep that evidence path defensible. Neither a report nor a service changes the employer's duties under California law.
Review your California employment AI evidence with Warden AI
No Robo Bosses Act and Bias Audits: Frequently Asked Questions
How do bias audits detect discrimination?
They compare outcomes across relevant groups at a defined decision stage. The reviewer calculates selection rates and impact ratios, then checks sample size and statistical uncertainty. The review looks at intersectional groups when the data support it, and asks whether the tool or the workflow explains the gap. The result is evidence for further review, not an automatic legal finding.
Does California SB 947 require a hiring bias audit?
No. SB 947 concerns automated decision systems used in discipline and termination, beginning July 1, 2027. It addresses sole reliance, human corroboration, a description of the employee's own data, notice, and related rights. It does not create a named annual bias-audit filing for hiring tools.
Does California require AI bias auditors to register?
Not yet. AB 1405, signed September 9, 2026, directs the state to create an AI auditor registry by January 1, 2029, with independence and reporting standards. It does not require any company to audit an AI system, and no California statute currently requires an independent bias audit of hiring tools.
Is the four-fifths rule a legal safe harbor?
No. Section 1607.4 treats a rate below four-fifths of the highest group's rate as a practical indicator of adverse impact. A result above that line does not make a process lawful. A smaller difference can still matter when it is significant in both statistical and practical terms. Validation and the rest of the employment process still matter.
What should a vendor include in an audit report?
The report should name the decision, population, product version, and dates. It should include groups, counts, selection rates, impact ratios, uncertainty analysis, controls, and limits. It should also state the findings, the remediation, and the retest plan, and whether the review covered only model outputs or later human decisions.



