California Limits Automated Discipline and Termination: What SB 947 Requires and What Remains Unclear

California has enacted Senate Bill 947, a law that stops employers from letting software make discipline and termination decisions on its own. The governor signed it on Sept. 30, 2026, and its requirements become operative on July 1, 2027. For employers managing workplace injuries, the law adds another consideration when attendance, productivity or other automated tools contribute to decisions affecting injured employees.

What The Law Requires

An employer may not rely solely on an automated decision system to discipline or fire an employee. If the employer primarily relies on the system’s output, a person must corroborate the decision using the data collected or used to produce that output, or other relevant supporting information. The statute expressly permits supervisory or managerial evaluations, personnel files, employee work product, peer reviews and witness interviews, which may include relevant online customer reviews. Output that cannot be corroborated, or that the reviewer finds inaccurate, incomplete or misleading, may not be used to make the decision. A manager who simply approves what the system recommends without that corroboration does not meet the standard.

The law also limits what these systems may be used for. An employer may not use one to prevent compliance with or violate labor, workplace safety, employment or civil rights laws. Nor may it use one to infer an employee’s protected status under the Fair Employment and Housing Act, such as race, sex, disability or other protected characteristics, or to predict and take adverse action against a worker for exercising legal rights.

Employees also get transparency when an employer primarily relies on an automated decision system to make a discipline or termination decision. At the time the employer informs the employee of the decision, it must provide a separate, plain-language written notice in the language it uses for routine employee communications. The notice must state that the employer primarily relied on an automated system, that a human reviewed the decision and corroborated the output and that retaliation for exercising rights under the law is prohibited. It also must provide contact information for a human the employee can contact for more information about the decision and the right to access a description of the employee’s own data the system used.

Employees may request a meaningful, objective description of their own data the system used in that decision. The description must anonymize the personal information of customers, other employees and other individuals. The law also expressly prohibits retaliation against employees who exercise its rights, complain about violations or assist in enforcing such rights.

The law covers private and public employers. The University of California is covered only if the regents agree. The collective bargaining exception requires an agreement that explicitly waives the law in clear and unambiguous terms, addresses wages, working conditions and other terms of work, and provides protection from algorithmic management. A separate, limited exception applies to uses required by or reasonably necessary to comply with federal statutes, regulations or binding federal contracts relating to aircraft development or national security, military, space or defense products or services.

Potential Implications For Workers’ Compensation

The requirements of SB 947 govern automated employment decisions; it is not a workers’ compensation statute. Its requirements, however, could overlap with workers’ compensation disputes when an automated system contributes to discipline or termination. The consequences would depend on the facts, the decision the system influenced and whether the system was used in a way covered by the statute. An automated recommendation does not eliminate the need to evaluate the employee’s injury, restrictions and protected activity under existing law.

Workers’ compensation discrimination. California’s workers’ compensation anti-discrimination provisions address specified conduct connected to an employee’s workers’ compensation claim. The relevant question is whether the employer took covered action for a reason prohibited by that statute. If an automated system was used to identify an employee because the employee filed, or intended to file, a workers’ compensation claim, the system’s role and the employer’s decision process could be relevant evidence.

Modified duty and temporary disability. A failure to return to work after an employer offers modified or alternative work could mean that an employee does not receive temporary disability benefits under some circumstances. In such instances, an employee might be disciplined or even terminated. If an automated tool informed the offer, assessed the employee’s restrictions or contributed to a later employment action, those facts could matter to the employment-law analysis and to the evidentiary record.

Psychiatric injury claims. A personnel action connected to an automated system could also be relevant to a psychiatric injury claim. For injuries within its scope, LC § 3208.3(h) bars compensation for psychiatric injury if it was substantially caused by a lawful, nondiscriminatory, good-faith personnel action. The employee first has the burden to establish a psychiatric injury. Then the employer has the affirmative burden to establish the defense. That requires evidence supporting the required elements, not merely an assertion that the action was routine or automated.

When SB 947 applies in psychiatric injury claims, compliance or noncompliance with its human-review and notice requirements could be relevant to whether a personnel action was lawful and how it was carried out. A violation could undermine the defense’s lawfulness element. Compliance, however, would not by itself establish that the action was lawful in every other respect, was nondiscriminatory, in good faith or a substantial cause of the injury. Those remain distinct questions under LC § 3208.3(h). The circumstances of the investigation, the employee’s opportunity to respond and the treatment of similarly situated employees could bear on the defense .

What Is Not Well Defined

Several aspects of SB 947 are likely sources of dispute, even though the enacted text resolves a number of threshold questions. Employers should distinguish between requirements the statute specifies and issues that remain open to interpretation.

One issue likely to be in dispute is the reach of the definition of an automated decision system. SB 947 supplies its own definition, covering computational processes derived from machine learning, statistical modeling, data analytics or artificial intelligence that produce a score, classification, recommendation or other simplified output used to assist or replace human discretionary decision-making and that materially affect individuals. It excludes calculators, databases, datasets and certain security and access-management tools. A points-based attendance system is an obvious test case — merely storing or totaling attendance data is different from using analytics to recommend discipline. The system’s actual functions, rather than its marketing label, will matter.

Another aspect of SB 947 vulnerable to dispute is what it means to “primarily rely” on an output. The corroboration duty applies when the employer primarily relies on the system’s output for a disciplinary or termination decision, so employers might be tempted to describe that output as one factor among several. How much weight makes that reliance primary is likely to be litigated. Labeling the output a secondary consideration will not necessarily resolve the question if the record shows that it drove the result.

A third aspect is that the statute identifies permissible sources of corroboration but does not prescribe a detailed review process, a training requirement or a documentation checklist. How searching the review must be, and when the same underlying data meaningfully corroborates an output, remain practical questions. A review that confirms the arithmetic but overlooks injury-related restrictions or protected absences could leave the employer exposed under other employment laws even if the calculation is accurate.

The prohibitions on inference and prediction raise their own questions. The hard cases involve proxies, such as a variable that correlates with disability or pregnancy without expressly identifying either characteristic. How intent and the system’s design affect liability will require interpretation. The law broadly protects the exercise of legal rights — employers should not assume that workers’ compensation claims, protected medical leave or wage complaints fall outside its reach.

The notice provisions specify both timing and required content. The data-request provision is less detailed — it does not set an express response deadline or fully explain how detailed a “meaningful, objective description” must be. Although the statute requires anonymization of other individuals’ information, practical questions remain about how employers will obtain usable explanations from vendors, and address asserted trade-secret concerns.

Enforcement is specified, rather than left entirely open. The Labor commissioner may investigate, issue citations and bring civil actions. Public prosecutors also may enforce the law. Violations carry a $500 civil penalty per violation, and the statute provides for additional relief, including attorney fees and costs and, in authorized civil actions, punitive damages. It also contains an employer burden-of-proof provision: In specified proceedings, once use of an automated system in a disciplinary or deactivation decision is demonstrated, the employer must show that it did not primarily rely on the system or that it complied with the applicable review and notice requirements. The statute does not expressly create a stand-alone private cause of action. How these provisions interact with other available claims and remedies, how separate violations will be counted and how relief will be applied remain important questions.

Other questions involve at-will employment and timing. An employer’s ability to terminate without stating a reason does not dispense with SB 947 when a covered system is used, and the burden-of-proof provision makes records of the actual decision process especially important. Although the collective bargaining and federal exceptions have stated conditions, their application to particular agreements and operations may be disputed. Employers should also plan for covered decisions made on or after July 1, 2027, even when the system uses data collected earlier — the operative date should not be treated as a blanket exemption for older data.

What Employers Should Do Now

Before July 1, 2027, employers should inventory the automated tools that feed discipline and termination decisions. They should build a documented human-review step, prepare the required written notices and create a procedure for data requests. For injured employees, that review should account for medical restrictions, protected absences and workers’ compensation activity rather than accepting attendance or productivity scores at face value. Employers also should confirm that vendors can supply the underlying information needed for corroboration and employee data descriptions.

Because application of the law will turn on how a decision was actually made, documentation should identify the reviewer, the information examined, any errors or limitations identified and the basis for the final decision. The most useful record will show a genuine assessment of the employee’s circumstances and the evidence supporting the action. That record could matter both to compliance with SB 947 and to the defense of a later workers’ compensation or employment claim.