Status note: SB 947 is enacted and chaptered California law, not a proposal. Governor Gavin Newsom approved it on September 30, 2026. It was filed with the Secretary of State that day. The chaptered text was published on October 1, 2026. The workplace ADS rules are not operative until July 1, 2027. No use, enforcement, or court result was found by October 4, 2026. This review cannot prove that no court activity exists.

Executive Summary

California has enacted SB 947. It limits how employers may use automated decision systems in employee discipline and firing. An automated decision system, or ADS, uses tools such as machine learning, statistics, data analysis, or AI. It makes an output that helps or replaces human judgment and has a major effect on people.

The law does not ban workplace AI. An employer may not use a covered ADS as the sole basis to discipline or fire an employee. A different rule applies when the employer “primarily relies upon an ADS output.” A human must corroborate the choice with supporting information. The output cannot be used if it cannot be corroborated. It also cannot be used if the human concludes that it is inaccurate, incomplete, or misleading. Separate provisions use slightly different triggers for notice and the employee-data description, as explained below.

The law does not define “primarily relies.” It sets no detailed review method. Nor does it set minimum standards for reviewer independence, qualifications, workload, or protection from pressure. A reviewer’s conclusion that the output is inaccurate, incomplete, or misleading bars its use. Whether workplace practice will support candid adverse findings is untested. The rules are not operative until July 1, 2027. Their effects remain unknown.

What Happened?

Governor Newsom approved SB 947 on September 30, 2026. It was filed with the California Secretary of State as Chapter 859 that day. California Legislative Information published the chaptered text on October 1, 2026. The first date is the event date; the second is the source publication date.

The new Labor Code part becomes operative on July 1, 2027. At the October 4, 2026 research cutoff, the rules were neither operative nor proven to work.

SB 947 covers some ADS uses in employee discipline and firing. It also bars an employer from using an ADS to prevent compliance with or violate any federal, state, or local labor, occupational-health-and-safety, employment, or civil-rights law or regulation. Another rule bars using an ADS to infer an employee’s protected status under California Government Code section 12940. A third bars using an ADS both to predict that a worker will exercise legal rights and to take adverse action against that worker for doing so. The final text does not define “worker,” so the reach of that third rule is unresolved.

Why This Matters

SB 947 draws a line between AI help and decisions made only by a machine. Employers may use covered systems. But an ADS output cannot be the sole basis for discipline or firing.

A different trigger applies to human review. When an employer “primarily relies upon an ADS output,” a person must check it against supporting information. The related provisions do not use identical words. The data-description right applies when the employer “primarily used an ADS.” The notice rule applies when the employer “primarily relied upon an ADS.” The law does not define these triggers. No agency or court has said whether their wording gives them different reach.

A “human in the loop” takes part in an automated process. Review quality matters more than the label.

How the Technology Works

The law defines an ADS by function. It may use machine learning, statistics, data analysis, or AI. It makes a score, label, recommendation, or similar output. That output helps or replaces human judgment. It must have a major effect on people.

This meaning is wider than generative AI, which makes new content. Coverage turns on a system’s function and use, not its sales label.

The law excludes spam filters, firewalls, antivirus software, and identity and access tools. It also excludes calculators, databases, datasets, and other data groups. A platform may include both an excluded security tool and a covered job tool. Coverage depends on the facts.

When an employer primarily relies on an ADS output, a human may check input data and other facts, such as manager reviews, personnel files, work products, peer reviews, and witness interviews. The output cannot support the choice if it cannot be corroborated or if the human reviewer concludes that it is inaccurate, incomplete, or misleading.

The Ethical Question

The main ethics issue is real human judgment in a major job choice. Does the person learn that automation was used? Can the person challenge bad data?

AI may make choices more even or find missed patterns. It can also use poor records, biased history, or weak measures. Review helps only if the person has enough time, skill, power, and evidence to disagree.

Potential Benefits

When the notice rule is triggered, the employer must give a written post-use notice. It is due when the employee is informed of the decision. The notice must state that the employer primarily relied on an ADS and that a human reviewed the decision and corroborated the output. It must name a human contact and explain the right to a data description. It must also state that retaliation is banned. It must use plain language and stand alone. It must use the language of routine employee communications and arrive through a simple written method.

When an employer primarily used an ADS for discipline or firing, the employee may ask for a data description. It must give a meaningful, objective account of the employee’s own data used by the ADS. The law does not define “objective” or set the needed detail. The employer must anonymize personal information about customers, other employees, and other people. The description may reveal a wrong record, mistaken identity, or other data error.

These are possible gains, not measured results. By October 4, 2026, the new part was not operative. The evidence does not show whether SB 947 reduces errors, bias, or unfair results.

Potential Risks

The first risk is review that acts as a rubber stamp. A person may accept machine advice too readily. The person may also be short on time. “Automation bias” means putting too much trust in machine advice.

A peer-reviewed study looked at past research. It found too much reliance on wrong machine advice. Most work came from health care and aviation. Several factors mattered. They included workload, trust, training, screen design, and accountability. This does not prove how California job reviews will work.

The ADS and reviewer may share an old, incomplete, biased, or wrongly linked record, so human review may repeat an error.

An employee-data description may involve personal information about customers, other employees, or other people. The employer must anonymize that information. Tight access limits are still needed.

Government Perspective

The Governor’s September 30 signing notice said SB 947 stops employers from relying only on AI for discipline or firing. This states the administration’s policy view. The chaptered law controls.

An Assembly committee analysis was prepared for a June 23, 2026 hearing. It recorded both sides. In its synopsis and background, the committee analysis discussed biased training data. It also discussed systems that can be hard to understand. Its summary of the supporter position emphasized secrecy, worker harm, mistakes, abuse, and insufficient human oversight. Opponents cited costs, lawsuits, limits on useful tools, and a possible chilling effect. A chilling effect means people may avoid a lawful act due to fear of risk or punishment. The analysis covered an earlier and wider bill version. It did not cover the exact final law. The PDF metadata shows creation and modification on June 22, 2026. The public posting or publication date is unknown.

The final text gives enforcement roles to the Labor Commissioner and public prosecutors. It sets a $500 civil penalty per violation and lists other possible relief. It does not clearly give employees a direct private right to sue under SB 947. How penalties add up, and whether other laws support private claims, remain untested.

AI Industry / Research Perspective

The cited rules place direct duties on employers, not AI vendors. Employers may still seek review tools. They may also seek vendor records on inputs, outputs, versions, and data. This expected business effect is not a direct vendor duty.

Human oversight can help. Yet a person’s presence does not prove safety. Employers can train reviewers and allow enough time. They can protect the power to disagree and check other sources. They can also track overrides, appeals, errors, and results.

The National Institute of Standards and Technology has an AI Risk Management Framework. This voluntary framework describes outcomes and practices for human roles, testing, tracking, appeals, and overrides. It is not a prescribed sequence of steps. SB 947 does not require it. Using it alone does not prove compliance.

Business Impact

California employers should find all systems that affect discipline or firing. Vendor labels should not decide coverage.

Each decision map should list the data, output, and users. It should show the weight given to the output and who can reject it. “Primarily relies” has no set meaning in the law. An in-house rule may reduce doubt until guidance or case law develops.

Reviewers need clear duties, enough time, and power to stop a choice. Employers should use routine employee languages and obtain needed vendor records and technical help.

Some businesses may also have California privacy duties. Union deals and some federal contracts affect coverage. One notice may not meet every rule. The same is true of one vendor promise or global AI policy. This is broad analysis, not legal advice.

Consumer Impact

SB 947 directly concerns employees and employers, not buyers of goods. Customer data may still appear in workplace records.

Employees may gain a human contact and a description of their own data. The law does not promise source code, model weights, raw datasets, a vendor’s name, or a full technical account. Model weights are numbers a trained AI system uses to shape its output.

The final text focuses on employees. It does not clearly cover job applicants or independent contractors just because of those roles. Cases near the employee-contractor line turn on facts. Parts of an earlier bill should not be assumed to remain in the final law.

What Happens Next?

The new part becomes operative on July 1, 2027. Its core duties apply to covered employers. Public agencies are included only when they meet the employer definition and other conditions. Vendors may face indirect requests for records or technical support. Unions matter in part through a limited, conditioned collective-bargaining exception. Before the operative date, affected groups can list systems and change processes. They can train reviewers, prepare notices, improve records, and test data corrections. Exact coverage remains fact-dependent.

Later answers may come from agencies, new laws, enforcement, or courts. The law does not define “primarily relies” or set minimum reviewer independence, training, or inquiry. Its terms also vary: one part uses “worker,” final definitions focus on “employee,” and an enforcement rule uses “deactivation” while other parts discuss firing. The legal effect is uncertain.

After July 1, 2027, useful proof may include human overrides and employee corrections. It may include results for different groups and agency views. Until then, claims that the law has worked or failed are predictions.

AI Coalition Network Perspective

AI Coalition Network supports AI that helps people make better choices. SB 947 does not ban workplace AI; it adds duties around discipline and firing.

The law’s promise turns on whether human checks are real. A signature after an automated process is not informed judgment. Strong programs will test systems and guard data. They will record human reasons, allow correction and appeal, and let reviewers say no.

Unclear rules should not make every tool look unsafe. Coverage turns on what a system does, how it is used, and how it affects people. Careful checks can guard employees while tools help managers.

Our view is clear but limited. The enacted plan creates real chances for accountability. Its results are unknown. Evidence after July 1, 2027, should show whether the safeguard works as planned.

Questions Leaders Should Ask

  1. Which systems affect discipline or firing, even if they are not sold as AI?
  2. When does our process “primarily rely” on a machine output, and how do we record that?
  3. Can the human reviewer reject the output without pressure or punishment?
  4. Does the reviewer check evidence beyond the data that made the output?
  5. How do we find data that is wrong, incomplete, old, biased, or linked to the wrong person?
  6. Can an employee reach a person, fix data, and challenge the choice?
  7. Are notices clear, separate, simple, and offered in routine employee languages?
  8. Can vendors provide records needed for review and disclosure?
  9. How do we guard coworker and customer facts during disclosure?
  10. Will we track overrides, appeals, errors, and results for different groups?

Sources

  1. California Legislative Information, SB 947 bill navigation and chaptered status. Event date: 2026-09-30. Source publication date: 2026-10-01. Status: official primary source; enacted and chaptered, not yet operative. https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=202520260SB947
  2. California Legislative Information, chaptered SB 947 PDF. Event date: 2026-09-30. Source publication date: 2026-10-01. Status: controlling official text for the cited requirements; implementation effects remain unverified. https://leginfo.legislature.ca.gov/faces/billPdf.xhtml?bill_id=202520260SB947&version=20250SB94789CHP
  3. Office of Governor Gavin Newsom, signing announcement. Event and publication date: 2026-09-30. Status: official corroboration of signing and administration perspective; statutory text controls. https://www.gov.ca.gov/2026/09/30/californias-nation-leading-ai-framework-just-got-stronger-governor-newsom-signs-more-first-in-the-nation-worker-protections-and-more/
  4. California Assembly Privacy and Consumer Protection Committee, analysis of an earlier SB 947 version prepared for the 2026-06-23 hearing. Event/hearing date: 2026-06-23. Document metadata creation and modification date: 2026-06-22. Source publication/public posting date: unknown. Status: government policy analysis and stakeholder viewpoints; not the controlling final text. https://apcp.assembly.ca.gov/system/files/2026-06/sb-947-mcnerney-apcp-analysis.pdf
  5. Goddard, Roudsari, and Wyatt, systematic review of automation bias in decision-support systems. Online publication date: 2011-06-16; issue date: 2012-01. Status: peer-reviewed contextual research, mainly healthcare and aviation; not California employment-outcome evidence. https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/
  6. National Institute of Standards and Technology, AI Risk Management Framework 1.0 Core. Publication date: 2023-01-26. Status: published voluntary guidance; not incorporated into SB 947 and not proof of compliance. https://airc.nist.gov/airmf-resources/airmf/5-sec-core/
  7. California Privacy Protection Agency, announcement on privacy and automated decisionmaking regulations. Event and publication date: 2025-09-23. Status: official government source for the separate privacy-regulation timeline; application to a specific employer requires legal analysis. https://www.cppa.ca.gov/announcements/2025/20250923.html