SB 947 makes human review of AI firing decisions a data problem

0:05Host:This is Leaders Insights. On the table: SB 947 makes human review of AI firing decisions a data problem. Most data leaders read "No Robo Bosses Act" and think HR problem, not my problem. Why are you telling me that reading is wrong?

0:20Expert:Because of one sentence in the statute. SB 947 defines an automated decision system as any computational process derived from machine learning, statistical modeling, data analytics, or artificial intelligence that issues a simplified output, including a score, classification, or recommendation, used to assist or replace human discretionary decision-making. Statistical modeling and data analytics. That is a regression in a spreadsheet. That is your attrition model.

0:53Host:Give me the basic facts first. Who signed what, and when does it bite?

0:57Expert:Governor Gavin Newsom signed SB 947 on Wednesday 30 September 2026. It bars employers from relying solely on an automated decision system when they discipline or fire a worker, and it applies from 1 July 2027. It was the final day he could sign or veto, and he signed 11 AI bills that day, closing a year with nearly 30 AI measures.

1:23Host:Nine months of runway. Is that generous or tight?

1:27Expert:Tight, if you have never inventoried your models. The law is not satisfied by a policy document. Ogletree Deakins reads it as requiring a human reviewer to independently corroborate the system's output, and Bloomberg Law reports employers must give affected workers notices about the tools used and summaries of the personal data considered in the decision. You cannot produce that summary from a dashboard screenshot.

1:52Host:So what exactly does a data team have to be able to prove?

1:55Expert:Three things, and they are all lineage questions. Lineage meaning the recorded path from raw data to the number a manager saw. Which system produced the score, which personal fields fed it, and which version of the model ran on that date. If your scoring pipeline overwrites yesterday's output, you cannot reconstruct the decision a lawyer asks about in 2029.

2:17Host:What counts as "the system decided" versus "a manager decided with help"?

2:22Expert:That is the contested line. Where an employer relies primarily on the system's output, a human reviewer has to run an independent investigation and gather evidence supporting or contradicting the recommendation: personnel files, work product, management evaluations, peer reviews, witness interviews. The word "primarily" is doing enormous work, and nobody has litigated it yet.

2:45Host:You said the corroboration has to be independent. In practice, managers rubber-stamp. Does the law stop that?

2:52Expert:On paper it tries. The written notice has to state four things: that the employer primarily relied on an automated decision system, that a human reviewed the decision and corroborated the output, contact details for a person who can explain it, and that the employer cannot retaliate against the worker for using these rights. Naming a person is the part that changes behaviour. Rubber-stamping is easier when nobody signs.

3:19Host:There is also a prohibition that sounds broader than discipline.

3:22Expert:Yes, and this one is squarely a modelling constraint. SB 947 bars using an automated decision system to infer a worker's protected characteristics or to predict and retaliate against a worker for exercising legal rights. Protected characteristics meaning race, disability, pregnancy and the rest. Inference is the operative word. You can trip it without ever collecting the attribute, through proxies like postcode or shift pattern.

3:50Host:What is the penalty? Is this a real risk or a rounding error?

3:53Expert:Covington's Global Policy Watch puts the civil penalty at $500, with the employer exposed to a civil action or administrative enforcement by the Labor Commissioner. Five hundred dollars per violation is small. The litigation and discovery exposure is not. My opinion, not a legal finding: the expensive part is being asked to produce records you never kept.

4:17Host:Second law. SB 951. Why does a layoff notice matter to a chief data officer?

4:23Expert:Because it forces attribution of cause. SB 951 expands what employers must disclose when a workforce reduction covered by Cal-WARN results from artificial intelligence or other automated technology, and it takes effect 1 January 2027. Cal-WARN is California's mass-layoff notice law. It requires 60 days' written notice, covers establishments with at least 75 employees, and a mass layoff generally means 50 or more people in a 30-day window.

4:55Host:And the new content of the notice?

4:57Expert:Fisher Phillips says the additional disclosures apply when the layoff, relocation or termination is caused "in whole or in substantial part" by an AI system or other automated technology replacing or automating positions, and Bloomberg Law reports the notice has to include the number, location and types of jobs displaced, plus the kinds of automation technology used.

5:22Host:Someone has to decide what "in substantial part" means. Who, in a real company?

5:27Expert:Finance writes the business case, HR writes the notice, and your team owns the only evidence either of them has. If your automation programme has a benefits tracker claiming 40 headcount avoided, that document now lives in a legal file. I would read every internal ROI deck as a future exhibit.

5:46Host:Give me the strongest argument that this is bad law.

5:50Expert:The California Chamber of Commerce makes it cleanly. In its opposition letter the Chamber wrote that the bill "broadly targets businesses of all sizes, across every industry, and regulates even low-risk applications of automated decision systems" and said it would drive up costs because any misstep leads to costly litigation for even the smallest employers. The scope point is fair. A scheduling tool that flags attendance is not a frontier model, and it gets caught the same way.

6:20Host:Does the bill's author have an answer?

6:22Expert:Senator Jerry McNerney told the Senate labour committee the bill requires human review when automated systems assist in discipline, termination or deactivation decisions, and bans predictive behaviour analysis that could punish people before wrongdoing. "Humans should be making these decisions," he said. Newsom's line was that AI should expand opportunity, not come at the expense of workers and families.

6:47Host:This is the second attempt. What got cut to make it survivable?

6:51Expert:Newsom vetoed the earlier version in October 2025, even though it had cleared both chambers with large majorities. McNerney reintroduced it in February 2026 and removed the provision Newsom named: SB 7 would have required notifying every worker foreseeably affected whenever such a system was in use, keeping an updated list of every system, and telling job applicants. Gig workers were also dropped.

7:18Host:The inventory requirement died. So the law now assumes companies know what they are running without asking them to write it down.

7:25Expert:Correct, and I think that is a mistake dressed as a concession. The inventory was the cheap part. Removing it means the first time most employers map their automated decision systems will be under a notice deadline rather than a planning cycle.

7:40Host:Is there evidence any of this gets enforced, or do these disclosure laws just sit there?

7:44Expert:That is the honest sceptic's case, and there is data. Bloomberg Law notes New York's governor imposed a similar AI-layoff disclosure by executive order in 2025, and no employers identified layoffs as AI-related within the rule's first year. Zero. Either nobody was displaced by automation in New York, or nobody wanted to write it down.

8:07Host:Which do you think it is?

8:09Expert:The second, obviously. But California added a feedback loop New York lacked. The Employment Development Department must publish summaries of technology-displacement notices and quarterly statewide reports, and deliver a report to the Legislature by 1 January 2028. Published aggregate counts create pressure. Silence becomes visible.

8:30Host:Where does a company outside California land on this?

8:33Expert:Treat it as the template. Connecticut already requires layoff notices to say whether cuts are tied to AI. California also built the audit plumbing: Newsom signed SB 813 creating a framework for independent verification organisations that assess AI systems for safety and risk, and AB 1405 creating a state registry for AI auditors. Independent assessors with a registry means somebody will eventually ask to see your records.

9:03Host:One concrete thing to do this week.

9:06Expert:Pull a list of every system in your stack that outputs a score, ranking or flag about a named employee: productivity metrics, quality scores, attendance models, attrition risk, call-centre QA. For each one, answer two questions in writing. Do we retain the input features and the output for that individual on that date, and can we name the human who saw it before an action was taken. Where the answer is no, you have until 1 July 2027, when SB 947 takes effect, to fix the logging. Start with the systems feeding performance management, because that is where discipline decisions originate.

9:45Host:What we read for this one: No Robo Bosses Act: California's Essential AI Firing Warning, Gov. Newsom wraps California term by enacting 11 more laws on AI safety — Transparency Coalition. Legislation for Transparency in AI Now., California Passes No Robo Bosses Act, California Governor Signs 3 Bills Targeting AI and Workplace Surveillance - Ogletree, New California Law Requires That Humans Decide Firings, Not AI, California Legislature Advances AI Employment Bills. End of episode. The CDO calculators are running at mba-training.com.