California AI Worker Protection Laws Mandate Human Review of Firing

For years, the question hanging over every AI policy debate has been some version of the same anxiety pollsters keep finding: will the machine take my job? California has now answered — not by banning the machine, but by insisting that a human stays in the loop when the machine decides who loses one.

On September 30, 2026, Governor Gavin Newsom signed a package of bills that, taken together, constitute the most aggressive workplace AI regulation in the United States. The new laws prohibit employers from relying solely on an automated decision system when firing or disciplining a worker, require disclosure when a mass layoff is driven by AI, and restrict AI-powered workplace surveillance1319. They arrive on top of a regulatory scaffolding the state has been assembling for two years — frontier model safety rules, content watermarking mandates, algorithmic hiring regulations, and automated decision-making privacy rules — that collectively makes California the closest thing America has to a functioning analog of the EU AI Act1216.

What the new laws actually do

The core of the package is a human-in-the-loop requirement. Employers may not treat an algorithm's output as the final word on discipline or termination. Where an employer primarily relies on an automated decision system, it must have a human corroborate the decision using supporting material — work product, manager evaluations, personnel files — and if that corroboration cannot be produced, the algorithmic output cannot be used at all510. Workers affected by such decisions must receive a written post-use notice explaining that an ADS assisted the decision, describing their right to access the data used about them, and providing contact information for a human being who can answer questions about it10.

A second measure extends the existing Cal/WARN layoff-notification regime into the automation era. When a qualifying mass layoff or relocation is caused in whole or in substantial part by AI or automated technology replacing workers, the notice must say so explicitly — identifying the number and classification of displaced jobs, the job functions being automated, and the specific category of AI system responsible513. That last detail is quietly remarkable: the state is effectively building a public paper trail of which companies are automating which roles.

The package also reaches surveillance. Employers are barred from using workplace surveillance tools that recognize or make inferences about a worker's emotional state, and from collecting neural data, with narrow exemptions for safety and certain federal defense-related work10. A separate provision bans the use of surveillance tools in workplace bathrooms13. Other measures require California public higher-education institutions to provide AI training and procurement guidance, and prohibit attorneys from fully delegating core legal work — drafting briefs, exercising legal judgment — to AI systems13.

Notably, the signing came a year after Newsom vetoed the "No Robo Bosses Act," a broader 2025 bill that would have imposed notice, oversight, and corroboration duties across a wider range of employment decisions1520. The 2026 package is the narrower second attempt, and its passage over that veto's objections suggests the legislature and the governor converged on a principle: automation-assisted management is acceptable, unaccountable automation is not.

A compliance stack, not a single statute

What makes California's approach distinctive — and burdensome — is that these workplace rules sit atop an already-dense stack of AI obligations with staggered deadlines. The Fair Employment and Housing Act regulations governing automated decision systems took effect October 1, 2025, applying to employers with five or more workers and covering everything from resume screening to targeted job advertising; they require anti-bias testing, pre-use and post-use notices, and four-year record retention, with liability extending to both employers and the vendors supplying the tools1217.

The California Privacy Protection Agency's automated decisionmaking technology regulations add another layer: businesses using ADMT for "significant decisions" — employment decisions are expressly identified as such — must conduct prescriptive, detailed risk assessments and provide opt-outs with genuine alternative pathways, with compliance required by January 1, 2027 for systems already in use1518. Meanwhile, AB 2013's training-data transparency requirements became effective January 1, 2026, and the AI Transparency Act (SB 942 as amended by AB 853) — requiring detection tools, visible labels, and embedded provenance metadata from generative AI providers with over one million monthly users — became operative August 2, 20262317.

The common planning error, one compliance analysis warns, is treating any single statute as the whole framework: a company can be fully compliant on content labeling while missing the more operationally demanding ADMT notice and risk-assessment duties entirely4.

The Brussels shadow

The EU AI Act hangs over all of this, sometimes explicitly. California's AB 853 deliberately delayed the AI Transparency Act's operative date to August 2, 2026, aligning with the EU AI Act's enforcement date for high-risk systems — a signal that the state is consciously building interoperability with the European regime rather than a silo4. The frontier-model law, SB 53, goes further: it contains a novel federal-deference provision under which a company that satisfies comparable standards elsewhere — including compliance with the EU AI Act — can have that compliance recognized in California rather than filing duplicate disclosures16.

There are substantive divergences. SB 53's compute threshold of 10^26 floating-point operations exceeds the EU AI Act's 10^25-FLOP trigger, meaning California's frontier regime captures fewer, larger developers16. And on enforcement, California is lighter than Brussels and far lighter than New York's RAISE Act, which carries penalties of up to $10 million for a first violation against California's $1 million ceiling for SB 53 disclosure failures16. But the direction of travel is unmistakable: roughly five to eight frontier developers — OpenAI, Anthropic, Google DeepMind, Meta, Microsoft — now face binding statutory duties, standardized safety frameworks, 15-day critical-incident reporting to the state emergency services office, and whistleblower protections with a private right of action and a burden-shifting evidentiary structure141516.

Why the employer response will define this era

The industry response so far has been measured. Anthropic publicly endorsed the final version of SB 53, saying its requirements largely match practices the company had already adopted; OpenAI and Google have not opposed it but have lobbied for federal preemption to avoid a state-by-state patchwork16. That is the fault line to watch. Employers deploying AI — as opposed to developing it — mostly fall outside SB 53's scope today, but the FLOP threshold will sweep in more companies as models grow, and the workplace package imposes duties on deployers directly, with civil penalties and Labor Commissioner enforcement attached1015.

For compliance teams, the practical read is this: California has effectively decided that the honest answer to "will AI take my job?" is sometimes yes — and that if it does, the worker is entitled to know it, to have a human confirm the exit decision, and to see the data the machine used. The state has bet that transparency obligations, layered from frontier labs down to HR departments, can generate the informational infrastructure future regulation will require without strangling the industry that accounts for 15.7% of the nation's AI job postings and more than half of global AI venture funding11.

The gamble's success depends on enforcement and on whether other states follow. What is already settled is the principle. As one summary of the 2026 package put it, the laws do not ban AI at work — they insist that a human stays in the loop and that workers get told when software is shaping decisions about their livelihoods19. After two years of vetoes, amendments, and phased deadlines, that is no longer a proposal in California. It is the operative law of the country's largest state economy, and given the deliberate EU alignment, increasingly the default for anyone building or buying AI that touches people's jobs.